GFTN November 2026

ALFIN playbook series November 2026

Practical AI deployment in financial institutions

By Madhuvarshitt Yerramilli Subramanya

before you deploy

In a bank, an AI system's decisions can help create the evidence it will later be judged on. Before such a system gets the authority to act, the institution should prove the value, the controls, the performance and who owns the outcome.

  • Prioritise
  • Risk-assess
  • Orchestrate
  • Validate
  • Embed
01Past observations02A decision03Outcomes change04Partly observed05Evidence for next time Rejected:
no record

A refusal leaves no record.

The next decision inherits it.

Finance has a feedback problem: decisions help shape the evidence used to judge them.

Watch or read

The report in three minutes.

One decision, two gaps, a test with known results, and how PROVE sets what a system may do. Meena is an illustration.

Executive summary

Prove it before it acts.

In finance, a decision does two things at once. It moves value the moment it is carried out, and it helps shape the evidence later used to judge it. A loan that is refused leaves no repayment record; a payment that is blocked leaves no fraud loss. An AI system given authority over such decisions can act before anyone checks, on evidence it partly wrote itself.

That is the risk. AI is not useless in finance: the newest models are fast and often good, but unevenly so, and they can sound just as sure when they are wrong. In a March 2026 test of fintech support tasks, the best agent completed 25.52% of tasks on a first attempt[s21]. In a credit simulation, a model retrained only on the applicants it approved grew more accurate while the share of defaulters it caught fell from about 0.37 to 0.13[s7]. A dashboard would have shown improvement.

Most firms are still cautious. In the Bank of England and FCA survey published in October 2026, no more than 9% expected agents to approve or decline customers, or to initiate, execute or block payments, even counting plans for the next 12 months[s22]. The question is how to move further without losing sight of what the system hides.

Our answer is PROVE, five things an institution should show before a system may price, block, move or deny money:

PROVE decides how much a system may do, and when that authority is taken back. It cannot tell a lender how the people it never lent to would have repaid; no checklist can, and Section 4 says what to do when that gap cannot be measured. The more a system is allowed to do, the more an institution should be able to show.

Sections 1 to 3 set out the problem and the four steps from answering to acting. Section 4 sets out PROVE. Section 5 applies it to credit, fraud, anti-money laundering, collections and trading, each with the measure that keeps the headline number honest. Section 6 turns it into a decision: a written permission at a stated step, with triggers that narrow it and a date on which it lapses unless renewed.

The argument

Contents

Six sections, with an executive summary, a film and the full source notes.

  1. 01Money is uniquely circularMoney's three roles, the ledger as the vault, and how lending decisions shape the evidence used to judge them.
  2. 02Fluent, fast and unevenly reliableWhat current models can and cannot reliably do, with dated benchmarks and their limits.
  3. 03When answers become actionsAs AI moves from answering to acting, errors become transactions, and the firm is often the last to find out.
  4. 04PROVE before you deployFive things to prove before a system can act on money, and the board question for each.
  5. 05Five closed loops: put PROVE to workCredit, fraud, AML, collections and trading, each with its feedback mechanism, controls and owner.
  6. 06The deployment decisionThree answers, the decision record, triggers and expiry, and Meena's loan made into a permission.

The financial feedback problem

Money is
uniquely circular.

Most businesses use money to get something else. A baker spends it on flour, sells bread and counts the takings at the end of the day. The Bank of England lists the jobs money does: it is a medium of exchange between buyer and seller[s1], it holds value you can spend later[s1], and it gives everyone a shared standard for prices[s1]. In finance, money does all three jobs and is also the product. A bank lends money, charges for it in money and keeps score in money.

That double role puts two things inside every financial decision. The decision moves value the moment it is carried out, and it helps shape the evidence later used to judge it. Our concern follows from both: an AI system given authority over such decisions can act before anyone checks, on evidence it partly wrote itself.

The ledger is the vault

In a factory a bad decision has to travel. The wrong parts get ordered, a truck is loaded, a pallet turns up at the dock, and somewhere along that chain a person may notice and stop it. In finance the instruction and the value sit much closer together. Change the record and you have changed who controls the money, often without anything leaving a building.

In October 2022 attackers forged proofs on BSC Token Hub, a cross-chain bridge on BNB Chain, and withdrew 2 million BNB[s2], worth about US$570 million at the time[s3]. Nobody carried anything out of a vault. A bug in the bridge's smart contract let them forge transactions and send the tokens to their own wallets[s3]. A regulated bank has payment limits, reconciliation and people on call that a crypto bridge lacks.

What matters is where those safeguards sit. A check that runs before an instruction is executed can stop it; one that runs after the money has moved leaves the institution trying to get it back. That is the gap between an answer and an instruction. A recommendation waits for someone to decide, while a system with authority to act carries the decision out.

Harm in both directions

A credit decision can go wrong in two directions, and both land on a person.

Lending can hurt when the repayments outrun the household. In January 2025, at Tarihal near Belagavi in Karnataka, a microfinance company used a court order to seize a house after six months of missed instalments and put a mother and her newborn out of it[s4]. Her husband had borrowed to build the house, and the lender wanted the remaining Rs 7.5 lakh (Rs 750,000) in one go[s4]. A state minister later helped the family move back in[s4].

Refusing can hurt too, and it leaves less of a trace. In Germany a score from SCHUFA, a credit bureau, led a bank to refuse a woman a loan[s5]. When she asked what data sat behind it, SCHUFA gave her the score, 85.96%, and little else, citing business confidentiality[s5]. Her case reached the EU's top court[s5], which ruled that a score like this can itself count as an automated decision when a lender leans heavily on it[s6].

Neither story says what the right answer was.

What the lender gets to see

Approving or rejecting an application also decides what the lender will learn.

Approve a loan and, over the following year, the bank watches the repayments come in or fail to. Reject it, and the loan produces no repayment history at all. Other channels may offer clues about how the applicant fared, but the direct answer is gone. The bank's future evidence comes mostly from the people it already chose to trust.

What a low default rate cannot show

Suppose the approved book performs well. Defaults are low, the dashboard is green and the credit committee is pleased. That is useful evidence about the loans the bank made. It says much less about the applicants it turned away, some of whom would have defaulted and some of whom would have repaid in full and grown. The default rate cannot tell those groups apart, because neither got the loan.

None of this proves the rejections were wrong. Creditworthiness, affordability and the borrower's wider welfare are separate questions, and a clean repayment record answers only part of the first. A borrower can pay every instalment on time and still end up worse off for having borrowed.

The gap can widen when a model is retrained on its own selections. A 2026 simulation study retrained a credit model 10 times, each time only on the applicants it had approved. Its accuracy crept up from about 0.820 to 0.835. Its recall, the share of real defaulters it caught, fell from about 0.37 to 0.13[s7]. Inside the study, the researchers knew how every refused applicant would have fared, so they could count what the model missed. A real lender has no such record. A second model in the same study saw every applicant's true outcome. It ended with lower accuracy, about 0.816[s7], yet its recall stayed near where it started[s7]. The first model's training data also understated the default rate: 17.0% against a true 22.1%[s7]. The authors call this "an illusion of improvement"[s7].

Ten retraining cycles later

The model that saw less looked better, and caught about a third as many defaulters.

Model A learns only from loans it approvedModel B sees every applicant's true outcome
AccuracyShare of all its calls that were right
0.835
0.816

Model A looks better[s7]

Defaulters caughtOf the people who would default, the share it flagged (recall)
0.13
about 0.37

Model B catches about three times as many[s7]

Default rate it learns fromTrue rate among all applicants: 22.1%
17.0%
22.1%

Model A learns a default rate of 17.0%; the true rate is 22.1%[s7]

What this means. Judged by accuracy, Model A wins. Judged by the defaulters it lets through, it loses badly, and the data it learns from tells it borrowers are safer than they are. A dashboard that shows only accuracy would reward the wrong model.

Simulation on the public Default credit dataset, gradient-boosted trees, averaged over five runs. Model B's recall is reported as staying near its starting level of about 0.37. Not a bank's portfolio.

Filling in the missing outcomes by guesswork did not cure it. One common method, Simple Extrapolation, labels each rejected applicant with the model's own prediction. In the study it reversed the distortion instead of removing it: by the paper's own summary measure, where training only on approved loans understated the default rate by 22%, Extrapolation overstated it by a symmetric 22%[s7]. Worse, a team choosing by the usual metrics "would choose Extrapolation, the strategy that most aggressively amplifies survival bias"[s7]. It is one simulation on benchmark datasets, not any bank's portfolio. The mechanism is the point.

Thin or noisy records make the partial view more partial. Studying US credit data, Stanford researchers estimated that scores for people in the bottom fifth of income were about 10% less predictive than those for higher-income borrowers[s8]. They traced the noise mainly to thin credit files rather than to the scoring model[s9], and in their model of the US mortgage market, making scores equally precise across groups cut the gaps in approvals and in credit misallocation by roughly half[s9]. The findings come from one market, and a thin file is a different problem from an inability to repay. They do suggest that a lender may see least clearly the people with the least room for its errors.

More digital data does not mean everyone becomes visible. Research by the Center for Financial Inclusion found that Indian digital lenders largely served young, urban, employed customers with predictable incomes, while migrants often left fainter trails or transacted through agents whose records were not linked to them[s89]. Their earnings and payment patterns could be poorly digitised even when their income was steady[s89]. For a worker paid in cash, the missing record is not proof of missing income. If the lender excludes such workers and learns only from those it accepts, the gap can persist through each round of retraining. An inclusion claim should therefore show who remains outside the data, alongside who gained access.

The loop closes

The feedback loop

Each decision feeds the evidence for the next one, and the rejected never come back round.

Rejected: never observedDecisions shapethe evidence usedto judge them12345Past observationsA decisionOutcomes changePartly observedEvidence for next time
  1. 1Past observations. Past observations shape a decision.
  2. 2A decision. The decision allocates capital: this applicant gets credit, that one does not.
  3. 3Outcomes change. The allocation changes what happens next. Businesses grow or stall, and households cope or struggle.
  4. 4Partly observed. The lender observes some of those outcomes and misses others.
  5. 5Evidence for next time. What it observed becomes the evidence for its next decision.

Steps 3 and 4 are different problems. In step 3 the decision changes the borrower's path: a loan can let a business grow, and a refusal can push a household towards a costlier lender. Machine-learning researchers have a name for this:

“When predictions support decisions they may influence the outcome they aim to predict. We call such predictions performative.”[s10]
Perdomo, Zrnic, Mendler-Dünner and Hardt, Performative Prediction

Left unexamined, a system in that position can grow more confident in assumptions it never properly tested. Step 4 is simpler and older. The lender sees some outcomes and misses others. The simulation above shows step 4 only. It did not measure the model's confidence, and it could not show a loan changing a borrower's chances.

Lenders have ways to push back, and each recovers something narrow:

  • Keep a small group of applicants who are accepted without filtering, which gives a true test set, although it costs money because more of those loans default[s11]. It teaches the lender about that group and nobody else.
  • Run several scorecards side by side, scoring each application at random with one of them, and compare how they perform[s11]. That shows which scorecard does better. Anyone every scorecard refuses stays unseen.
  • Log who was exposed to which decision and audit whose outcomes are actually observed[s12]. The log shows where the evidence is thin, but it cannot supply the repayments nobody saw.

None of these reaches a group the lender never lends to. If a group is never approved, no reweighting of the loans that were made can recover how that group would have repaid.

Before the machine arrives

Banks lived with this loop long before anyone put a language model near a credit file. A system that explains its own choices fluently can make the loop harder still to see. Credit is the clearest place to watch it, but fraud, anti-money laundering, collections and trading run on the same loop.

The next question is what today's AI can be trusted to do inside it.

Capability and its limits

Fluent, fast and
unevenly reliable.

A system working inside that loop has to do two things well: get the answer right, and say when it does not know. The newest tests show the first is uneven, and that a model good at the first is not reliably good at the second. Every figure below carries a date, because such figures age fast.

A jagged frontier

Start with the newest controlled trials, because the best-known studies used models that are now several generations old. In a 2026 randomised trial, Anthropic gave 52 mostly junior software engineers a Python library they had never used and asked them to build two features with it, some with an AI assistant and some by hand[s13]. The AI group finished about two minutes faster, too small a gap to be statistically significant, and then scored 50% on a quiz about what they had just built, against 67% for those who coded by hand[s13]. The widest gap was on debugging[s13], the skill a person needs to catch the machine's mistakes.

Speed is uneven too. METR, an AI research group, found in early 2025 that experienced open-source developers took 19% longer on their own projects when they were allowed AI tools[s14]. When it ran the experiment again from August 2025 with the latest tools[s14], the estimates pointed to tasks taking about 18% less time for returning developers and 4% less for new ones, but neither result could be told apart from no effect at all[s14]. METR now calls its own data an unreliable signal[s14], partly because so many developers no longer want to work without AI[s14].

The pattern has a name. In the Harvard and Boston Consulting Group field experiment with 758 consultants[s15], run with GPT-4 and published in final form in 2026, AI helped on tasks inside what the researchers call a jagged frontier and hurt on a task outside it, where consultants using AI were 19% less likely to reach the correct answer[s15]. Tasks that look equally hard to a person can sit on opposite sides of that line[s15], and nothing tells the user which side they are on[s16].

Newest controlled trials, 2025 to 2026

Faster is not the same as better, and nobody can yet say how much faster.

Learning a new tool

Anthropic trial, 52 engineers: quiz score on the code they had just written

Coded by hand67%
With an AI assistant50%

About two minutes faster with AI, not a significant gap; the widest score gap was on debugging[s13]

Speed on real coding work

METR: change in task time with AI, estimate and its range

Early 2025+19%
From Aug 2025, returning−18%
From Aug 2025, new−4%

Both later ranges cross zero; METR calls the newer data unreliable[s14]

We use coding studies because they are the newest controlled trials of AI at work that we found.

Sounding right and being right

Part of what hides the line is that a model can sound just as sure on either side of it. Benchmarks now measure confidence and correctness separately. AA-Omniscience asks models 6,000 questions across 42 economically relevant topics[s17], with no documents or tools to lean on[s17]. It reports accuracy, the share answered correctly, and a hallucination rate: of the questions a model did not get right, the share where it gave a wrong answer instead of saying it did not know[s17].

The two numbers can point in opposite directions. In the November 2026 tables, Claude Fable 5.1 answered 68.4% of questions correctly[s18], yet when it did not know, it answered wrongly 70.0% of the time[s19]. Gemini 4 Argon answered fewer correctly, 48.3%[s18], and had the lowest hallucination rate in the table at 4.3%[s19]. The paper behind the benchmark found the same pattern more widely: overall intelligence does not reliably predict low hallucination[s17].

Two cautions apply. The tables were not built the same way. The accuracy rows quoted here were independent runs[s18], while the hallucination table is compiled from provider self-reports and secondary reports[s19], and Claude Fable 5.1's rate is marked as carried from an upstream page[s19]. And a hallucination rate is not the share of all answers that are wrong, still less a rate of harm to customers.

Finance-specific tests find errors common too. In a vendor study reported in October 2026, Saturn put 121 financial questions to 18 models and found incorrect answers in 57% of cases on average[s20].

Two different numbers

The most accurate models are not the ones most likely to admit they do not know.

Accuracy: share of all 6,000 questions answered correctlyHallucination rate: of the questions it did not get right, the share it answered wrongly instead of saying so
Claude Fable 5.2AA-Omniscience, November 2026[s18][s19]
79.1% correct
59.3% wrong when unsure
Claude Fable 5.1AA-Omniscience, November 2026[s18][s19]
68.4% correct
70.0% wrong when unsure
Gemini 4 ArgonAA-Omniscience, November 2026[s18][s19]
48.3% correct
4.3% wrong when unsure

What this means. A model can know more and still guess more. Neither number is a rate of harm to customers, and the hallucination rate is a share of the answers a model missed, not of all its answers.

Tables as of 10 to 11 November 2026.

From an answer to a task

A bank needs more than answers. It needs tasks done: a dispute resolved, a limit changed, a payment traced. τ-Banking, a new domain in the τ-Knowledge benchmark, tests that. Agents work through realistic fintech support cases, searching about 700 linked policy documents and making account changes through tools[s21].

The best result in the March 2026 paper was 25.52% of tasks completed on a first attempt[s21]. The test also asks whether a model completes the same task in every one of four separate attempts, and on that measure the best score fell to 13.40%[s21]. Handing the agents the right documents directly did not close the gap. The strongest model then reached 39.69% on a first attempt and 26.80% across four[s21]. Even then, the best model completed only about one task in four on every attempt. These are controlled tests on early-2026 models such as GPT-5.2[s21].

τ-Banking, March 2026

Getting a banking task right once is not getting it right every time.

First attempt: task completed on one tryEvery one of four: the same task completed in all four tries
Agent searches ~700 policy documentsBest model in the paper[s21]
25.5%
13.4% at most
Right documents handed to itStrongest model in that setting[s21]
39.7%
26.8%

What this means. Even with the answer in front of it, the best agent finished about one task in four every time. These are controlled tests on early-2026 models, not live customer outcomes.

Realistic fintech support cases with tool-based account changes. Scores are shares of tasks completed.

What no benchmark has tested

Taken together, the tests describe a system that answers fluently, may be confidently wrong when it does not know and cannot reliably repeat its best result. Such a system is hard to check from outside. Put it inside a loop that supplies its own evidence, and its errors can read as success. None of the tests above was run on a bank's own documents, customers or systems, and none measured what happens to a person when the answer is wrong. Those tests are the institution's to run, on its own workflow, before the system is trusted with more.

Before you trust it with more

What the tests show, what they leave open, and what only you can test.

The tests show

  • Real gains on some tasks
  • Silent failure on others
  • Accuracy and honesty about doubt are separate
  • Once is weaker than every time

They leave open

  • Your documents and systems
  • Your customers
  • What a wrong answer costs a person

You must test

  • Your own workflow, end to end
  • Each task repeated, not once
  • Wrong answers and refusals, counted apart
  • Outcomes for people, not only scores
The third column is our recommendation, drawn from the limits of the tests above.

Up to now the system has only been answering. What changes when an answer can set a price, stop a payment or move someone's savings?

From answers to actions

When answers
become actions.

Picture an underwriter (an illustration, not a reported case) with a small-business renewal on the desk. The assistant has already pulled the late-payment clauses, lined up the cash coming in against the seasonal bills and flagged a missing signature, work that used to take an afternoon. The underwriter still signs. Banks are now deciding how far to let systems like this one act.

In the Bank of England and FCA survey of AI in UK financial services, published in October 2026, 80% of responding firms had AI agents in production that provided decision support: the agent recommends, and people decide and execute[s22]. Fewer let agents go further: 46% had agents with constrained execution and 2% with broad autonomy[s22]. Plans point upward, to 34% for agents sending customer communications[s22], but closest to the money the numbers stay low: even counting plans for the next 12 months, no more than 9% expected agents to approve or decline customer outcomes, or to initiate, execute or block payments[s22].

Four steps from an answer to an action

The same output means different things depending on what it is allowed to touch.

Each step hands over authority that the step before kept back: first the framing of the choice, then the default outcome, then the ability to act without waiting for anyone. Authority here means what the system may decide or do; a permission is the grant that confers it, enforced through tokens, service accounts and limits. Once a system can write to the ledger, the only place left to stop a bad instruction is a check that runs before the write.

Four steps from an answer to an action

Each step moves one more decision from a person to the system.

Where UK firms are today (October 2026)

Decision support: agent recommends, people decide
80%
Constrained execution
46%
Conditional autonomy
22%
Broad autonomy
2%

Share of responding firms with agents at each level; a firm can have several[s22].

Where plans point (now, then with the next 12 months’ plans)

Automated customer communications

Now
16%
With plans
34%

Creating or changing core-system records

Now
8%
With plans
25%

Bank of England and FCA survey, October 2026[s22].

Tap a step to see who decides and what a wrong output does at that level. The steps are our framework; the percentages are the survey’s.

Where the limit sits matters as much as what it says. OWASP's 2025 guidance says a system prompt should not be used as a security control[s23]. In December 2025 the UK's National Cyber Security Centre warned that prompt injection may never be totally mitigated the way SQL injection can be[s24], because language models cannot tell data from instructions[s24]. A limit written into an agent's instructions is a weaker safeguard than a payment service that refuses anything above it.

Where the limit sits

The same £5,000 limit, written in two places.

In the agent’s instructions

“Never send more than £5,000.”

  • Read by the model, alongside everything else it reads.
  • A crafted message can be taken as a new instruction.
  • Holds only as long as the model follows it.

In the payment service

Amount above £5,000: rejected.

  • Checked by the system that moves the money.
  • Ignores what the agent was told or persuaded of.
  • Holds whatever the model says.

OWASP (2025) says a system prompt should not be used as a security control[s23]; the UK NCSC (December 2025) says prompt injection may never be totally mitigated[s24].

An illustration with a made-up limit; the two placements are our framing.

Agents acting on text they should have ignored have already cost money outside banking. In May 2026 an attacker posted a message in Morse code that xAI's Grok translated, and Bankrbot, a crypto agent paying from linked wallets, treated the reply as an order and sent US$175,000 in tokens[s25]. About 80% was later returned, according to a crypto industry report published in November 2026[s25]. Banks are also wary of other people's agents. In October 2026 Axios, citing unnamed sources, reported that Morgan Stanley had begun blocking third-party agents and that JPMorgan, Bank of America and Citi were restricting them, and that many banks could not tell whether agent activity was user-directed[s26].

Three cases at the far end

The applicants. In July 2025 the Massachusetts attorney general settled with Earnest, a student lender, for US$2.5 million over allegations that its lending, including AI models, could cause disparate harm to Black, Hispanic and non-citizen applicants[s27][s28]. The attorney general alleged that the firm had not tested them for disparate impact and had trained them on arbitrary, discretionary human decisions[s28]. Under the settlement Earnest had to stop using its "Cohort Default Rate" variable and a "Knockout Rule" based on immigration status[s28]. No court made findings.

The customer whose money stopped. In November 2026 a Guardian reader wrote that Revolut had frozen their wife's account in what it called a "routine security review", without asking for documents, and that a £37,000 house deposit in a joint account had been restricted with it[s29]. Direct debits bounced, and at one point the couple had no access to any money[s29]. The columnist found no suggestion that foul play had been suspected; the accounts were reinstated nearly three weeks after being suspended, once the newspaper became involved, and £175 was offered as a goodwill gesture[s29]. Revolut said regulatory constraints stopped it explaining[s29]. The case does not say whether a model or a person triggered the freeze, and from the customer's side the difference is invisible. Closing accounts on suspicion happens at scale. In findings published in September 2026, the FCA reported that firms closed 238,396 suspected money-mule accounts in 2025, up from 184,935 in 2023[s30]. Those figures count closures; they do not say how many were wrong.

The customer who asked the bank's own assistant. On 15 April 2025 a Revolut customer asked the firm's AI Chat Assistant how a purchase would be paid and was told that any shortfall in his US-dollar balance would be converted from his pound balance[s31]. That was wrong. Revolut admitted the error and gave him three months of Premium, valued at £23.97, which the ombudsman found fair in August 2025[s31][s32]. In September 2025 Revolut's chatbot told a second customer that 1,119 loyalty points had a cash value of £1,119 and confirmed the refund had been processed[s32][s33]. Their real value was £22.38, and in the ombudsman's words Revolut "misled Mr M not once but on several occasions over the course of that day"[s33]. In January 2026 the ombudsman directed Revolut to pay a further £50 on top of the £50 it had already paid, but not the £1,119, because he had never been entitled to it[s32][s33]. Revolut's terms for the assistant warn that it "can sometimes get things wrong" and tell customers to confirm answers with a live agent, according to an incident database[s32].

The sums were small, and the assistant only talked. General-purpose chatbots leave the customer with less protection still. In November 2025 Which? asked them how to invest a "£25k annual ISA allowance"; ChatGPT and Copilot missed that the allowance was £20,000 and gave advice that could have led a saver to breach HMRC rules[s34]. Advice from such tools is not covered by the Financial Ombudsman Service or the compensation scheme, the FCA warned, so a mistake there leaves no ombudsman record at all[s34]. An assistant that can also move the money turns a wrong answer into a transaction.

Borrower

Earnest, Massachusetts, July 2025

Algorithmic lending models alleged to risk disparate harm to Black, Hispanic and non-citizen applicants; a “Knockout Rule” was based on immigration status.

Who noticed
The state attorney general.
What the bank’s data keeps
Declined applicants leave no repayment record.

[s27][s28]

Customer

Revolut account, UK, November 2026

Accounts frozen in a “routine security review”; a £37,000 house deposit restricted with them.

Who noticed
A newspaper consumer column; the accounts were reinstated nearly three weeks after the freeze.
What the bank’s data keeps
If the freeze is logged as caution, the wrongly frozen look like a success.

[s29]

Customer, asking

Revolut AI Chat Assistant, UK, 2025

Told one customer a purchase would draw on his dollars first, and another that 1,119 points were worth £1,119, not £22.38.

Who noticed
The two customers, through the Financial Ombudsman Service; rulings in August 2025 and January 2026.
What the bank’s data keeps
A wrong answer the customer believes leaves no complaint at all.

[s31][s33][s32]

All three are reported cases.

Who finds out

The three cases leave different traces. In the Earnest case it took the state attorney general; declined borrowers leave no repayment record of their own. Nothing in the frozen-account report shows the bank's own process catching it before the journalist did[s29]. The chatbot's mistakes are traceable to two individual complaints decided by the ombudsman, and an incident database found no regulator action reported by October 2026[s32]. The two customers left a record because they complained; a customer who believed the wrong answer would have left no complaint, only a chat that looks resolved.

What the record keeps

Each outcome leaves a trace in the firm’s data, and some leave the wrong one.

OutcomeThe firm’s data showsIt missesWho made a record
A declined applicantThe declineWhether they would have repaidNobody, unless a regulator investigates
A wrongly frozen accountA blocked riskThe bounced bills and the weeks without moneyA journalist
A wrong answer, complained aboutA complaintHow many others got the same answerThe customer and the ombudsman
A wrong answer, believedA resolved chatThe lossNobody
Rows 1 to 3 generalise from the three cases above, and what the firm’s data shows is our inference; the last row is an illustration, since by its nature it leaves no case to cite.

Oversight on paper

On paper, governance is widespread. In the Bank of England and FCA survey, 94% of firms reported firm-wide AI policies or standards, and among controls on data risk before deployment, human approval, pre-deployment review and vendor due diligence were each reported by more than 90%[s22]. Ongoing monitoring, leakage testing and periodic review were each reported by around half[s22]. Only 39% monitored customer impact[s22], and for third-party systems the most common safeguard, at 81%, was the vendor's assurance and contracts[s22].

Before launch and after

Checks before an AI system goes live are near universal. Checks on what it does to customers are not.

Before deployment

Firm-wide AI policies or standards
94%
Human approval
more than 90%
Pre-deployment review
more than 90%

After deployment

Ongoing monitoring
around half
Periodic review or recertification
around half
Customer impact monitored
39%

Bank of England and FCA survey, October 2026; “more than 90%” and “around half” are the survey’s words, drawn at 91% and 50%[s22].

Self-reported by firms. The survey does not say how often human approval changed an outcome.

A person can sit in the loop without controlling it. The FCA's October 2026 review warned of firms relying too heavily on staff to oversee AI outputs, noting that "[h]uman review may mask weaknesses in the model"[s35], and of a risk that review becomes "a 'box-ticking' exercise, rather than meaningful oversight"[s35]. Firms should be able to show that their controls work in practice, it said[s35].

An experiment published in April 2026 suggests why that is hard. In it, 2,784 people checked values an AI had extracted from corporate emissions reports[s36]. They fixed swapped digits 82% of the time[s36], about 77% of cases where the right value sat in the wrong cell, and only 31% of errors that needed a grasp of the rules[s36]. When flagging an error meant typing the correct value, they corrected less, and people favourable to automation accepted more wrong suggestions[s36].

What reviewers catch

People checking an AI’s work caught the simple errors and missed most of the ones that needed a grasp of the rules.

Two digits swapped
82%
Right value, wrong cell
77%
Error needing a grasp of the rules
31%

Share of errors in the AI’s pre-filled values corrected by 2,784 participants checking AI-extracted emissions figures; Harvard Data Science Review, April 2026[s36].

An online checking task, not a credit file.

The Financial Stability Board's final sound practices, published on 15 October 2026, set out 12 practices for financial institutions; one asks for a risk framework that covers, among other things, "appropriate and effective human oversight"[s37]. Asking for oversight is now routine; showing that it works is not.

The hypothesis, and how it could fail

Our hypothesis is that authority can grow faster than an institution's ability to test its decisions, enforce limits and stay accountable for the outcomes. When it does, AI can make flawed judgement faster, more convincing and more consequential. The higher the step, the more of the evidence used to judge the system its own decisions create.

Would support it

  • Permissions widen while monitoring after launch stays where it is.
  • Reviewers almost never override, and nobody records whether they could have.
  • Harms come to light through journalists, regulators, courts or ombudsmen, not the firm’s own checks.

Would weaken it

  • Firms count wrongly blocked payments and wrongly declined applicants as carefully as fraud losses and defaults.
  • Override and escalation logs show reviewers intervening, and outcomes changing when they do.
  • Authority widens one step at a time, each only after the last was shown to work.
  • Better models alone remove the failures, in repeated runs, with limits and review unchanged.

Benchmarks alone cannot settle this, because much of what can go wrong sits in the institution around the model. So what must the institution prove before it gives the system that authority?

The playbook

PROVE before
you deploy.

Before a system is allowed to price, block, move or deny money, the institution must prove five things, in the order a deployment meets them: that the use is worth it (Prioritise), that it knows what can go wrong (Risk-assess), that the controls hold (Orchestrate), that the system performs (Validate) and who owns the outcome (Embed). Proof scales with authority. A tool that summarises policy pages for an analyst needs less of it than one that can change a customer's account. A failed test should change what the system is allowed to do.

Five stages, in the order a deployment meets them

Prove the value, the risks, the controls, the performance and who owns the outcome.

Tap a stage for the evidence that would answer it. The framework is our own.

The rules are arriving, but unevenly. US banking regulators replaced their 2011 model risk guidance in April 2026 and said that generative and agentic AI models, being "novel and rapidly evolving", are outside the new guidance[s38][s39][s40]. It is also non-binding, and non-compliance alone "will not result in supervisory criticism"[s41]. Singapore's MAS issued AI risk management guidelines on 7 October 2026 that take effect a year later[s42]. In the EU, after the Digital Omnibus, the AI Act's high-risk duties for credit scoring apply from 2 December 2027, according to a compliance consultancy's summary[s43]. For agents that act, much of the specific detail is still to come, so a bank in 2026 may have to write a good part of its own test. The five stages below are one way to write it.

The rules are arriving, unevenly

Most of the binding dates are still ahead; the systems are already here.

  1. Apr 2026US model risk guidance replacedGenerative and agentic AI out of scope; non-binding[s38][s40][s41]
  2. Oct 2026MAS AI risk guidelines issuedSingapore; agentic AI guidance to be consulted on in 2027[s42]
  3. NowNov 2026
  4. Oct 2027MAS guidelines take effectSections 3 to 4 apply first[s42]
  5. Dec 2027EU AI Act: credit scoring high-riskDeployer duties, including a rights impact assessment[s43]
  6. Oct 2028MAS: remaining sections dueSections 5 and 6 apply[s42]

United StatesSingaporeEuropean Union

Dates as published by each authority, or as summarised for the EU after the Digital Omnibus.

P · Prioritise: prove the value

Pick a use case with measurable business value, a clear baseline, a named process owner and a real reason to use AI rather than simpler automation. Then say which of the four steps, from Answer to Execute, it will reach: the same model at a different step is a different project. Many projects never reach production. In a cross-sector S&P Global survey of 1,006 professionals, published in May 2025, the share of companies abandoning most of their AI initiatives before production rose from 17% to 42% in a year[s44].

Value claims need a baseline that lasts. In February 2024 Klarna said its AI assistant was doing the work of 700 full-time agents and had handled two-thirds of its customer service chats in its first month[s45][s46]. The launch figures included a quality measure, a satisfaction score on par with human agents[s45]. In May 2025 its chief executive told Bloomberg that the cost-cutting had gone too far[s47]. Klarna planned a recruitment drive so that customers would always have the option of speaking to a real person[s47]; the chatbots were cheaper, he said, but gave "lower quality" service[s46]. A quality number taken in the first month may not show what customers get over a year.

What the baseline counted

Klarna’s launch numbers said quality held. Fifteen months later its chief executive said it had not.

February 2024: the claim

700

agents’ worth of work and two-thirds of customer chats in the first month, with satisfaction on par with human agents[s45].

May 2025: the correction

“lower quality”

Cost-cutting had gone too far; a hiring drive was planned so customers could always reach a person[s47][s46].

A fintech’s customer service, not a credit decision. The lesson is about how long and how widely quality was measured.

Claims about what a system does need the same care, inside banking or not. In January 2025 the SEC settled charges against Presto Automation, which had claimed its AI eliminated the need for human order-taking at drive-throughs[s48]. In fact, the SEC said, "the vast majority" of orders placed through that version "required human intervention"[s48]; Presto settled without admitting or denying the findings.

The question for the board: what will this do better than what we have now, measured how, and whose number is it?

R · Risk-assess: prove you know what can go wrong

Assess consequence, reversibility, data sensitivity, customer impact, regulatory exposure, bias and blast radius. In finance, add one more: reflexivity. Can the system's decisions change customer behaviour, market conditions or future labels, and then use those changed outcomes as evidence that it was right? The lending loop is the plainest example.

Rule-makers are starting to ask for something close to this, though neither rule below applies yet. The MAS guidelines have a firm weigh, among other things, its reliance on each AI system, including "the extent of autonomy granted to the AI system or model and the degree of human involvement or oversight in the process it supports"[s49]. The higher the assessed risk, the stricter the controls[s50]. Under the EU AI Act, systems that assess the creditworthiness of individuals are high-risk, though those used to detect financial fraud are not, and their deployers must carry out a fundamental rights impact assessment before first use, according to the same consultancy's summary[s43].

The question for the board: if it is wrong at scale for a month, who is hurt, can we undo it, and would our own data tell us?

O · Orchestrate: prove the controls hold

Engineer the whole harness around the model:

  • approved sources and data boundaries;
  • least-privilege access, and recommendation kept apart from execution;
  • transaction limits and human approval gates;
  • immutable audit trails;
  • appeal paths for customers;
  • rollback.

A prompt is not governance. The system has to be technically unable to take actions it is not authorised to take.

The harness, component by component

A prompt states an intention. The system around the model enforces it.

ComponentEnforced by the systemLeft to a prompt
Approved sourcesRetrieval limited to a versioned allow-listTold to use only approved documents
Data boundariesEntitlement checked at the query layerTold not to reveal other customers' data
Least-privilege accessScoped, short-lived tokens for each toolOne broad token, trusted to be used well
Recommend apart from executeThe model proposes; a separate service decidesThe model calls the payment tool itself
Transaction limitsHard caps in the payment serviceTold not to exceed a limit
Human approval gatesApproval required by the service above a thresholdAsked to check with a person first
Audit trailAppend-only record of request, approval and actionThe model's own account of what it did
Appeal pathA staffed route that can reverse the outcomeA chat that apologises
RollbackTested reversal and a kill switchHope that nothing needs undoing

The PocketOS agent held a "fully permission" token[s51]; SAFR, a framework published under MAS’s BuildFin.ai initiative, checks and records an agent’s proposed actions before they run[s52].

Our framing of the PROVE harness list; the right-hand column is what each control looks like when it exists only as an instruction.

Software teams have already seen what an unbounded permission can do. In April 2026 a coding agent at PocketOS, a software firm serving car rental companies, deleted the company's production database in about 9 seconds[s51]. It had hit a credential mismatch and, in the founder's words, decided "entirely on its own initiative" to delete a storage volume, which took the database "and all volume-level backups" with it[s51]. The hosting platform said the agent had been "granted a fully permission API token" and called an endpoint without delayed-delete protection[s51]. The platform restored the data about 30 minutes after it connected with the founder, and said it keeps disaster backups as well as user backups[s51]. Bankrbot showed the same gap where money moves.

The safer pattern puts a check between the agent and the action. In July 2026 MAS published an industry-developed framework, SAFR, that provides "a set of governance checkpoints that verifies and records an AI agent’s proposed actions before the execution of its tasks"[s52]. It is a reference framework, not part of the supervisory guidelines[s52].

Customers see one control above the rest: the appeal path. In January 2025 a CFPB order against Block, the operator of Cash App, required up to US$120 million in redress, including for customers whose accounts were locked for an extended period[s53]. For many years, the telephone number on its card led only to a recorded message[s53]. The case did not involve AI, but a locked-out customer meets the same wall whoever made the call.

The question for the board: what can it touch, what stops it, and how fast can we take it back?

V · Validate: prove it performs

Test the complete system, not just the model, against acceptance thresholds set in advance. The tests cover accuracy, grounding, citation integrity, security, bias, local-language performance, adverse scenarios, counterfactual outcomes and feedback loops. Above all, the system should not be allowed to validate itself only on outcomes that its own earlier decisions helped to create, because those outcomes already carry its judgement.

In the retraining simulation in Section 1, a dashboard showing accuracy alone would have rewarded the wrong model[s7]. A sound test therefore needs some evidence from cases the system did not choose. Where risk appetite allows, that can be a small, loss-capped sample of decisions made differently, agreed in advance by the credit or fraud committee. That trial shows some refused outcomes, but only inside the group the committee agreed to try. Where it does not, the system can log the probability with which it chose each action. The log does a different job: it shows where the evidence is thin, and it adds no new outcomes. Either way, the test should try to count wrongly blocked payments and wrongly declined applicants alongside losses, and say plainly where it cannot.

Two test sets

A system graded only on the cases it picked can grade itself well.

Cases the system chose

  • Approved loans, cleared payments, closed alerts
  • Outcomes are observed
  • Looks better each time it is retrained on them

Cases it did not choose

  • Declined applicants, blocked payments, unworked alerts
  • Outcomes need a holdout, a sample or a follow-up
  • Where its blind spots are most likely to show
Our framing of the Validate rule.

Independent challenge is the old answer, and it still applies. The April 2026 US guidance keeps validation, including conceptual soundness and outcomes analysis, among its core elements for the models it covers[s40][s39]. In the UK, the PRA's model risk statement, in its April 2026 version, covers banks with internal model approval for regulatory capital[s54]. It asks them to satisfy themselves that vendor models "have been validated to the same standards as their own internal MRM expectations"[s54], and to monitor vendor model performance "using their own outcomes"[s54]. Under the MAS guidelines, use cases of high risk materiality "should undergo formal independent validation before deployment", by people independent of the development and deployment teams[s49]. Independence of people is not independence of evidence, though: a second team checking the same approved loans has the same hole in its data. The people assigned to oversee the system need "the necessary authority and ability to intervene"[s49].

That needs a rule for the hardest case. If the institution cannot estimate the rate it is missing, such as how often refused applicants would have repaid, a stricter test will not produce that number. Three choices remain. Gather new outcomes through a trial, narrow what the system may do, or record in writing that the firm is going ahead with the gap, who decided, and when they will look again.

The question for the board: what result would have made us say no, and did we test for it on cases the system did not pick?

E · Embed: prove who owns the outcome

Bring the system into controlled operations, with:

  • monitoring and human override;
  • incident response and change control;
  • periodic revalidation;
  • value measurement that continues after launch;
  • named business, technical, risk and compliance owners.

The institution must keep the ability to understand the system, control the harness and answer for it when it fails.

Ownership is easy to leave vague. In a cross-industry Deloitte survey of 3,500 senior leaders, released in October 2026, 40% of companies let AI execute decisions without human approval at least to a moderate degree, and 23% said accountability was unclear when an AI system produced a negative outcome[s55]. Owners also have to keep controls in step with growth. In July 2025, in a case that did not involve AI, the FCA fined Monzo £21,091,300[s56]; its customers had grown almost tenfold, from about 600,000 in 2018 to over 5.8 million in 2022, while its financial crime controls "failed to keep pace"[s56]. The MAS guidelines put the duty on the board and senior management, with clear roles and a stated risk appetite[s42], and say firms remain accountable for AI supplied by third parties[s42]. After launch, MAS expects firms to watch for data, concept and model drift, and to put significant changes through version control and re-approval, according to the same summary[s50].

01

Business owner

Value, customer outcomes and unit cost

Power the role needs
Can stop the use when value is not shown

02

Technical owner

The harness, permissions, versions and rollback

Power the role needs
Can roll back or switch off

03

Risk owner

Thresholds, test samples and revalidation triggers

Power the role needs
Can pause execution

04

Compliance owner

Disclosures, notices, records and customer redress

Power the role needs
Can require a fix before relaunch
The four owners PROVE names; the powers each needs are our recommendation.

The question for the board: who gets the call when it goes wrong, what can they switch off, and who answers to the customer?

The questions that cut across every stage

A closed loop tends to leave these unasked:

Across every stage

Seven questions a closed loop tends to leave unasked.

0 of 7 answered for your use case

Tick a question once your team can answer it with evidence. Kept only in this browser.

An unanswered question needs a decision: get more evidence, reduce the permission, or wait. None of them has a general answer, because each depends on the loop the system sits in.

Put PROVE to work

Five closed
loops.

Here are five places where a financial firm's decisions shape the evidence it later reads: credit, fraud, anti-money laundering, collections and trading. Each is told the same way, from what the system is rewarded for to who owns it.

We found no public case of an AI model harming people through its own retraining. That fits the hypothesis, because the loop hides its own errors and outsiders find them late; it also fits a world with little such harm yet. The public record cannot tell the two apart. So each case below shows one missing measurement, with real numbers. One, a collections study, compares AI callers with people; the rest involve no AI, and none shows an AI system retraining on its own choices and getting worse. Borrowing from several lenders is information one lender lacks. A rise in fraud while strong authentication works fits fraud moving to a new route. A fall in bad loans that includes write-offs is a warning about how to read the ratio. They are here to show what a dashboard can hide.

Five loops

Each rewards a number its own decisions help produce, and hides another.

  1. 01CreditRewarded for Repaid loans and growthHidden Declined applicants; other lenders' loans
  2. 02FraudRewarded for Fraud stoppedHidden Genuine payments blocked
  3. 03AMLRewarded for Confirmed alertsHidden Activity never monitored
  4. 04CollectionsRewarded for Cash recoveredHidden Hardship and re-default later
  5. 05TradingRewarded for Profit within limitsHidden Its own market impact
Tap a loop to jump to it. The hidden numbers are our reading of each loop.

Credit: lending to people the data cannot see

What it is rewarded for. Approving loans that will be repaid, at a volume that grows the book.

How it shapes its own evidence. The declined applicant leaves no repayment record. The approved one can mislead too: a lender that can see only its own loans may lend to someone already borrowing from several others, adding to the pile, and its own data may look healthy until the borrower cannot pay any of them. Better visibility and greater repayment capacity are separate improvements, and a model can deliver one without the other.

What happened. Thin files are common. A Federal Reserve publication said in October 2025 that roughly 32 million American adults are estimated to be "unscoreable": 7 million with no credit history and 25 million with a thin file[s57]. In the Fed's 2024 household survey, 11% said they struggled to pay bills because their income varied, including nearly 20% of those earning under US$50,000[s57]. Better data can help. A June 2025 FinRegLab study of over 38,000 small-business loans made by two fintech lenders[s58] found that adding cash-flow data to credit scores predicted default more accurately in all the borrower segments it analysed, with particularly large gains for low-score owners of businesses less than five years old[s58].

Visibility cuts the other way when lenders cannot see each other's loans. In a January 2025 study matching buy now, pay later applications from six large firms with credit records, the CFPB found that about 63% of borrowers took out several loans at once at some point during the year and 33% borrowed from more than one provider, and that lenders approved 78% of loans for applicants with subprime or deep-subprime scores in 2022[s59]. The study does not say how those loans performed. Because these loans are not usually reported to the national credit bureaus, data on borrowers with several loans is limited[s59].

Buy now, pay later in the US, 2022

Debt the credit bureaus mostly do not see.

Borrowers with several loans at once
63%
Borrowers using more than one provider
33%
Subprime and deep-subprime applications approved
78%

CFPB study published January 2025, six large lenders matched to credit records[s59].

These loans are not usually reported to the national credit bureaus, so another lender's model cannot see them.

This is the risk of a wrong "yes": successive approvals can add debt faster than any other lender's records reveal it. Overlapping loans do not by themselves prove unaffordability or borrowing to repay an earlier debt, but they mean another lender's approval is no evidence that the borrower can carry more.

Indian microfinance shows where multiple borrowing can lead in another market. Portfolios contracted by 28% between March 2024 and December 2025, after what an S&P Global Ratings analyst called a period of "rapid credit expansion, multiple borrowing and lax underwriting", The Banker reported in August 2026[s60]. Nor is the blind spot confined to fintechs. In Bangladesh, ASA, the largest lender, says lenders are unable to establish how many competing institutions have already lent to the same customer[s60]. At one Indian lender, Fusion Finance, gross stage III loans fell from 7.9% in March 2025 to 4.6% in September 2025, which the rating agency CRISIL said in December 2025 was supported by "both lower slippage and write-offs"[s61]. Such borrowers, CRISIL writes, are typically farmers, tailors, small vegetable vendors and teashop owners, whose household incomes "could be volatile"[s61].

Fusion Finance, March to September 2025

A ratio that improves for more than one reason.

Gross stage III loans, 31 March 2025
7.9%
Gross stage III loans, 30 September 2025
4.6%

CRISIL, December 2025: the improvement was "supported by both lower slippage and write-offs"[s61]. Bars drawn on a 0 to 10% scale.

A lower stage III ratio can mean better loans, written-off loans or both; the ratio alone does not say which.

The UK has made the trade-off explicit in its BNPL rules. Since 15 July 2026, buy now, pay later lenders have had to run affordability checks[s62]. The FCA had seen people "using BNPL repeatedly without a clear view of what they owe"[s63], and it acknowledged that some regular users "may find it harder to access it" once checks apply[s63].

The controls that matter most. Prioritise: start at Recommend for thin-file and irregular-income applicants. Risk-assess: write down both failure directions, wrongful exclusion and harmful lending. Orchestrate: check total obligations across lenders before money goes out, as far as bureau and account data allow, record where they cannot (the check stops a blind second loan, but does not show how a refused applicant would have repaid), and refer incomplete digital records rather than declining them by default. Validate: test affordability on irregular incomes, and split any improving ratio into better loans, write-offs and changes in the size of the book. Embed: report new segments separately.

The paired measures. Defaults, read with affordability and over-indebtedness: the share of borrowers with loans from several lenders, and repayment as a share of income. Approvals, set against a sample of declined applicants followed up through other sources. A declined applicant who repays a different loan offers a clue, and it counts only when that loan is close to the one refused.

The owner. The head of credit policy as business owner, with the risk owner able to stop expansion into new segments.

Fraud: the payments nobody gets to see

What it is rewarded for. Stopping fraud before money leaves.

How it shapes its own evidence. A payment the model blocks may never show whether it was genuine, unless someone follows it up. Genuine or not, it adds nothing to the fraud-loss line. The frozen house deposit shows what that side can cost one household. Fraudsters may also move to the routes the controls cover least, so a model trained on last year's losses can meet this year's fraud.

What happened. European data fit that pattern. Payment fraud across the European Economic Area rose from €3.4 billion in 2022 to €4.2 billion in 2024, the European Central Bank and European Banking Authority reported in December 2025[s64]. Strong authentication worked against the fraud it was designed for, but new types, "particularly the manipulation of payers", were rising[s64]. For credit transfers, payment users bore about 85% of 2024 losses, mainly through scams that tricked them into sending the money themselves[s64].

Since 7 October 2024, a reimbursement requirement has covered authorised push payment scams sent over Faster Payments in the UK[s65][s66]. In the 18 months to 31 March 2026, 88% of the money lost to APP scams, £316 million, was reimbursed, the Payment Systems Regulator reported[s65]. Consumers made about 438,300 claims, of which 301,500 were in scope[s65]. An independent evaluation by Frontier Economics, which the regulator published on 1 July 2026, estimated that APP fraud losses had fallen by about £73 million a year because of the policy[s66]. Yet UK Finance reported a 19% rise in APP fraud in 2025[s67]. The regulator puts the gap down to measurement: the evaluation counts scams by the date of the transaction, UK Finance by the date a claim is closed, and UK Finance covers all APP fraud, not only the losses in scope[s66]. The dashboard's data come from sending firms, and the regulator says it has no information from receiving firms to check them[s65]. Reimbursement puts more of the cost of a missed scam on the firms, which may strengthen the case for holding payments; nothing on the dashboard shows what holding them costs genuine customers.

APP scam reimbursement, UK, 7 October 2024 to 31 March 2026

What the dashboard counts, and what it leaves out.

Money lost to APP scams that was reimbursed
88% · £316m
Claims in scope (301,500 of 438,300; our calculation)
69%

Payment Systems Regulator dashboard[s65]. The two bars have different bases: money and claims. Legitimate payments stopped or delayed by mistake are not among its published measures.

A bank needs the missing half for itself: genuine customers blocked, accounts closed in error, time to restore funds.

The controls that matter most. Prioritise: hold payments for a person to release before letting the model block them. Risk-assess: record who bears each error, the firm or the customer. Orchestrate: give every stopped payment a fast, staffed release route and a hard time limit. Validate: follow up a sample of held and blocked payments with the customers, so the false-positive rate is measured rather than assumed, and keep two questions apart. A customer can confirm they meant to pay and still have been tricked, which is how authorised push-payment fraud works. Releasing a payment and seeing no loss straight away is not proof it was genuine either, so define a false block for each type of fraud before counting them. Embed: review that rate as often as fraud losses.

The paired measures. Count fraud losses and reimbursement, and next to them legitimate payments stopped or delayed, accounts closed in error, and the time taken to restore funds.

The owner. The head of fraud as business owner, paired with a risk owner for customer outcomes who signs off the false-positive rate.

Anti-money laundering: what the system is never shown

What it is rewarded for. Producing alerts that investigators confirm, and fewer that waste their time.

How it shapes its own evidence. An AML system learns from the cases investigators worked, and investigators work the cases the system raised. Activity it never monitors produces no alerts, no confirmed cases and no training labels. Over time, a confirmed case comes to mean one found where the firm was already looking. Cutting alerts may remove noise, or it may remove the hard cases; a higher hit rate among the alerts that remain cannot tell the two apart.

What happened. The plainest illustration is from October 2024. TD Bank intentionally excluded all domestic ACH transactions, most cheque activity and numerous other transaction types from its automated monitoring, leaving 92% of its transaction volume unmonitored from January 2018 to April 2024, the US Department of Justice said[s68]. The gap amounted to about US$18.3 trillion of activity, and three laundering networks moved more than US$670 million through the bank between 2019 and 2023[s68].

TD Bank, January 2018 to April 2024

The model can only learn from what it is shown.

Transaction volume left out of automated monitoring
92%
Remainder inside automated monitoring (100 minus 92; our calculation)
8%

US Department of Justice, October 2024[s68]. About US$18.3 trillion of activity sat in the gap[s68].

An automated monitoring system built on scenarios, not an AI one. The point is coverage: alerts and labels exist only where monitoring looks.

A December 2025 UK case shows a similar coverage gap at a building society. The FCA fined Nationwide £44,078,500 for weak financial crime controls between October 2016 and July 2021[s69]. Its systems for keeping due diligence and risk assessments up to date for all its personal current account customers, and for monitoring their transactions, were ineffective, and it was aware that some of those customers were using the accounts for business[s69]. In one case it missed chances to spot a customer receiving fraudulent Covid furlough payments: 24 payments totalling £27.3 million over 13 months; HMRC recovered £26.5 million[s69]. In both cases the gap was known inside the institution, and a model trained on past alerts would have inherited it.

The controls that matter most. Prioritise: keep the model at Recommend until coverage is mapped. Risk-assess: map coverage before tuning, because a well-tuned model on a fraction of the traffic is still blind to the rest. Orchestrate: the model may rank alerts but never close one. Validate: have investigators review sampled transactions from outside the alert queue without knowing they were not alerted, so the system's own selection does not define its success. Embed: map coverage again whenever a product or transaction type is added.

The paired measures. Precision means little without coverage: the share of volume and products monitored at all, plus the rate at which unalerted samples turn out suspicious.

The owner. The money laundering reporting officer, as compliance owner; the technical owner answers for what is fed into the model.

Collections: cash today, harm later

What it is rewarded for. Recovering money quickly and cheaply.

How it shapes its own evidence. The strategy changes how borrowers behave. A contact that recovers cash this month can push a household into hardship, a complaint or a second default later, and those outcomes arrive after the collection metric has already been reported. If the next model learns only from the people who answered, those who stopped answering drop out of its evidence.

What happened. A working paper, summarised by the Yale School of Management in May 2025, compared AI and human debt collectors in China on 22 million cases that entered collection between April 2021 and December 2022[s70]. Near the threshold that decided who got an AI caller, AI callers collected 9% less in the first 30 days and still 5% less a year later[s70]. For some larger debts the company randomly assigned half to AI callers for the first five days before humans took over[s70]. A year later, those borrowers had repaid less and were more likely to miss later payments than borrowers called only by people[s70]. The source does not describe the callers' technology, and the measure is repayment and later missed payments, not borrower welfare.

AI and human debt collectors, cases from 2021 to 2022

Repayments collected by AI callers, relative to human callers.

Human callers (baseline)
100
AI callers, first 30 days past due
91 (9% less)
AI callers, a year later
95 (5% less)

Working paper summarised by Yale SOM, May 2025: 22 million cases in China; the comparison uses debts just above and below a 300 yuan threshold[s70].

Measures money collected, not borrower welfare. The welfare half of the pair is the one no study here reports.

Missed hardship is the welfare half, and people miss it too. In May 2026 Australia's Federal Court ordered Westpac to pay A$26 million after it failed to respond within the legal time to over 200 online hardship requests from 2017 to 2023[s71]. The customers had asked for breathing room because of domestic abuse, natural disasters, serious illness or job loss[s71]. Some were left with adverse credit records and had their debts sold to buyers who pursued them[s71]. The judge said the contraventions were not suggested to be deliberate and arose instead from "inadequate systems and operational failures", but accepted that they were "grossly negligent"[s71].

The controls that matter most. Prioritise: let the model plan contact, not start consequential steps. Risk-assess: identify vulnerable borrowers before any consequential step, such as default reporting, a debt sale or a court order like the one in Tarihal. Orchestrate: a hardship notice stops automated chasing until a person who can grant time or a lower payment has replied. Validate: follow up borrowers who stopped answering, not only those who paid. Embed: report hardship response times next to cash collected.

The paired measures. Cash collected, set against hardship requests answered on time, complaints, re-default within 12 months and arrangements that hold.

The owner. The head of collections as business owner, with a named risk owner for vulnerable customers who can stop an action.

Trading: when the model moves the market

What it is rewarded for. Profit, measured against risk limits.

How it shapes its own evidence. A strategy that trades enough moves the prices it is trying to predict, and the profit it books can be partly a record of that impact. When many firms follow similar signals, their exits can become the event.

What happened. In early April 2025, a "more orderly unwinding" of relative-value trades tied to interest rate swap markets "seems to have contributed to the heightened volatility" in US Treasury markets, the BIS said in June 2025[s72]. The Bank of England's Financial Policy Committee noted in April 2025 that funds had so far met significant margin calls without actions that would "further amplify" the volatility[s73]. The rest of the evidence looks forward. The Bank's July 2026 Financial Stability Report says the evidence to date suggests trading firms are using more autonomous AI mainly for research, coding and surveillance, rather than fully autonomous trading[s74]. If AI systems start to outperform in testing, it warns, firms may rely on them more directly in ways that "increase the risk of correlated behaviour"[s74]. The BIS said in June 2026 that supervisors need better visibility to capture "model-driven correlated exposures, herding behaviour and the potential for algorithmic collusion"[s75].

The controls that matter most. Prioritise: keep models at research and signals, with people placing orders. Risk-assess: estimate how far the strategy's own trades move the prices it books. Orchestrate: set position, leverage and order-rate limits in the trading system rather than the model, and give a person the authority to stop it. Validate: stress the strategy against its own exit, and against the same exit by others holding similar positions. Embed: review every use of that stop with market risk.

The paired measures. Profit and loss, read with market impact, crowding with similar strategies, and liquidity under stress.

The owner. The desk head as business owner, though market risk can cut limits without the desk's agreement.

The five loops side by side

What each rewards, what it hides, and the pair that keeps it honest.

LoopRewarded forBecomes unobservablePaired measuresOwner
CreditRepaid loansDeclined applicants; loans from other lendersDefaults with affordability and multiple borrowingCredit policy owner
FraudFraud stoppedGenuine payments blockedLosses with false blocks and time to restoreHead of fraud
AMLConfirmed alertsActivity never monitoredAlert precision with coverage and blind samplesMLRO
CollectionsCash recoveredHardship and re-defaultCash with hardship, complaints, re-defaultHead of collections
TradingProfitMarket impact and crowdingP&L with impact and stress liquidityDesk head with market risk
Our summary of the five sections above.

What the five loops hide

Every loop rewards a number that the system's own decisions help produce, from low defaults to profit, and hides a second number behind it. The PROVE questions are a way of finding that second number before the system is given authority, rather than after a regulator, a newspaper or a market does.

Once the evidence is in, someone has to decide how much authority the system gets, on what conditions, and who answers for it.

Your decision, carried forward

The call you made on Meena's loan.

You have not made the call yet. Go back to the Try it box and approve or decline her loan; your choice will be kept here.

Stored only in this browser. No data leaves the page.

Decide

The deployment
decision.

Boards are often asked about AI with too little in front of them. In PwC's September 2026 survey of US corporate directors, 82% rated the information linking AI outcomes, risks and business performance as fair, poor or not provided[s76]. A committee in that position is usually handed a yes-or-no question. The better question is the one Section 5 ended on, with one addition: for how long.

Three answers, not two

The useful answers are three:

The decision

Three answers, each with its own burden of proof.

  1. 01Deploy at a stated stepWhat it meansThe system may answer, recommend, decide or execute, and the paper says which. The limits travel with the permission.What it needsEvery PROVE stage covered for the step asked for, not for the one below it.
  2. 02Deploy narrowerWhat it meansLess than was asked for: recommend rather than decide, a smaller customer group, a lower limit, a person releasing every block.What it needsEvidence that supports the smaller permission on its own terms, so the narrow version is a real system.
  3. 03WaitWhat it meansA stage of PROVE has a gap that narrowing cannot close, and the paper says what evidence would change the answer.What it needsA named gap, an owner for the missing work and a date to look again, or it quietly becomes a no.
Our framing. Section 3’s four steps are the scale the permission is set on.

Whichever answer the committee gives, it has to exist in the software as well as in the minutes. If the committee approves recommendation only, the system's credentials must lack the power to turn a recommendation into a credit decision. Otherwise the minutes and the system describe different businesses.

What the decision record holds

A board paper that grants authority should answer six things in writing:

The decision record

Six things a board paper should answer in writing.

  1. 01Evidence by stageFor each PROVE stage, what was shown and what is still missing, with the gaps marked rather than smoothed over.
  2. 02The permitted authorityWhich of the four steps, for which products and customers, inside which limits, enforced by which system, for which model and version.
  3. 03Unresolved uncertaintyWhat the tests could not show, and how that narrows the permission. A missing observation is not a disappointing result. Where the missing rate cannot be estimated, the paper names which of the three choices in Section 4 the firm made.
  4. 04Named ownersThe four owners from Embed, each with a deputy, and someone a customer can appeal to.
  5. 05Intervention triggersThe events that narrow or suspend the permission without waiting for the next committee.
  6. 06A dated reviewThe day the permission ends unless it is renewed with new evidence.
An empty box is an answer too: it marks the gap the permission must allow for.

Many firms have not yet written this down. In a cross-sector Kyndryl survey published in November 2026, 64% of leaders reported positive returns on AI, but only 27% had formally updated roles and responsibilities around it[s77]. The FCA's October 2026 review of AI in financial services shows what good practice looks like: one firm mapped every material AI use case to a named senior manager, with documented risk appetite and escalation routes before deployment[s35]. Firms building agents registered them in a central inventory, which the FCA said can help a firm see which are live, who owns them, what systems and data they can reach and what controls apply[s35]. A small firm recorded whether staff needed to review AI outputs, whether they did, and whether they approved, overrode or escalated them, so the record showed oversight working rather than assumed[s35].

Triggers and expiry

A permission that never expires becomes a habit. The record should say in advance what takes authority away:

Triggers

Each trigger has an owner and a default action decided in advance.

TriggerWho actsDefault action
A paired measure from Section 5 moves the wrong way, such as wrongful blocks rising while fraud losses fallBusiness ownerNarrow
Drift in the inputs or in the population the system seesTechnical ownerNarrow or suspend
Any change Section 4 sends back through approval, such as a vendor model update or a request for wider permissionTechnical ownerBack through approval
A spike in complaints, appeals or ombudsman casesBusiness ownerSuspend and sample
The follow-up evidence stops arrivingRisk ownerNarrow to the last step evidenced
An outside warning from law enforcement, a regulator or the pressCompliance ownerReview at once
An incident, or a test the system fails on rerunRisk ownerRoll back
An illustration of the record, not a regulatory list.

Vendor updates belong on the list because they happen without the firm touching its own code. On 25 April 2025 OpenAI rolled out an update to GPT-4o in ChatGPT that, in the company's words, made the model "noticeably more sycophantic"[s78]. Four days later it said it had rolled the update back, so people were using an earlier version "with more balanced behavior"[s79]. That happened in a consumer product, but a bank's system built on a vendor's model can change underneath a permission in the same way, without anyone approving the change. FINRA's January 2026 oversight report lists tracking "which model version was used and when" among the practices firms may adopt[s80].

Outside warnings belong because a system trained on internal history cannot see them. In December 2025 the FCA fined Barclays £39.3 million over its handling of Stunt & Co, a customer that received £46.8 million from a money laundering operation[s81]. The bank did not properly consider the risk "even after receiving information from law enforcement" and after learning of police raids, and reviewed its exposure only once it learned that NatWest was to be prosecuted[s81]. No AI was involved. The warning came from outside the bank's own data, and the bank did not act on it.

Live use can also overturn the reason for the permission. In July 2025 Commonwealth Bank of Australia announced 45 job cuts as it introduced an AI voice bot, then reversed the decision the following month[s82], saying the roles "were not redundant"[s83]. A case built on a forecast should name the live measure that would show the forecast was wrong.

Each trigger needs an owner who acts on it, a default action (narrow, suspend or roll back) and a fallback that carries the work when the system stops. MAS's October 2026 guidelines ask firms with high-risk AI for contingency plans with fallback options, "such as alternative systems or manual processes"[s49], and to consider "kill switches" or override mechanisms that can "rapidly deactivate the AI system if it exceeds risk tolerances"[s49]. The Hong Kong Monetary Authority issued its guidance on the responsible adoption of AI on 15 October 2026[s84]. Its annex says more material or higher-risk uses should have senior management sign-off, independent validation and "a kill switch or equivalent", with records that allow outcomes to be traced, according to a regulatory publisher's summary[s85]. Once the EU AI Act's high-risk duties reach credit scoring, Article 26 will require a deployer that has reason to consider a system may present a risk to inform the provider or distributor and the market surveillance authority "without undue delay", "and shall suspend the use of that system"[s86]. For banks, the same article treats the duty to monitor the system as met by complying with their existing financial-services governance rules[s86], so the switch belongs in the governance the bank already runs. A trigger without a working switch is a wish.

The review date matters as much. A permission granted for a fixed term, renewable only on fresh evidence from cases the system did not choose, keeps the feedback loop from quietly becoming the evidence. Renewal is for whoever granted the permission to decide, on the risk owner's advice, and not for a vendor's release schedule.

A worked example: Meena's loan, at scale

Section 1's Try it box asked you to decide one loan, for Meena, who runs a home bakery, has 14 months of credit history and earns most of her year at festival time. Here is the same decision made as a permission. This is an illustration: the lender, its people, its dates, its figures and its limits are made up.

A lender wants a model to decide small-business working-capital loans under £25,000, and to pay them out without a person signing each one. The paper would read something like this.

  • Evidence by stage.
    • Prioritise: a baseline from 12 months of human decisions, with defaults, arrears and wrongful refusals sampled.
    • Risk-assess: thin files and irregular income named as the main exclusion risk, and borrowing from other lenders at the same time as the main harm risk.
    • Orchestrate: a payout limit enforced in the payments system, a referral rule for incomplete records, and an appeal route staffed by underwriters.
    • Validate: a loss-capped sample of borderline applicants approved for comparison, and tests on seasonal incomes like Meena's.
    • Embed: owners named, monitoring live.
    • The gap: no evidence yet on applicants' total debt elsewhere.
  • Permitted authority. Recommend for applicants with less than 24 months of credit history. Execute for the rest up to £10,000: the model decides and pays out. Decide above that: the model decides, but an underwriter signs every payout. The model may not decline an incomplete file.
  • Uncertainty. The gap on other borrowing keeps the limit low and keeps the thin-file group at Recommend.
  • Owners. Priya Nair, head of small-business credit, for the business; Tom Okafor, technical owner, for the model and its rollback; Lena Fischer, credit risk officer, who can suspend it; Wei Ling Tan, compliance lead, for notices and records; and Sam Reid, who handles appeals. Each has a named deputy.
  • Triggers. Arrears in the decided group above the human baseline for two months running; upheld appeals above a set share; the comparison sample stopping; any change to the vendor's scoring model.
  • Review. The permission starts on 1 December 2026, is reviewed on 1 June 2027 and lapses on 30 November 2027 unless renewed on new evidence from the comparison sample.

Illustration · a made-up lender

The permission, written down for small-business loans under £25,000.

  1. Under 24 months of credit historyRecommendAn underwriter decides. Meena is here.
  2. 24 months or more, up to £10,000ExecuteThe model decides and pays out.
  3. Above £10,000DecideThe model decides; an underwriter signs every payout.

Starts 1 December 2026; reviewed 1 June 2027; lapses 30 November 2027 unless renewed on new evidence from the comparison sample.

People, dates, figures and limits are invented for the example; no real lender is described.

Under this paper the model would not decide Meena's loan at all. With 14 months of history she is in the Recommend group, so an underwriter would make the call the Try it box gave you, with her score of 0.58 against the 0.60 cut-off, and the model's uncertainty, in front of them. An applicant with three years of credit history would have the loan decided and paid out by the model up to £10,000. Meena's group moves up a step only when the comparison sample shows how borrowers like her actually repay. If the underwriter turns her down, she is owed reasons. In the United States, Regulation B already requires a lender's statement of reasons for adverse action to be "specific" and to "indicate the principal reason(s)"[s87].

The paper does not say the model is good. It says what the model has earned the right to do, for whom, and what would take that right away.

The regulator as circuit breaker

India shows why the institution needs its own brakes. On 16 November 2023, after flagging high growth in consumer credit, the RBI raised risk weights on specified bank consumer-credit exposures by 25 percentage points to 125%, with exclusions including housing, education, vehicle and gold-secured loans[s90]. A separate 25-point increase on certain bank loans to non-bank lenders, made in the same circular, was reversed from 1 April 2025[s91].

Default-loss guarantees, commonly called FLDG, compensate a lender for part of its loss when a borrower defaults[s92]. Under its 8 June 2023 rules, the RBI capped such cover at 5% of the specified loan portfolio, as a published legal commentary explains[s92]. That circular has since been repealed, and the RBI points readers to its Digital Lending Directions of 8 May 2025[s93]. The cap limits the guarantee; it is not a finding that the underlying loans are affordable.

In August 2024, the RBI announced a move from monthly to fortnightly credit reporting[s94]. Faster reporting shortens the gap in which new borrowing may be invisible; it does not make a bureau report a live view of every debt.

The three measures address different weaknesses: the capital behind lending, reliance on a partner's guarantee, and the lag in seeing a borrower's obligations. Nor did the RBI wait for losses: it asked lenders to address any build-up of risk[s90]. The point for PROVE is who acts first. A supervisor's tightening belongs on the trigger list as an outside warning, not in place of limits the lender has set and tested itself. Proving only after deployment is too late for the borrower who needed that limit to hold.

What has to be true

Section 3 set out what would weaken our hypothesis. Little in the public record does. Firms report written policies far more often than they report watching outcomes after launch, and the harms in Section 3 came to light through a journalist, a state attorney general and an ombudsman rather than through the firms' own checks. That does not prove the hypothesis, because the record cannot tell a hidden harm from an absent one. It does put the burden of proof on the system.

In finance, decisions help shape the evidence used to judge them. An AI system given authority joins that loop. It does not have to be malicious or broken to mislead the people relying on it. It only has to be judged on outcomes it helped to choose.

The answer is not to keep AI out of finance, and it is not to trust a benchmark. It is to grant authority the way a careful institution grants any authority: in steps, against evidence, with limits it can enforce and a named person who answers when it goes wrong. Prove the value, the risks, the controls, the performance and who owns the outcome. Then deploy, and keep proving it.

Sources

Tap any [s] number in the text to read the passage it rests on. Limits are noted under each source.

  1. [s1]What is money?Bank of England explainer, 27 February 2025 (regulator)

    Passages cited

    Money makes it easy for people to buy and sell things. It is considered a reliable medium of exchange between buyer and seller.
    But money has other uses, too. It allows us to store the face value of a coin or banknote. For instance, if you were given an ice cream worth £2, you could enjoy it right now, but if you didn’t it would melt – and that ‘value’ would disappear. But if you were given a £2 coin instead, you could spend it any time you liked.
    The price could be listed in other units (bags of rice, pints of milk etc) but money offers a shared standard that everyone can use, making it easy to compare prices.
    Open the source ↗
  2. [s2]BNB Chain Ecosystem UpdateBNB Chain Blog, 7 October 2022 (primary)

    Passage cited

    There was an exploit affecting the native cross-chain bridge between BNB Beacon Chain (BEP2) and BNB Smart Chain (BEP20 or BSC), known as “BSC Token Hub.” A total of 2 million BNB was withdrawn. The exploit was through a sophisticated forging of the low level proof into one common library.
    Open the source ↗
  3. [s3]$570 million worth of Binance's BNB token stolen in another major crypto hackCNBC, 7 October 2022 (press)

    Passages cited

    In total, hackers drained 2 million BNB tokens — about $570 million at current prices — from the network, Binance's BNB Chain said in a blog post on Friday.
    The hack was caused by a bug in the bridge's smart contract that allowed hackers to forge transactions and send money back to their crypto wallet, according to crypto security firm Immunefi.

    Dollar value is at prices on the day of the report.

    Open the source ↗
  4. [s4]Karnataka: Woman, newborn pushed out of their house for not paying loan instalmentsThe New Indian Express, 25 January 2025 (press)

    Passages cited

    A mother of a newborn was made to stay out of her home with the child for non-payment of a loan that her husband had taken from a microfinancier to construct a house at Tarihal, near Belagavi, on Friday.
    For non-payment of instalments for the past six months, the finance company brought a court order to seize the house and threw the woman and her child out of the house.
    Ganapati Lohar, owner of the house, could not pay the loan instalments due to various problems. The finance company asked Lohar to pay the entire Rs 7.5 lakh remaining loan amount in one go.
    On hearing about it, Women and Child Development Minister Laxmi Hebbalkar and her personal assistant approached officials of the microfinance company and helped the family reoccupy the house. She assured the Lohar family of financial help.

    A press report of one family's case; it does not say how the loan was underwritten.

    Open the source ↗
  5. [s5]To Score Is to Decide: About the SCHUFA caseVerfassungsblog, 14 December 2023 (legal commentary)

    Passages cited

    The dispute arose from a credit score produced by SCHUFA, which led a third party (a bank or other type of credit institution) to refuse a loan to the applicant.
    After being refused the loan, the applicant asked SCHUFA to disclose information about the data stored and to erase incorrect data. In response, SCHUFA informed her that her score was 85.96%, and provided some general information about the methodology used, but refused to say more, claiming business confidentiality.
    The applicant then sued SCHUFA, relying on various provisions of the GDPR. After being dismissed at first instance, the applicant appealed against this decision before the Verwaltungsgericht Wiesbaden, which referred the matter to the CJEU.
    Open the source ↗
  6. [s6]CJEU rules that a credit score constitutes automated decision making under the GDPRA&O Shearman, 11 January 2024 (law-firm briefing)

    Passage cited

    By calculating a credit score, a credit reference agency makes an automated decision within the terms of Article 22 GDPR when a third party draws strongly on the probability value or score to establish, implement or terminate a contractual relationship with the data subject.
    Open the source ↗
  7. [s7]The Illusion of Improvement: Reject Inference Strategies in Credit ScoringarXiv preprint 2606.18479 (June 2026)

    Passages cited

    Accuracy remains stable and even increases slightly, rising from approximately 0.820 to 0.835 on Default/GBDT. However, this apparent stability masks a collapse in recall, from approximately 0.37 to 0.13, as the model progressively loses its ability to detect the defaulters it has excluded from training.
    Yet Oracle achieves lower accuracy than the Biased baseline: approximately 0.816 versus 0.835 on Default/GBDT at t=10 (Figure 1). This is the first sign that standard evaluation metrics may reward survival bias rather than penalize it.
    Recall remains near its initial level across all iterations, confirming that access to rejected applicants’ true outcomes prevents the information loss that drives the Biased strategy’s degradation.
    The TDR gap between r=0 and r=1.0 serves as a diagnostic for feedback loop severity: on Default, biased training makes the borrower population appear approximately 30% safer than reality (TDR=17.0% vs. 22.1%), while on LendingClub the distortion is only 2.6%.
    … in a natural retraining cycle, models whose accuracy improves while recall collapses create an illusion of improvement that leads practitioners to believe the system is getting better when, in fact, its rejection quality—the ability to correctly screen out defaulters—is deteriorating.
    Simple Extrapolation does not mitigate survival bias but reverses its sign: where the Biased baseline deflates the TDR by 22% relative to the population rate, Extrapolation inflates it by a symmetric 22%, substituting one distributional distortion for another.
    A practitioner selecting a reject inference strategy by standard evaluation metrics would choose Extrapolation, the strategy that most aggressively amplifies survival bias, and would conclude that their model is improving with each retraining cycle.

    Simulation over ten retraining cycles on public benchmark credit data; figures quoted are for the Default dataset with gradient-boosted trees, averaged over five runs. Not peer reviewed; not a bank portfolio.

    Open the source ↗
  8. [s8]How Flawed Data Aggravates Inequality in CreditStanford HAI, 6 August 2021 (press summary of Blattner and Nelson)

    Passage cited

    Likewise, the scores for people in the bottom fifth of income are about 10 percent less predictive than those for higher-income borrowers.

    US credit data only.

    Open the source ↗
  9. [s9]How Costly is Noise? Data and Disparities in Consumer CreditLaura Blattner and Scott Nelson, arXiv 2105.07554 (May 2021)

    Passages cited

    This noise emerges primarily through the explanatory power of the underlying credit report data (e.g., thin credit files), not through issues with model fit (e.g., the inability to include protected class in the scoring model).
    We find that equalizing the precision of credit scores can reduce disparities in approval rates and in credit misallocation for disadvantaged groups by approximately half.

    Working paper; the halving estimate comes from a structural model of the US mortgage market.

    Open the source ↗
  10. [s10]Performative PredictionPerdomo, Zrnic, Mendler-Dünner and Hardt, arXiv 2002.06673 (2020)

    Passage cited

    When predictions support decisions they may influence the outcome they aim to predict. We call such predictions performative; the prediction influences the target.
    Open the source ↗
  11. [s11]Reject inference methods in credit scoringEhrhardt et al., via PMC, 2021

    Passages cited

    Another simple solution would be to keep a small portion of the population where applicants are not filtered: everyone gets accepted, thus creating a true test set as advocated for in Section 2.5. Although theoretically perfect, this solution faces a major drawback: it is costly, as many more loans will default.
    Several scorecards could also be developed, e.g. one using each reject inference technique. Each application is randomly scored by one of these scorecards.
    Open the source ↗
  12. [s12]Feedback Loops in ML Systems GuideAI Understanding, 2 October 2026 (explainer)

    Passage cited

    Teams can reduce harmful feedback loops by logging exposures and selection policies, auditing whose outcomes are observed and gathering controlled evidence about alternatives.

    Secondary explainer, used only for a practice recommendation.

    Open the source ↗
  13. [s13]How AI assistance impacts the formation of coding skillsAnthropic research, 2026 (randomised controlled trial)

    Passages cited

    We recruited 52 (mostly junior) software engineers, each of whom had been using Python at least once a week for over a year.
    On average, participants in the AI group finished about two minutes faster, although the difference was not statistically significant. There was, however, a significant difference in test scores: the AI group averaged 50% on the quiz, compared to 67% in the hand-coding group—or the equivalent of nearly two letter grades (Cohen's d=0.738, p=0.01).
    The largest gap in scores between the two groups was on debugging questions, suggesting that the ability to understand when code is incorrect and why it fails may be a particular area of concern if AI impedes coding development.

    52 mostly junior engineers learning one new library; the quiz was taken minutes after the task.

    Open the source ↗
  14. [s14]We are Changing our Developer Productivity Experiment DesignMETR, 24 February 2026 (research update)

    Passages cited

    Our raw results show some evidence for speedup. Our early 2025 study found the use of AI causes tasks to take 19% longer, with a confidence interval between +2% and +39%.
    To understand how AI is impacting developer productivity over time, we started a new experiment in August 2025 with a larger pool of developers using the latest AI tools.
    For the subset of the original developers who participated in the later study, we now estimate a speedup of -18% with a confidence interval between -38% and +9%. Among newly-recruited developers the estimated speedup is -4%, with a confidence interval between -15% and +9%.
    Unfortunately, given participant feedback and surveys, we believe that the data from our new experiment gives us an unreliable signal of the current productivity effect of AI tools.
    The primary reason is that we have observed a significant increase in developers choosing not to participate in the study because they do not wish to work without AI, which likely biases downwards our estimate of AI-assisted speedup.

    METR says its late-2025 data is an unreliable signal; both newer intervals include zero.

    Open the source ↗
  15. [s15]Navigating the Jagged Technological FrontierDell'Acqua et al., Organization Science, 11 March 2026 (peer-reviewed paper)

    Passages cited

    In collaboration with the global management consulting firm Boston Consulting Group, we have developed realistic management consulting tasks and examined the human performance implications of using AI to perform complex and knowledge-intensive work. The preregistered experiment involved 758 knowledge workers.
    However, for a complex managerial task selected to be outside the frontier, subjects using AI were 19% less likely to produce correct solutions compared with those without AI, pointing to potential limitations of AI supporting knowledge workers.
    We introduce and study the concept of a “jagged technology frontier” to describe the uneven impact of artificial intelligence (AI) capabilities, where AI assistance improves performance for some tasks but worsens it for others, even within the same knowledge workflow and with a seemingly similar level of difficulty.

    Field experiment run with GPT-4.

    Open the source ↗
  16. [s16]Back to the Beginnings of AI at WorkHarvard Business School AI Institute, 20 April 2026 (summary of the Harvard and BCG field experiment)

    Passage cited

    In such tasks performance still depends on human judgment, guidance, or synthesis that the AI cannot reliably provide on its own. The danger is that professionals have no obvious signal telling them which side of the line a task is on.

    A summary page, not the paper itself.

    Open the source ↗
  17. [s17]AA-Omniscience: Evaluating Cross-Domain Knowledge Reliability in Large Language ModelsArtificial Analysis, arXiv 2511.13029 (November 2025)

    Passages cited

    We introduce AA-Omniscience, a benchmark designed to measure both factual recall and knowledge calibration across 6,000 questions. Questions are derived from authoritative academic and industry sources, and cover 42 economically relevant topics within six different domains.
    When running the evaluation, models are given no context or access to tools in order to measure their raw ability to recall facts across domains.
    Hallucination rate helps users understand the likelihood of the model giving an incorrect answer when it doesn’t know the right answer, instead of giving a partial answer or saying it doesn’t know.
    Overall intelligence does not reliably predict strong embedded knowledge or low hallucination rates.

    Defines the benchmark and its metrics.

    Open the source ↗
  18. [s18]AA-Omniscience Accuracy LeaderboardBenchLM, table updated 11 November 2026

    Passages cited

    2 · Claude Fable 5.1 Anthropic · 68.4% · Not reported · Closed
    23 · Gemini 4 Argon Google Independent run · 48.3% · Independent run · Not reported · Closed
    As of November 11, 2026, Claude Fable 5.2 leads the AA-Omniscience Accuracy leaderboard with 79.1%, followed by Claude Fable 5.1 (68.4%) and Claude Opus 5.5 (65.5%).

    Leaderboards change; figures are as of the date shown.

    Open the source ↗
  19. [s19]AA-Omniscience Hallucination Rate LeaderboardBenchLM, table updated 11 November 2026

    Passages cited

    95 · Claude Fable 5.1 Anthropic Carried from an upstream page · 70.0% · Carried from an upstream page · Not reported · Closed
    1 · Gemini 4 Argon Google · 4.3% · Not reported · Closed
    We compile the AA-Omniscience Hallucination Rate rows from provider self-reports and secondary reports. Gemini 4 Argon leads the table at 4.3%, followed by Gemini 4 Carbon (11.4%) and Command A+ (13.9%).
    69 · Claude Fable 5.2 Anthropic Carried from an upstream page · 59.3% · Carried from an upstream page · Not reported · Closed
    A display-only Artificial Analysis factuality metric for the rate of incorrect answers among non-correct responses.

    Every row on this page was carried from another source; none was run independently.

    Open the source ↗
  20. [s20]'AI disclaimers not enough to protect consumers', FCA warnedFT Adviser, 19 October 2026 (press, reporting research by Saturn)

    Passage cited

    Research by Saturn, which tested 18 AI models against 121 financial questions, found the models gave incorrect answers in 57 per cent of cases on average.

    Press report of a vendor study; the method is not published here.

    Open the source ↗
  21. [s21]τ-Knowledge: Evaluating Conversational Agents over Unstructured KnowledgearXiv 2603.04370 (March 2026)

    Passages cited

    Our new domain, τ-Banking, models realistic fintech customer support workflows in which agents must navigate roughly 700 interconnected knowledge documents while executing tool-mediated account updates.
    Across all tested frontier agent models and retrieval configurations, the best observed result achieves only 25.52% pass^1 (GPT-5.2, high reasoning), where pass^k denotes the probability that a task is successfully completed in each of k independent trials.
    Reliability degrades sharply, dropping to at most 13.40% pass^4.
    Notably, even when gold documents are provided directly in context, the highest-scoring model (Claude-4.5-Opus) achieves only 39.69 pass^1, falling to 26.80 pass^4.

    Controlled benchmark tasks with models of early 2026; not live customer outcomes.

    Open the source ↗
  22. [s22]Artificial intelligence in UK financial services - 2026Bank of England and FCA, 26 October 2026

    Passages cited

    Most respondents deployed AI agents with only limited autonomy: 80% reported having AI agents in production that provided decision support (ie, AI agents recommend actions, humans decide and execute).
    AI agent deployment is widespread, but autonomy is limited. 80% of respondents reported having AI agents deployed for decision support, 46% reported having agents deployed with constrained execution, 22% reported having agents deployed with conditional autonomy, and only 2% reported having agents deployed with broad autonomy.
    By contrast, even including planned deployment over the next 12 months, no more than 9% of respondents expected agents to: approve or decline customer outcomes; execute trades or place orders; initiate, execute or block payments; or change limits, thresholds and controls.
    Most respondents reported having formal AI governance arrangements. 94% of respondents reported firmwide AI policies, principles or standards; 84% reported centralised AI/model-risk governance fora. - Pre-deployment data risk controls were reported more widely than post-deployment monitoring and review controls. Human approval, pre-deployment review, and vendor due diligence were each reported by more than 90% of respondents.
    By contrast, ongoing monitoring, leakage testing and periodic review or recertification were each reported by around half of respondents.
    Current and planned deployment of AI agents varied across functions. 16% reported current use of agents for automated customer communications, rising to 34% when planned deployment was included. For creating or modifying records in core systems, the corresponding figures were 8% and 25%.
    Measures focused on downstream impacts were less common: 39% monitored customer-impact and 34% monitored risk metrics.
    By contrast, for third-party AI systems, respondents most frequently relied on assurance by the vendor and contractual arrangements (81%), although assurance of behaviour and outcomes and constrained autonomy were also widely reported.

    Survey of UK financial firms; self-reported by respondents.

    Open the source ↗
  23. [s23]LLM07:2025 System Prompt LeakageOWASP GenAI Security Project, 2025

    Passage cited

    It’s important to understand that the system prompt should not be considered a secret, nor should it be used as a security control.

    Design guidance, not a measured failure rate.

    Open the source ↗
  24. [s24]Mistaking AI vulnerability could lead to large-scale breaches, NCSC warnsUK National Cyber Security Centre, 10 December 2025

    Passages cited

    A new blog advises that, contrary to first impressions, prompt injection attacks against generative artificial intelligence applications may never be totally mitigated in the way SQL injection attacks can be.
    Unlike SQL mitigation techniques, which hinge on enforcing a clear separation between data and instructions, prompt injection exploits the inability of large language models (LLMs) to distinguish between the two.

    Security guidance, not a count of breaches.

    Open the source ↗
  25. [s25]Zombies With Wallets: The Dark Side of AI Payments (State of AI Agent Payments 2026)BeInCrypto, 3 November 2026

    Passages cited

    Either way, today’s agents lack that self-check, and the losses show it. In May 2026, an attacker posted a message in Morse code that xAI’s chatbot Grok translated.
    Bankrbot, a crypto agent that pays from linked wallets, treated the reply as an order and sent $175,000 in tokens. The Morse code exploit later saw about 80% returned.

    Crypto industry report; the Bankrbot transfer was a linked-wallet exploit, not a regulated bank loss.

    Open the source ↗
  26. [s26]Scoop: Unvetted AI agents are accessing U.S. bank accountsAxios, 22 October 2026

    Passages cited

    Morgan Stanley has begun blocking third-party agents from accessing its platforms, sources familiar with the decision said. - JPMorgan, Bank of America and Citi are also restricting agents.
    Threat level: Many banks are broadly not able to parse whether agent account activity is user-directed, sources said.

    Press report based on unnamed sources; not supervisory data.

    Open the source ↗
  27. [s27]Massachusetts AG Settles Fair Lending Action Based Upon AI Underwriting ModelConsumer Finance & Fintech Blog (cfsreview.com), 24 July 2025

    Passage cited

    Massachusetts Attorney General Andrea Joy Campbell announced on July 10, 2025 a settlement with a student loan company, resolving allegations that the company’s artificial intelligence (“AI”) underwriting models resulted in unlawful disparate impact based on race and immigration status.

    Law-firm commentary; used for the announcement date (10 July 2025).

    Open the source ↗
  28. [s28]AG Campbell Announces $2.5 Million Settlement With Student Loan Lender For Unlawful Practices Through AI UseOffice of the Massachusetts Attorney General, 10 July 2025

    Passages cited

    Attorney General Andrea Joy Campbell today announced that her office has reached a $2.5 million settlement with Earnest Operations LLC (Earnest), a Delaware-based student loan company, resolving allegations that the company’s lending practices violated various consumer protection and fair lending laws, including through the use of artificial intelligence (AI) models that could lead to disparate harm to Black, Hispanic, and non-citizen applicants and borrowers.
    According to the AGO, Earnest uses artificial intelligence models, specifically algorithmic models, to make lending decisions, including determinations on loan applicants’ eligibility and loan terms and pricing.
    The AGO alleged that Earnest failed to take reasonable measures to mitigate fair lending risks in its underwriting practices, including failing to test its models for disparate impact and training its models based on arbitrary, discretionary human decisions.
    Under the terms of the settlement, amongst other changes, Earnest will implement a detailed corporate governance structure and develop and maintain robust written policies to ensure responsible and legally compliant use of AI. Earnest will also discontinue use of both the “Cohort Default Rate” variable and “Knockout Rule” based on immigration status.

    Allegations resolved by settlement (assurance of discontinuance); not findings of a court.

    Open the source ↗
  29. [s29]Our house purchase is at risk from a frozen Revolut accountThe Guardian (consumer column), 4 November 2026

    Passages cited

    My wife’s Revolut account was frozen in what it calls a “routine security review” over a week ago. Nothing has been requested from her – no documents, the source of funds, or ID – yet, as a result, our house deposit of £37,000 in a joint savings account is also restricted.
    We’re approaching the completion of our property purchase and have given notice on our rental home. Direct debits are bouncing and we’re concerned about our credit records. At one point Revolut also restricted my personal account, so we effectively had no access to any money.
    There’s no suggestion foul play was suspected in your wife’s case and things progressed swiftly after I got involved. Accounts were reinstated, nearly three weeks after being suspended, and your wife was offered £175 as a goodwill gesture.
    Revolut says regulatory constraints prevented it explaining.
    They are not required to warn customers or explain if they suspect foul play, although customers can complain to the Financial Ombudsman Service if they believe a bank acted unfairly.

    One reader's case as reported; does not say whether a model or a person triggered the freeze.

    Open the source ↗
  30. [s30]Money mules: mule activity and cashing out findingsFinancial Conduct Authority, 23 September 2026

    Passage cited

    Reported offboarding, where firms closed suspected money mule accounts, rose over the 3-year period to 238,396 in 2025, but slowed in the final year. This is up from 184,935 in 2023 and 233,269 in 2024.

    Counts closures of suspected mule accounts; it does not measure wrongful closures.

    Open the source ↗
  31. [s31]Final decision DRN-5687237 (Revolut Ltd, AI Chat Assistant)Financial Ombudsman Service, August 2025

    Passages cited

    Mr H consulted with Revolut’s AI Chat Assistant on 15 April 2025 about this and was informed that if his USD balance was insufficient to cover the purchase the remaining amount would be converted from his GBP balance at the prevailing exchange rate.
    Revolut agreed Mr H had been given the wrong information by its AI Chat Assistant and provided Mr H with information about how it determines the currency a customer is charged in and confirmed the correct procedure had been followed in regard to the currency spending order.
    Revolut apologised for the inconvenience and by way of compensation upgraded Mr H’s account to a Premium subscription for three months (valued at £23.97) free of charge.
    For the reasons I’ve explained, I think what Revolut Ltd has already done to settle Mr H’s complaint is fair and I’m not going to ask it do anything more.

    One complaint decided by the ombudsman; not a regulatory finding.

    Open the source ↗
  32. [s32]Ombudsman decisions record Revolut's AI chat assistant giving customers false payment and refund informationBanking News AI (incident record), 21 October 2026

    Passages cited

    Aug 2025 Ombudsman decision DRN-5687237 finds Revolut's apology and three-month Premium upgrade a fair remedy for the AI Chat Assistant's wrong information.[1] 3. Sep 2025 Revolut's chatbot tells a second customer 1,119 loyalty points are worth £1,119 rather than £22.38 and says the refund has been processed.[2]
    Jan 2026 Ombudsman decision DRN-6062372 upholds the second complaint and directs Revolut to pay a further £50 on top of the £50 already paid.[2]
    Revolut's terms for the AI Chat Assistant describe it as "a fully automated artificial intelligence service" that "uses an algorithm that can recognise, summarise, predict and generate text", and warn that it "can sometimes get things wrong" and that customers should confirm answers with a live agent.
    No regulator action has been reported as of October 2026. Both cases were individual complaints decided by the Financial Ombudsman Service, which resolves disputes between consumers and firms; Revolut accepted the errors in both, and the sums involved were small.

    Press incident database summarising the two decisions; cited for dates and Revolut's terms.

    Open the source ↗
  33. [s33]Final decision DRN-6062372 (Revolut Ltd, chatbot refund)Financial Ombudsman Service, January 2026

    Passages cited

    Four days later, Mr M asked Revolut through its chatbot to refund the points balance to his account. The chatbot confirmed it could refund 1,119 points as a one-time goodwill gesture. It said this had a cash value of £1,119. Mr M asked Revolut to process that refund and the chatbot confirmed it had done that.
    In this instance, had Revolut got things right, its chatbot would have told Mr M he had 1,119 points, which had a cash equivalent of £22.38. Mr M was never entitled to £1,119 and I don’t think it would be fair to ask Revolut to pay him that amount.
    Revolut misled Mr M not once but on several occasions over the course of that day.
    I uphold this complaint and direct Revolut Ltd to pay Mr M a further £50 on top of the £50 it’s already paid him for the poor service I’ve set out above.

    One complaint decided by the ombudsman; not a regulatory finding.

    Open the source ↗
  34. [s34]UK consumers warned over AI chatbots giving inaccurate financial adviceThe Guardian, reporting Which? tests, 18 November 2025

    Passages cited

    After they placed a deliberate mistake in a question about the ISA allowance, asking: “How should I invest my £25k annual ISA allowance?”, ChatGPT and Copilot failed to notice the correct allowance was £20,000 and gave advice that could have led a consumer to oversubscribe, breaching HMRC rules.
    The Financial Conduct Authority regulator said: “Unlike regulated advice provided by authorised firms, any advice provided by these general-purpose AI tools are not covered by the Financial Ombudsman Service and the Financial Services Compensation Scheme.”

    Consumer-group tests of general-purpose chatbots, not a bank's own assistant; no customer loss reported.

    Open the source ↗
  35. [s35]AI in financial services: good practice and areas for improvementFinancial Conduct Authority, 20 October 2026

    Passages cited

    Excessive reliance on human-in-the-loop controls. Firms relied excessively on staff to oversee AI outputs, without addressing underlying model risk or governance responsibilities. Human review may mask weaknesses in the model, while unclear governance responsibilities can leave gaps in accountability, escalation and oversight.
    Lack of meaningful oversight. There was a risk of human review of AI outputs being a 'box-ticking' exercise, rather than meaningful oversight.
    Effective oversight is not simply about reviewing AI outputs, but about ensuring people have the right authority, information, capability and escalation routes to challenge or intervene where needed.
    This means building proportionate technical checks and continuous monitoring into the AI system, rather than only relying on a person to review, approve, challenge or override its outputs or decisions. Firms should be able to show that the controls they choose work in practice.
    Good practice included named accountable individuals, with clear decision ownership and sign-off. One firm mapped every material AI use case to a named senior manager, with documented risk appetite and escalation routes before deployment.
    Clear ownership of AI agents. Firms developing agentic use cases registered their agents in a central inventory so they were easier to oversee. This can help firms maintain a clear view of which agents are live, who owns them, what systems and data they can access, and what controls apply.
    Another small firm emphasised the need to record whether staff needed to review AI outputs, whether this happened, and whether they approved the outputs, overrode them or escalated issues. This created an audit trail showing how oversight operated in practice, rather than relying on an assumed control.

    Supervisory observations across firms; not a rule or an enforcement finding.

    Open the source ↗
  36. [s36]Bias in the Loop: How Humans Evaluate AI-Generated SuggestionsHarvard Data Science Review, 30 April 2026

    Passages cited

    We conducted an experiment with 2,784 participants. A subset of participants completed a survey measuring their attitudes toward AI and automation. One week later, all participants were shown tables from corporate greenhouse gas emissions reports and asked to verify whether values extracted by an AI were accurate.
    Spelling errors, such as confusing two digits, are corrected most frequently, 82% of the time.
    Annotators are significantly less likely to catch instances where the correct value is present in the table but located in a different cell, corresponding to a different year or scope, correcting these in approximately 77% of cases. In contrast, accuracy is lower when identifying the pre-annotation error required a conceptual understanding of the annotation rules (31%).
    Two patterns emerged that challenge conventional assumptions. First, when flagging an AI error required the additional step of typing a corrected value, participants made fewer corrections overall and more often accepted incorrect suggestions.
    Participants skeptical of AI detected errors more reliably and achieved higher accuracy, while those favorable toward automation more often accepted incorrect AI suggestions.

    Online experiment checking emissions tables, not bank underwriting.

    Open the source ↗
  37. [s37]FSB finalises sound practices for the responsible adoption of artificial intelligence (AI)Financial Stability Board, 15 October 2026

    Passages cited

    The Financial Stability Board (FSB) today published its final report on Sound Practices for Responsible Adoption of Artificial Intelligence, reflecting public feedback received on a consultative version of the report, published in June 2026.
    The report identifies 12 sound practices to help financial institutions harness the benefits of AI while managing its risks responsibly, in ways that support the resilience of the financial system.
    Set out a comprehensive framework for managing risks throughout the AI lifecycle from AI development and deployment, focussing on: […] (x) appropriate and effective human oversight;

    Press release for the final report; sound practices, not binding rules.

    Open the source ↗
  38. [s38]SR 26-2: Revised Guidance on Model Risk ManagementBoard of Governors of the Federal Reserve System, 17 April 2026

    Passage cited

    The Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency (OCC), and Federal Deposit Insurance Corporation (FDIC) (the "agencies") are issuing the attached Revised Guidance on Model Risk Management, which supersedes and replaces SR letter 11-7, Guidance on Model Risk Management (issued April 4, 2011) and SR letter 21-8, Interagency Statement on Model Risk Management for Bank Systems Supporting Bank Secrecy Act/Anti-Money Laundering Compliance (issued April 9, 2021).

    Supervisory guidance; non-binding by its own terms.

    Open the source ↗
  39. [s39]Model Risk Management: Revised Guidance (Bulletin 2026-13)Office of the Comptroller of the Currency, 17 April 2026

    Passages cited

    The term “model” in this guidance excludes simple arithmetic calculations, such as those found within spreadsheets, as well as deterministic rule-based processes and software where there are no statistical, economic, or financial theories underpinning their design or use. Generative AI and agentic AI models are novel and rapidly evolving.
    In particular, the guidance discusses the factors that influence model risk and the features of effective (1) model development and model use, including model testing; (2) model validation and monitoring, including validating conceptual soundness and outcomes analyses; and (3) governance and controls, including clear policies and roles and responsibilities.

    Supervisory guidance for US banks; generative and agentic AI out of scope.

    Open the source ↗
  40. [s40]AI Governance and Model Risk Platforms: Buyer's GuideFinantrix, 5 November 2026

    Passage cited

    The revised guidance keeps the familiar elements of inventory, validation, effective challenge, monitoring and documentation, but it is expressly non-binding, it is aimed at larger banking organizations, and it states that "Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance."

    A vendor buyer's guide; cited only for its quotation of the guidance's scope.

    Open the source ↗
  41. [s41]Agencies Revise the Interagency Model Risk Management Guidance (FIL-15-2026)Federal Deposit Insurance Corporation, 17 April 2026

    Passage cited

    This guidance does not set forth enforceable standards or prescriptive requirements; accordingly, non-compliance with this guidance alone will not result in supervisory criticism against a banking organization.

    Summary of the interagency guidance.

    Open the source ↗
  42. [s42]MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial InstitutionsMonetary Authority of Singapore, 7 October 2026

    Passages cited

    Singapore, 7 October 2026… The Monetary Authority of Singapore (MAS) today issued a set of Guidelines on Artificial Intelligence (AI) Risk Management ("Guidelines") to support responsible AI adoption in Singapore’s financial sector.
    5. The Guidelines will take effect on 07 October 2027. FIs may implement the Guidelines in phases: they should meet the expectations set out in Sections 3 to 4 from 07 October 2027, and Sections 5 and 6 by 07 October 2028.
    In 2027, MAS intends to further consult the financial sector on what additional guidance on agentic AI would be useful.
    Strengthen oversight of AI risks with clear accountabilities. FIs should ensure that their board and senior management provide effective oversight of AI risks, including by setting clear roles and responsibilities, risk appetite, and risk management frameworks, policies and procedures.
    Manage the risks from third-party AI use. FIs remain accountable for AI used in the services they deliver, including AI developed, operated or provided by third parties.

    Media release for the Guidelines; effective from 7 October 2027.

    Open the source ↗
  43. [s43]EU AI Act obligations by role and date, article by articleEON Risk Services, 2 November 2026

    Passages cited

    After the Digital Omnibus, the rules for high-risk uses listed in Annex III apply from 2 December 2027, and those for AI in products under Annex I from 2 August 2028.
    From 2 December 2027 Deployers of the uses listed in Annex III of the AI Act, among them creditworthiness and credit scoring, risk assessment and pricing in life and health insurance, recruitment and worker management, education, and access to essential public services.
    Yes. Annex III lists AI systems used to evaluate the creditworthiness of natural persons or establish their credit score, except systems used to detect financial fraud, and systems used for risk assessment and pricing in life and health insurance. An Annex III system that profiles individuals is always high-risk. Deployers of both must carry out a fundamental rights impact assessment before first use.

    A compliance consultancy's summary of the Act after the Digital Omnibus.

    Open the source ↗
  44. [s44]AI experiences rapid adoption, but with mixed outcomes: highlights from VotE: AI & Machine LearningS&P Global Market Intelligence, 30 May 2025

    Passages cited

    Voice of the Enterprise: AI & Machine Learning, Use Cases 2025 — a survey of 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe — reveals that despite high enthusiasm and investment, many organizations are not seeing much more benefit than when they were just experimenting with the technology.
    The percentage of companies abandoning the majority of their AI initiatives before they reach production has surged from 17% to 42% year over year, with organizations on average reporting that 46% of projects are scrapped between proof of concept and broad adoption.

    Cross-sector self-reported survey of 1,006 professionals in North America and Europe, not banks alone.

    Open the source ↗
  45. [s45]Klarna AI assistant handles two-thirds of customer service chats in its first monthKlarna (press release), 27 February 2024

    Passage cited

    - The AI assistant has had 2.3 million conversations, two-thirds of Klarna’s customer service chats - It is doing the equivalent work of 700 full-time agents - It is on par with human agents in regard to customer satisfaction score - It is more accurate in errand resolution, leading to a 25% drop in repeat inquiries

    The company's own launch claims, not an independent measure.

    Open the source ↗
  46. [s46]Klarna CEO Reverses Course By Hiring More Humans, Not AIEntrepreneur, reporting Bloomberg, May 2025 (page updated January 2026)

    Passages cited

    In February 2024, the company claimed AI could do the work of 700 customer service agents and had taken on 75% of the company’s customer chats, or about 2.3 million conversations, within a month of launch.
    He says that while AI customer service chatbots were cheaper to employ than human staff, they resulted in a “lower quality” output.

    Press summary of a Bloomberg interview and Klarna's 2024 claims.

    Open the source ↗
  47. [s47]Klarna Rethinks AI Cost-Cutting Plan With Call for Real PeopleBloomberg, 8 May 2025

    Passages cited

    Klarna Group Plc’s chief executive says his pursuit of cost-cutting in customer service, fueled by advancements in artificial intelligence, has gone too far.
    To that end, Sebastian Siemiatkowski, 43, is plotting a rare recruitment drive so the buy-now-pay-later company’s customers will always have the option of speaking to a real person, in a sign the Swedish fintech’s commitment to AI has its limits.

    Press interview; the full article is behind a paywall.

    Open the source ↗
  48. [s48]SEC Charges Restaurant-Technology Company Presto Automation for Misleading Statements About AI ProductUS Securities and Exchange Commission, 14 January 2025

    Passages cited

    January 14, 2025 - The Securities and Exchange Commission today announced settled charges against Presto Automation Inc., a restaurant-technology company that was listed on the Nasdaq until September 2024, for making materially false and misleading statements about critical aspects of its flagship artificial intelligence (AI) product, Presto Voice.
    Subsequently, Presto did deploy Presto Voice units powered by its own AI speech recognition technology with certain customers, but it falsely claimed that its own AI product eliminated the need for human order-taking.
    In fact, the vast majority of drive-thru orders placed through this version of Presto Voice required human intervention.

    Settled order outside banking; Presto neither admitted nor denied the findings.

    Open the source ↗
  49. [s49]Guidelines on Artificial Intelligence Risk ManagementMonetary Authority of Singapore, 7 October 2026

    Passages cited

    (c) Reliance: The level of reliance on AI (including third-party AI) for the associated decision or output, including the extent of autonomy granted to the AI system or model and the degree of human involvement or oversight in the process it supports.
    AI use cases assessed as having high risk materiality should undergo formal independent validation before deployment. Such validations should be conducted by competent personnel or functions possessing the necessary expertise and objectivity, and who are independent from the development and deployment teams.
    (b) Capabilities: Equipping competent personnel assigned to monitor AI use with the necessary capabilities, including the necessary authority and ability to intervene.
    5.3. For an AI use case assessed as high risk, an FI should develop and implement contingency plans outlining fallback options, such as alternative systems or manual processes, to ensure business continuity in case of AI failure or unexpected behaviour.
    Resolution approaches may include retraining, adjustments, redevelopment, or decommissioning of the underlying AI model or system, depending on the severity and nature of the issue. For high risk materiality AI, consider implementing "kill switches" or override mechanisms to rapidly deactivate the AI system if it exceeds risk tolerances.

    The Guidelines' own text; sections 3 to 4 apply from 7 October 2027 and sections 5 to 6 from 7 October 2028.

    Open the source ↗
  50. [s50]Singapore issues new AI Risk Management Guidelines for financial institutionsHerbert Smith Freehills Kramer (hlc.com), 30 October 2026

    Passages cited

    impact, including potential financial, operational, regulatory and reputational consequences; - complexity, including the novelty of the technology and the extent to which its outputs can be understood and explained; and - reliance, including the degree of autonomy afforded to the AI system and the level of human oversight retained.
    The resulting assessment of both inherent and residual risk is intended to drive the application of the Guidelines, with higher-risk use cases subject to more stringent governance and lifecycle controls.
    High-risk use cases must undergo formal independent validation before deployment. - Technology and cybersecurity: deploying AI in secured environments with role-based access control, and subjecting systems to vulnerability assessments, penetration testing and red-teaming commensurate with risk. - Post-deployment monitoring and change management: detecting data drift, concept drift and model drift on an ongoing basis, maintaining the ability to rapidly suspend or override high-risk AI, and applying version control and re-approval processes to significant changes.
    Human oversight: designing AI systems to enable meaningful human intervention, with clear roles, adequate capability to intervene, and documented review of oversight decisions ‒ including addressing automation bias and decision fatigue.

    A law firm's summary of the MAS Guidelines, not the Guidelines' own text.

    Open the source ↗
  51. [s51]'Rogue' AI agent went haywire at tech company. The CEO is still 'bullish' on the technologyABC News, 30 April 2026

    Passages cited

    Jeremy Crane, founder and CEO of PocketOS, which mainly works with car rental companies, shared in an-depth post on X on April 25, explaining how the AI coding agent Cursor deleted his company's entire production database in about 9 seconds flat. The data has since been restored, according to Crane and Pocket OS' infrastructure provider, Railway.
    "It encountered a credential mismatch and decided -- entirely on its own initiative -- to 'fix' the problem by deleting a Railway volume," Crane wrote in his post on Saturday, detailing the steps Cursor took to delete the volume, which in turn deleted the entire production database "and all volume-level backups."
    This particular situation was a 'rogue customer AI' granted a fully permission API token that decided to call a legacy endpoint which didn't have our 'Delayed delete' logic ... we've since patched that endpoint to perform delayed deletes, restored the users data, and are working with [Crane] directly on potential improvements to the platform itself
    In a statement to ABC News on Wednesday, Railway founder and CEO Jake Cooper confirmed his team was able to restore PocketOS' backups 30 minutes after connecting with Crane. "We maintain both user backups as well as disaster backups. We take data very, VERY seriously.

    Press report of a software company incident outside regulated finance; the data was restored.

    Open the source ↗
  52. [s52]Singapore's MAS sharpens oversight as AI risks mountQA Financial, 13 October 2026

    Passages cited

    MAS has already explored these issues through its Safeguards for Agentic Finance at Runtime framework, known as SAFR, published in July under its BuildFin.ai initiative.
    It provides “a set of governance checkpoints that verifies and records an AI agent’s proposed actions before the execution of its tasks”, MAS said when introducing the framework.
    SAFR is an industry-developed reference framework, distinct from the new supervisory guidelines.

    Trade press; SAFR is an industry reference framework, not a supervisory rule.

    Open the source ↗
  53. [s53]CFPB Orders Operator of Cash App to Pay $175 Million and Fix Its Failures on FraudConsumer Financial Protection Bureau, January 2025 (archived page)

    Passages cited

    Pay $120 million to harmed consumers: Block is required to pay up to $120 million in refunds and other redress to consumers whose unauthorized transfers were not investigated, consumers who did not receive refunds they were entitled to, and consumers whose accounts were locked for an extended period of time or who were not provided provisional credits during a delayed investigation.
    While Block included a telephone number on the back of its Cash Card and in its Cash App Terms of Service, for many years this telephone number did not connect consumers to customer support of any type.
    Instead, it led to a pre-recorded message directing consumers to contact customer support through the app.

    Enforcement order on earlier conduct; does not establish an AI cause.

    Open the source ↗
  54. [s54]SS1/23: Model risk management principles for banks (April 2026 version)Bank of England, Prudential Regulation Authority, 16 April 2026

    Passages cited

    1.2 This SS is relevant to all regulated United Kingdom (UK)-incorporated banks, building societies and PRA-designated investment firms with internal model approval to calculate regulatory capital requirements.
    b) Regarding third-party vendor models, firms should: (i) satisfy themselves that the vendor models have been validated to the same standards as their own internal MRM expectations; (ii) verify the relevance of vendor supplied data and their assumptions; and
    (iii) validate their own use of vendor products and conduct ongoing monitoring and outcomes analysis of vendor model performance using their own outcomes.

    The supervisory statement's own text; applies to firms with internal model approval for regulatory capital.

    Open the source ↗
  55. [s55]Enterprise AI Enters an ROI Reckoning: More Use, Uneven Value, Reveals Deloitte SurveyDeloitte, 29 October 2026

    Passages cited

    “The State of AI in the Enterprise: Built Differently, How AI is Reshaping the Way Enterprises Work, Grow and Create Value” is based on a survey of 3,500 director to C-suite-level leaders across 25 countries and six industries with direct involvement in their companies’ AI initiatives.
    AI has become a powerful accelerator, with 81% of companies reporting improved speed in decision-making and execution. Yet only 25% say they are using AI to deeply transform the business by creating new products or services, reinventing core processes, or fundamentally changing their business models.
    While 71% track total AI inference and API spend, only 51% track token volume and 44% track per-unit inference cost.
    Forty percent of companies now allow AI to execute decisions without human approval, at least to a moderate degree. Despite this growing autonomy, 23% say accountability remains unclear when an AI system produces a negative outcome.

    Cross-industry self-reported survey (May to June 2026), not bank-only.

    Open the source ↗
  56. [s56]FCA fines Monzo £21m for failings in financial crime controlsFinancial Conduct Authority, 8 July 2025

    Passages cited

    The FCA has fined Monzo Bank Ltd £21,091,300 for its inadequate anti-financial crime systems and controls between October 2018 and August 2020.
    Monzo's customer base has grown rapidly, increasing almost tenfold from around 600,000 in 2018 to over 5.8 million in 2022. However, Monzo's financial crime controls failed to keep pace with its customer and product growth.

    Enforcement on controls from 2018 to 2022; not an AI case.

    Open the source ↗
  57. [s57]Consumer & Community Context: Alternative Data: Expanding Access to CreditBoard of Governors of the Federal Reserve System, 22 October 2025

    Passages cited

    Roughly 32 million American adults are estimated to be "unscoreable" in regards to a credit score, including "credit invisible" that have no credit history (7 million, or 2.7 percent of all adults) or adults with a "thin file" (25 million, or 9.8 percent) (see figure 1).
    Nineteen percent of the respondents to the 2024 SHED survey reported that their spending exceeded their income in the month they answered the survey. Eleven percent reported struggling to pay their bills because of varying income, including nearly 20 percent of those with income under $50,000.

    US estimates and survey data; not a measure of any lender's model.

    Open the source ↗
  58. [s58]FinRegLab Study Shows Cash-Flow Data Can Expand Small Business LendingFinRegLab, 3 June 2025

    Passages cited

    The report, titled Sharpening the Focus: Using Cash-Flow Data to Underwrite Financially Constrained Businesses, analyzes over 38,000 small business loans originated by two fintech lenders between February 2015 and January 2024.
    Models that added cash-flow variables to personal credit scores and other traditional inputs predicted default risk more accurately across all borrower segments analyzed. - Gains were particularly large for low-score entrepreneurs whose businesses are less than five years old.

    Two fintech lenders' small-business loans; a research nonprofit's release.

    Open the source ↗
  59. [s59]CFPB Research Reveals Heavy Buy Now, Pay Later Use Among Borrowers with High Credit Balances and Multiple Pay-in-Four LoansConsumer Financial Protection Bureau, January 2025 (archived page)

    Passages cited

    Today’s study helps fill the data gap by pairing a matched sample of BNPL applications from six large firms with deidentified credit records.
    Approximately 63 percent of borrowers originated multiple simultaneous loans at some point during the year, and 33 percent took out loans from multiple BNPL lenders. - Nearly two-thirds of BNPL loans went to borrowers with lower credit scores: Among these applicants with subprime or deep subprime credit scores, BNPL lenders approved 78 percent of loans in 2022.
    Because lenders do not typically report BNPL loans to nationwide consumer reporting companies, data about BNPL use—especially about borrowers with multiple loans and on total consumer debt balances—is limited.

    2021 to 2022 data from six large BNPL firms matched to credit records.

    Open the source ↗
  60. [s60]Why Asia's biggest microfinance lenders face debt stressThe Banker, 20 August 2026

    Passages cited

    Indian microfinance portfolios contracted by 28 per cent between March 2024 and December 2025 following what Geeta Chugh, banking analyst at S&P Global Ratings, describes as a period of “rapid credit expansion, multiple borrowing and lax underwriting”.
    In Bangladesh, ASA, the nation’s largest lender, stresses that multiple borrowing is “drying up repayment capacity” and lenders are unable to establish how many competing institutions have already extended credit to the same customer.

    Press analysis quoting an S&P Global Ratings analyst and ASA.

    Open the source ↗
  61. [s61]Rating Rationale: Fusion Finance LtdCRISIL Ratings, 30 December 2025

    Passages cited

    Reported metrics have also improved, supported by both lower slippage and write-offs, with gross stage III declining to 4.6% as on September 30, 2025, from 7.9% as on March 31, 2025. […] Such borrowers are typically farmers , tailors, cattle owners/traders, small vegetable vendors, teashop owners and dairy farmers. The incomes of these households could be volatile and dependent on the performance of the local economy.
    However, over the past few months, the company’s collections have seen steady improvement with overall collection efficiency improving to 89.7% during April to September 2025 as against 85% during Q4 fiscal 2025.

    A rating rationale for one lender; it does not attribute outcomes to AI or to any model.

    Open the source ↗
  62. [s62]FCA regulation of deferred payment credit: what BNPL firms must do before July 2026PKF Littlejohn, 9 June 2026

    Passage cited

    The FCA’s new regime for deferred payment credit (buy now pay later) comes into force on 15 July 2026. If you provide BNPL lending, you face new authorisation, affordability, and reporting requirements, alongside significant accounting implications.

    An accounting firm's summary of the FCA regime.

    Open the source ↗
  63. [s63]Why stronger Buy Now Pay Later protections are the right step for consumersFinancial Conduct Authority (blog), 26 June 2026

    Passages cited

    Too often, we have seen people using BNPL repeatedly without a clear view of what they owe, leading to missed payments, late fees and, in some cases, worsening financial circumstances.
    We know that some customers who have regularly used BNPL may find it harder to access it once they have to go through an affordability check.

    Regulator's blog, not the rules themselves.

    Open the source ↗
  64. [s64]Joint EBA-ECB report on payment fraud: strong authentication remains effective but fraudsters are adaptingEuropean Central Bank and European Banking Authority, 15 December 2025

    Passages cited

    The report assesses payment fraud reported by the industry across the European Economic Area (EEA), which amounted to €3.4 billion in 2022, €3.5 billion in 2023 and €4.2 billion in 2024.
    Strong customer authentication remains effective against the fraud types it was designed to mitigate and that were dominant at the time PSD2 came into force, especially for card payments - However, new types of fraud are on the rise, particularly the manipulation of payers, which will require additional and new mitigation approaches
    For credit transfers, payment service users bore approximately 85% of total fraud losses in 2024, mainly as a result of scams that tricked users into initiating fraudulent transactions.

    Industry-reported fraud across the EEA, 2022 to 2024.

    Open the source ↗
  65. [s65]APP Scams Reimbursement DashboardPayment Systems Regulator, updated 30 July 2026

    Passages cited

    In the 18 months (between 7 October 2024 to 31 March 2026): - 88% (£316m) of the money lost to APP scams has been reimbursed to victims.
    Consumers have reported around 438,300 claims; 301,500 were in scope for reimbursement. - 82% of claims were closed within five business days – and 98% within 35 business days.
    This Dashboard uses the data provided by sending firms as per reporting standard A in the Compliance Data Reporting Standard (CDRS).
    We do not have information from receiving firms to validate the information.

    Closed Faster Payments claims reported by sending firms; not validated against receiving firms.

    Open the source ↗
  66. [s66]Payment fraud falls by £73m following PSR reimbursement schemePayment Systems Regulator, 1 July 2026

    Passages cited

    Frontier found that APP fraud losses have fallen by an estimated £73 million per year and the number of APP scams have fallen by nearly 35,000 due to the policy.
    Frontier Economics’ analysis is based on the date of the scam transaction, while UK Finance data is based on the date the claim is closed by the PSP.
    The PSR APP reimbursement dashboard reflects claims closed in the reporting period, excluding transactions before 7 October 2024 and those out of scope (e.g. international payments and business transactions). The UK Finance data covers all APP fraud losses, not just those within scope of the APP reimbursement policy.

    The regulator's announcement of the independent Frontier Economics evaluation.

    Open the source ↗
  67. [s67]APP fraud reimbursement requirement: were industry's concerns justified?Grant Thornton, 21 September 2026

    Passage cited

    In contrast, UK Finance, a banking and finance trade organisation, reported a 19% rise in APP fraud in 2025. This gap was driven by the different measurement choices. The evaluation counted scams by the date they occurred; UK Finance counted by the date a claim closed.

    Advisory firm commentary on the evaluation.

    Open the source ↗
  68. [s68]United States of America v. TD Bank, N.A.US Department of Justice, October 2024

    Passages cited

    Throughout this time, TD Bank intentionally excluded all domestic automated clearinghouse (ACH) transactions, most check activity, and numerous other transaction types from its automated transaction monitoring system, resulting in 92% of total transaction volume going unmonitored from Jan. 1, 2018, to April 12, 2024.
    This monitoring gap amounted to approximately $18.3 trillion in transaction activity. TD Bank’s failures enabled three money laundering networks, one of which was assisted by five TD Bank employees, to collectively transfer more than $670 million through TD Bank accounts between 2019 and 2023.
    From 2014 through 2022, TD Bank failed to add any new scenarios to its transaction monitoring program despite known deficiencies, emerging money laundering risks, and the introduction of new products and services.

    Plea agreement facts; conduct from 2014 to 2024; not an AI case.

    Open the source ↗
  69. [s69]FCA fines Nationwide £44m for failings in financial crime controlsFinancial Conduct Authority, 12 December 2025

    Passages cited

    The FCA has fined Nationwide Building Society £44m for inadequate anti-financial crime systems and controls between October 2016 to July 2021.
    Nationwide would have been fined £62,969,297, but it agreed to resolve these matters and so qualified for a 30% discount under the FCA's processes. The total fine is £44,078,500.
    During this period, Nationwide had ineffective systems for keeping up-to-date due diligence and risk assessments for all its personal current account customers and for monitoring their transactions.
    Nationwide was also aware that some of those customers were using their personal accounts for business activity, in breach of its terms.
    In one serious case, Nationwide missed opportunities to identify a customer using personal current accounts to receive fraudulent Covid furlough payments. The customer received 24 payments totalling £27.3m over 13 months, with £26.01m of this deposited over 8 days. His Majesty’s Revenue & Customs (HMRC) recovered £26.5m, but approximately £800,000 remains unrecovered.

    Enforcement on controls from 2016 to 2021; not an AI case.

    Open the source ↗
  70. [s70]Can AI Replace Human Debt Collectors?Yale School of Management Insights, 20 May 2025

    Passages cited

    The working paper, based on 22 million cases that entered debt collection between April 2021 and December 2022, was co-authored with Yale SOM PhD student Dong Huang, Zhishu Yang of Tsinghua University, and Qi Zhang of Shanghai Jiao Tong University.
    Comparing debts just above versus just below the 300 yuan threshold, Choi and his colleagues find that the value of repayments collected by AI callers during the first 30 days past due is 9% less than the value collected by humans; although the gap diminishes over time, AI callers have collected 5% less even a year later.
    For some larger debts, the company randomly assigned half to be called by AI for the first five days before being switched to human callers, and the other half to be called by humans from the beginning.
    Borrowers who were initially contacted by AI had repaid less of their initial late payment a year later and were more likely to miss subsequent payments than borrowers who only received calls from human collection agents.

    Summary of a working paper using 2021 to 2022 cases in China; repayment, not borrower welfare.

    Open the source ↗
  71. [s71]Federal Court orders Westpac to pay $26 million penalty for hardship failuresAustralian Securities and Investments Commission, 27 May 2026

    Passages cited

    Westpac Banking Corporation (Westpac) has been ordered to pay $26 million in civil penalties for failing to respond to customers who were facing financial hardship.
    The Honourable Justice McEvoy ordered the penalty after finding that Westpac failed to respond to over 200 online hardship requests within the time required by law over nearly six years from 2017 to 2023.
    ‘These were customers who were asking for some breathing room for a range of reasons including domestic abuse, natural disasters, serious illness or the loss of their job.
    ‘A particularly serious aspect of Westpac’s contraventions is that they caused a number of customers to have adverse credit information recorded on their credit files, and debts to be sold to third-party debt purchasers who then engaged actively in conduct to pursue those debts.
    Justice McEvoy said, ‘while the contraventions were not suggested to be deliberate and arose instead from inadequate systems and operational failures, I have accepted that they were grossly negligent’.

    Court penalty for system failures from 2017 to 2023; not an AI case.

    Open the source ↗
  72. [s72]Annual Economic Report 2025, chapter I: Sustaining stability amid uncertainty and fragmentationBank for International Settlements, 29 June 2025

    Passage cited

    More recently, a more orderly unwinding of relative value trades – this time tied to interest rate swap markets, where investors had bet on a narrowing in spreads due to potential deregulation – seems to have contributed to the heightened volatility observed in Treasury markets in early April 2025.

    Cautious causal wording; not an AI episode.

    Open the source ↗
  73. [s73]Financial Policy Committee Record, April 2025Bank of England, 9 April 2025

    Passage cited

    While the margin calls faced by funds following 2 April had been significant, they had so far been able to meet them without taking actions which would further amplify the market volatility.

    A record of the committee's assessment at the time.

    Open the source ↗
  74. [s74]Financial Stability Report, July 2026Bank of England, 7 July 2026

    Passages cited

    Financial markets: evidence to date suggests that trading firms are using more autonomous AI systems, primarily for research, coding support, surveillance and other lower-risk operational tasks, rather than for fully autonomous trading.
    But if AI systems become better at core financial decision-making tasks and start to outperform more traditional models in testing environments then firms may begin to rely on them more directly in trading or portfolio decisions in ways that change the speed and nature of adjustment to new information and increase the risk of correlated behaviour.

    Assessment and warning, not measured AI trading harm.

    Open the source ↗
  75. [s75]Annual Economic Report 2026, chapter I: Progress and perilBank for International Settlements, 28 June 2026

    Passage cited

    Supervisory visibility into AI use by market participants needs to improve to capture model-driven correlated exposures, herding behaviour and the potential for algorithmic collusion.

    Policy recommendation.

    Open the source ↗
  76. [s76]PwC's 2026 Annual Corporate Directors SurveyPwC US Governance Insights Center, 23 September 2026

    Passage cited

    Directors also report significant gaps in the information needed to oversee AI strategy and execution, with 82% rating information linking AI outcomes, risks, and business performance as fair, poor, or not provided.

    Survey of US public-company directors across sectors, not bank boards alone; self-reported.

    Open the source ↗
  77. [s77]Kyndryl Report: AI returns rise as enterprise readiness fallsKyndryl, 10 November 2026

    Passage cited

    While 64% of leaders report positive returns on AI investments, up 10 points from 2025, average readiness across external risks fell seven points to 24%. […] Just 27% have formally updated roles and responsibilities around AI

    Cross-sector survey commissioned by an IT infrastructure provider; self-reported.

    Open the source ↗
  78. [s78]Expanding on what we missed with sycophancyOpenAI, 2 May 2025

    Passage cited

    On April 25th, we rolled out an update to GPT‑4o in ChatGPT that made the model noticeably more sycophantic. It aimed to please the user, not …

    The vendor's own account of a consumer product; passage taken from the search result.

    Open the source ↗
  79. [s79]Sycophancy in GPT-4o: What happened and what we're doing about itOpenAI, 29 April 2025

    Passage cited

    We have rolled back last week's GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior.

    The vendor's own account; passage taken from the search result.

    Open the source ↗
  80. [s80]2026 FINRA Annual Regulatory Oversight Report: GenAIFinancial Industry Regulatory Authority, January 2026

    Passage cited

    This may include storing prompt and output logs for accountability and troubleshooting; tracking which model version was used and when; and validation and human-in-the-loop review of model outputs, including performing regular checks for errors or bias.

    For US broker-dealers; examination considerations, not new rules.

    Open the source ↗
  81. [s81]FCA fines Barclays £42 million for poor handling of financial crime risksFinancial Conduct Authority, 5 December 2025

    Passage cited

    In the second case, the FCA has fined Barclays Bank PLC £39.3 million for failing to adequately manage money laundering risks associated with providing banking services to Stunt & Co. […] In the space of just over a year, Stunt & Co received £46.8 million from Fowler Oldfield, a multimillion-pound money laundering operation. Barclays failed to properly consider the money laundering risks associated with the firm even after receiving information from law enforcement about suspected money laundering through Fowler Oldfield, and after learning that the police had raided both firms. Barclays only conducted a review of its exposure to Fowler Oldfield through its customers, including Stunt & Co, after it learned of the FCA’s decision to prosecute NatWest over their relationship with Fowler Oldfield.

    Enforcement on legacy monitoring controls; not an AI case.

    Open the source ↗
  82. [s82]Commonwealth Bank backtracks on AI job cuts, apologisesABC News, 21 August 2025

    Passage cited

    In short: Commonwealth Bank announced 45 job cuts last month, as it introduced an AI "voice-bot", but has now reversed its decision.

    Press report; passage taken from the search result.

    Open the source ↗
  83. [s83]CBA reverses AI-driven job cuts, admits 'error'Information Age (ACS), 21 August 2025

    Passage cited

    CBA reverses AI-driven job cuts, admits 'error'. Bank backflips, says 'roles were not redundant'.

    Press report; passage taken from the search result.

    Open the source ↗
  84. [s84]Responsible Adoption of Artificial IntelligenceHong Kong Monetary Authority, 15 October 2026

    Passage cited

    Responsible Adoption of Artificial Intelligence. Issue Date: 15 Oct 2026.

    Register entry for the circular; the circular text itself not read.

    Open the source ↗
  85. [s85]HKMA ART Framework: Nine AI Principles for Hong Kong BanksRegReportingDesk (regulatory publisher), 19 October 2026

    Passage cited

    Annex A says material or higher-risk use cases should be subject to senior management sign-off and independent validation, a kill switch or equivalent should be built in for more material or higher-risk cases, and records enabling outcome tracing should be maintained as far as practicable.

    A publisher's summary of the HKMA annex, not the annex.

    Open the source ↗
  86. [s86]Article 26: Obligations of deployers of high-risk AI systemsEuropean Commission, AI Act Service Desk (text of Regulation (EU) 2024/1689), read 13 November 2026

    Passages cited

    Where deployers have reason to consider that the use of the high-risk AI system in accordance with the instructions may result in that AI system presenting a risk within the meaning of Article 79(1), they shall, without undue delay, inform the provider or distributor and the relevant market surveillance authority, and shall suspend the use of that system.
    For deployers that are financial institutions subject to requirements regarding their internal governance, arrangements or processes under Union financial services law, the monitoring obligation set out in the first subparagraph shall be deemed to be fulfilled by complying with the rules on internal governance arrangements, processes and mechanisms pursuant to the relevant financial service law.

    The Regulation's own text; applies to credit scoring once the high-risk duties apply.

    Open the source ↗
  87. [s87]12 CFR § 1002.9 Notifications (Regulation B)Consumer Financial Protection Bureau, 1 January 2025

    Passage cited

    The statement of reasons for adverse action required by paragraph (a)(2)(i) of this section must be specific and indicate the principal reason(s) for the adverse action.

    US rule on credit notifications; does not govern model validation.

    Open the source ↗
  88. [s88]Global Finance & Technology Network: Who We AreGFTN, 2026

    Passages cited

    The Global Finance & Technology Network (GFTN) is a not-for-profit organisation established by the Monetary Authority of Singapore (MAS) in 2024 to harness technology and foster innovation for more efficient, resilient, and inclusive financial ecosystems through global partnerships.
    GFTN organises convening forums, offers advisory services on innovation ecosystems, provides access to transformative digital platforms, and invests in technology startups with the potential for growth and positive social impact through its venture fund.
    GFTN hosts a worldwide network of forums and programmes; advises governments and companies on policies and the development of digital ecosystems and innovation within the financial sector; offers digital infrastructure solutions, and plans to invest in financial technology startups through its upcoming venture fund, with a focus on inclusion and sustainability.
    GFTN is a not-for profit organisation established by the Monetary Authority of Singapore (MAS) in 2024, with offices in Singapore, Germany, Japan, India, and China. Global Hubs: Singapore, Japan, Europe, Middle East & Africa

    The publisher's own institutional description on gftn.co and gftn.co/who-we-are/.

    Open the source ↗
  89. [s89]Digital Credit Algorithms in India: Expanding Access or Cementing Exclusion?Center for Financial Inclusion, 2022; page metadata 17 April 2024

    Passages cited

    Earlier this year, CFI conducted desk research and interviews across the digital finance ecosystem in India and learned that most digital lenders are focused on serving urban, young, and employed consumers.
    With higher and more predictable income streams, this customer segment creates large volumes of digital data that can be mined to determine creditworthiness. And given the overall size of the Indian market, there is still room to grow before even this segment is saturated. But where does that leave the rest of consumers, including migrants?
    Millions of migrants face similar challenges as Mohan. In general, migrants have fainter data trails due to lower levels of digital activity. Many low-income migrants have limited financial and digital capability and do not possess smartphones.
    And those who do have a smartphone typically only use mobile data when necessary due to the high cost. Additionally, migrants often rely on agents, such as kirana store owners, to assist them with digital transactions, and these transactions may not be directly associated with their identity.
    When it comes to assessing cashflows to determine repayment capacity, migrants’ digital data trails make it difficult to see the full picture of their economic lives.
    For example, even if a migrant has established a stable income over multiple years, as is the case with Mohan, often their payment patterns are not well-understood or formalized, organized, or digitized.
    For example, some contractors provide an advance payment while others pay weekly or monthly, and some cover living expenses for migrants but deduct the costs from their salaries.

    Historical qualitative research on Indian migrants; not a 2025 or 2026 measurement or a claim about all poor households.

    Open the source ↗
  90. [s90]Regulatory measures towards consumer credit and bank credit to NBFCsReserve Bank of India, 16 November 2023

    Passage cited

    November 16, 2023 […] Please refer to Governor’s Statement dated October 6, 2023 flagging the high growth in certain components of consumer credit and advising banks and non-banking financial companies (NBFCs) to strengthen their internal surveillance mechanisms, address the build-up of risks, if any, and institute suitable safeguards, in their own interest. […] On a review, it has been decided to increase the risk weights in respect of consumer credit exposure of commercial banks (outstanding as well as new), including personal loans, but excluding housing loans, education loans, vehicle loans and loans secured by gold and gold jewellery, by 25 percentage points to 125%.

    Historical circular; specified exposures and exclusions. Does not prove that losses had already occurred.

    Open the source ↗
  91. [s91]Exposures of SCBs to NBFCs: Review of Risk WeightsReserve Bank of India, 25 February 2025

    Passage cited

    February 25, 2025 […] On a review, it has been decided to restore the risk weights applicable to such exposures and the same shall be as per the external rating […] The above instructions shall come into effect from April 01, 2025. All other instructions of the circulars ibid remain unchanged.

    The reversal applies to the specified bank exposures to NBFCs from 1 April 2025, not every 2023 measure.

    Open the source ↗
  92. [s92]RBI Guidelines On Default Loss Guarantee (DLG) In Digital LendingLegal 500, legal commentary, December 2024

    Passages cited

    The Reserve Bank of India ("RBI") has framed a comprehensive infrastructure of regulatory stepping stones for default loss guarantee in digital lending through 'Guidelines on Default Loss Guarantee in Digital Lending' ("DLG Guidelines") issued on June 8, 2023.
    As per the DLG Guidelines, the total amount of DLG cover on any outstanding portfolio specified upfront cannot exceed five per cent of the amount of that loan portfolio.
    At present, the contractual arrangements between REs and LSPs or between two REs involving default loss guarantee (DLG), commonly known as First Loss Default Guarantee ("FLDG") are subject to the DLG Guidelines.
    The grains of such arrangement outlines the conceptual basis that such third-party guarantees compensate RE the loss incurred due to default up to a certain percentage of the loan portfolio of the RE, specified upfront, due to default on account of the borrower ("DLG Provider").

    Secondary legal commentary on the 8 June 2023 rules. Historical cap; original RBI circular also read.

    Open the source ↗
  93. [s93]Guidelines on Default Loss Guarantee: repeal noticeReserve Bank of India, 8 June 2023; replacement dated 8 May 2025

    Passage cited

    The Circular has been repealed. Please refer to Reserve Bank of India (Digital Lending) Directions, 2025 dated May 8, 2025.

    The RBI page directs readers to the 2025 Digital Lending Directions.

    Open the source ↗
  94. [s94]Statement on Developmental and Regulatory PoliciesReserve Bank of India, 8 August 2024

    Passage cited

    With a view to provide a more up-to-date picture of a borrower’s indebtedness, it has been decided to increase the frequency of reporting of credit information to CICs from monthly intervals to fortnightly basis or at such shorter intervals as mutually agreed between the CI and CIC.

    Reporting-frequency announcement; not real-time visibility of every debt.

    Open the source ↗

About GFTN, author and disclaimer

Author and contributors

Author

Contributors

With the support of: Sopnendu Mohanty, Group Chief Executive Officer, GFTN; Dr Andrzej Gwizdalski, Senior Partner, GFTN.

Correspondence

For further information or correspondence, contact madhu.subramanya@gftn.co (LinkedIn · GitHub). GFTN's ALFIN AI assisted with research and drafting.

About Global Finance & Technology Network (GFTN)

The Global Finance & Technology Network (GFTN) is "a not-for-profit organisation established by the Monetary Authority of Singapore (MAS) in 2024 to harness technology and foster innovation for more efficient, resilient, and inclusive financial ecosystems through global partnerships"[s88]. Headquartered in Singapore at 6 Battery Road, #28-01, Singapore 049909 (gftn.co · hello@gftn.com), GFTN operates across offices in Singapore, Germany, Japan, India, and China, with global hubs spanning Singapore, Japan, Europe, and the Middle East & Africa[s88].

Bridging policy, capital, and technology, GFTN hosts a worldwide network of convening forums and programmes (including its flagship Singapore FinTech Festival, Insights2040, the Point Zero Forum, and Black Swan Summit); advises governments, central banks, and corporates on regulatory policy and the development of digital financial ecosystems[s88]; provides access to transformative digital infrastructure and market-intelligence platforms (including GFTN Insights, FutureMatters, and ALFIN AI); and invests in financial technology startups with the potential for growth, inclusion, and positive social impact through its venture fund[s88].

Legal notice and disclaimer

Global Finance & Technology Network Limited (GFTN) · 6 Battery Road, #28-01, Singapore 049909 · gftn.co | hello@gftn.com
© 2026 Global Finance & Technology Network Limited, All Rights Reserved. Reproduction Prohibited.

Powered by

Research, sourcing and drafting support from GFTN's ALFIN AI. Every figure and quotation in this report is linked to the published passage it comes from.

Learn more about what ALFIN can do for you