On 22 October 2026 the software company SAS presented a study under the headline “Banks expect to hand decisions to AI – even if they can’t explain them”. The figure that has been repeated since appears in the press release: 76% of banks, it says, expect autonomous credit and risk decisions by artificial intelligence by 2031. The underlying report, “Why AI at Scale Needs Governance at Speed”, was produced by FT Longitude, an agency within the Financial Times Group, and sponsored by SAS and Microsoft. In June and July it surveyed 300 credit, treasury and risk executives at banks with total assets of at least USD 20bn across 31 countries.

Most readers take the number as a sign that machines will be deciding on loans within a few years. The report itself does not support that reading. What it measures is an expectation, it defines neither “autonomous” nor “at scale”, and according to its own conclusion the experts interviewed for the study consider such a rapid change unlikely. A question the study raises in passing is more productive than the date: how does a bank switch off a model that has crossed its limits? In securities trading that question has had a binding answer since 2018. For credit and risk decisions, European law requires limits and, depending on the type of decision, a human being, yet no rule on switching the system off. Above all, it is still unclear what state a credit process returns to once it has been stopped.

The analysis draws on the study report, the press release, the survey on artificial intelligence in UK financial services that the Bank of England and the Financial Conduct Authority (FCA) published on 26 October, and the relevant legal texts in banking supervision, data protection and securities law. Applying the trading rules to lending is an inference from trading practice, neither a forecast nor a statement by any authority. One point on the source: SAS sells software for risk management and AI governance. It also sponsors the RiskMinds International conference, which opens in London on the day this piece appears.

In brief

What: the study “Why AI at Scale Needs Governance at Speed” by FT Longitude, sponsored by SAS and Microsoft, presented on 22 October 2026

Finding: 76% of the 300 respondents expect autonomous credit or risk decisions “at scale” within one to five years, 10% say this is already happening

Today: for the credit process, 7% of respondents say AI decides autonomously with ongoing monitoring, 33% report AI decisions with human oversight or control mechanisms

Law: limits for automated credit decisions under the European Banking Authority (EBA) Guidelines EBA/GL/2020/06 and the German Minimum Requirements for Risk Management (MaRisk), human judgement in ratings under the internal ratings-based (IRB) approach under Article 174(e) of the Capital Requirements Regulation (CRR), a right to human intervention for general consumer loans from 20 November 2026, the stop button of the EU AI Act from 2 December 2027

Open: a binding rule for switching off with a defined safe state, as securities trading has had since 2018

The 76% is the sum of two answers to a question about expectations

The figure comes from a single question on page 12 of the report, which reads: “When do you expect AI to make autonomous risk or credit decisions at scale in your organization?” 10% of respondents said this was already happening, 25% expect it within one to two years, 51% within three to five years and 14% further out. The 76% is the sum of the two middle answers. The press release turned “within five years” into “by 2031”, although no answer in the report carries that year. Nor does the report say anywhere what “autonomous” or “at scale” means. Each respondent therefore put a date on their own idea of autonomy.

When do respondents expect autonomous AI on credit and risk? Answers from 300 credit, treasury and risk executives 10% already 25% in 1–2 years 51% in 3–5 years 14% beyond 5 years 76% which the press release turns into “by 2031”
Answers from 300 credit, treasury and risk executives to the question of when AI will make autonomous credit or risk decisions “at scale” in their organisation. The headline 76% is the sum of the answers “within one to two years” and “within three to five years”, excluding the 10% who say it is already happening. The study does not define “autonomous” or “at scale”. Source: FT Longitude, sponsored by SAS and Microsoft, October 2026, p. 12, survey June/July 2026.

Where banks stand today is the subject of a different question in the report, and the numbers there are much lower. For the credit process, 7% of respondents say AI already decides autonomously with ongoing monitoring (chart on page 11), for asset and liability management (ALM) the figure is 4% and for the organisation as a whole 3%. A further 33% report AI decisions in the credit process that run under human oversight or with control mechanisms. Some 70% of respondents say they place moderate or more extensive limits on the use of AI for autonomous decisions. The report is careful to draw that distinction itself: machines handle 23% of decision-making processes on average, “but this is not the same as making 23% of decisions”.

The report does, admittedly, show that autonomy in lending is already possible. A senior credit risk executive at an unnamed Canadian bank says that in retail credit “more than 90% are AI-based decisions with very little human-in-the-loop methods”. For high-risk portfolios, by contrast, the bank has kept a human in the loop on purpose until it has more confidence in the technology. The example shows a pattern that fits the caution of the other respondents. What gets automated is small-ticket volume business, and the human stays wherever a single wrong decision becomes expensive. Marilin Pikaro, Director of Innovation, Conduct and Consumers at the European Banking Authority (EBA), paints a similar picture in her interview with the study’s authors. The authority’s surveys, she says, show that AI, especially in high-risk areas, “is still applied as a supporting tool, and not making individual decisions”.

The experts we interviewed for this study assess such a rapid change as unlikely. FT Longitude, “Why AI at Scale Needs Governance at Speed”, sponsored by SAS and Microsoft, October 2026

The study is still more than a mood survey, because it counts gaps in governance as reported by the banks themselves. Only 18% of respondents say they have fully embedded AI governance processes, 19% monitor the drift of their models in real time, and 42% have no defined “kill switches” for shutting down models whose behaviour moves beyond acceptable limits. That a vendor of governance software sponsored the survey does not invalidate it. The figures nevertheless remain self-reported answers in a sponsored survey whose recruitment the report does not describe. What carries this piece is therefore less the size of the 42% than the concept the study uses for it: the kill switch that shuts a model down.

UK supervisors find few agents that decide on customers

Four days after the SAS study, on 26 October 2026, the Bank of England and the FCA published their fourth survey of artificial intelligence in UK financial services. According to the report, 86% of participants use AI, up from 75% in 2024. The supervisors also asked about AI agents and their degree of autonomy. Some 80% of participants run agents that recommend actions while humans decide and execute, and only 2% report agents with broad autonomy. Even when planned deployment over the next 12 months is included, the report finds that “no more than 9%” of participants expect agents to approve or decline customer outcomes, execute trades, initiate payments or change limits and controls. In commercial banking and in trading, not a single participant reports an agent with broad autonomy.

The two surveys cannot be equated. The supervisors analysed 196 responses from across UK financial services, insurers and asset managers among them. The figures on agents come from the 105 full questionnaires, and the horizon is 12 months. The SAS study asked 300 credit, treasury and risk executives at banks in 31 countries about their expectations over five years. Read side by side, they still reveal a tension that the headline hides. When a supervisor asks about deployment and concrete plans, it finds single-digit shares, and when a vendor asks about expectations, it finds three quarters. In its workshops with significant institutions, the European Central Bank (ECB) found a similar picture, with automation in credit scoring limited at most to small retail loans, as the analysis of the ECB’s supervisory AI statistics from October shows.

The law requires a stop button in lending only from December 2027, and only for scoring individuals

How much autonomy the law permits depends on the type of decision, and the room is narrowest where banks calculate their capital. For ratings under the internal ratings-based (IRB) approach, Article 174(e) of the Capital Requirements Regulation (CRR) requires institutions to complement the statistical model with “human judgement and human oversight”, to review the model’s assignments and to document how human judgement and model results are combined. The sentence has been in the regulation since its original 2013 version. A rating assigned by a model without this review structure would not comply with it. The wording leaves open whether reviewing a sample is enough.

For loan origination itself, the EBA Guidelines EBA/GL/2020/06, which apply to significant institutions, require the credit risk policy to identify the “products, segments and limits” for which automated decision-making is allowed (paragraph 38(g)) and the decision-making powers of automated models (paragraph 65). For less significant institutions, the conditions for automated decision-making sit in section BTO 1.2 of the German Minimum Requirements for Risk Management (MaRisk). Limits for automated credit decisions have therefore been mandatory for years, as the analysis of Deutsche Bank’s corporate credit decisions sets out in detail.

For consumers, a person enters the picture. Since the judgment of the Court of Justice of the European Union (CJEU) on SCHUFA of 7 December 2023, a score that plays “a determining role” in granting credit is itself an automated decision under Article 22 of the General Data Protection Regulation (GDPR). From 20 November 2026, four days after this piece appears, borrowers taking out a general consumer loan with an automated creditworthiness assessment can demand the intervention of a person under Section 30(6) of the German Federal Data Protection Act (BDSG). The analysis of the FCA case studies covers the provision at length. From 2 December 2027, assessing the creditworthiness of natural persons also counts as a high-risk use under the EU AI Act. Its Article 14(3) then scales human oversight to the risks, the “level of autonomy” and the context of use. It does not prohibit autonomy. It asks for more oversight the more autonomy a system has.

Article 14(4)(e) goes further and requires a capability that banking supervision has not so far explicitly demanded for credit processes. Human oversight must be able “to intervene in the operation of the high-risk AI system or interrupt the system through a ‘stop’ button or a similar procedure that allows the system to come to a halt in a safe state”. The obligation applies only to creditworthiness assessments and credit scoring of natural persons. Annex III of the regulation does not cover loans to legal entities, ALM or treasury, and the position of sole traders and individual guarantors is unsettled. For IRB ratings, only Recital 58 mentions an exception for systems used to calculate capital requirements, while the annex itself excludes only fraud detection.

On the reading offered here, the legal latitude is widest in ALM and treasury, where neither Article 22 GDPR nor the AI Act applies. It is bounded there by the limit system, segregation of duties and the governance of model risk. According to the study, AI also decides autonomously less often there than in the credit process. The logic of switching off is not new in Germany, however. The Federal Financial Supervisory Authority (BaFin) outlined it in June 2021 in its principles for the use of algorithms in decision-making, as a tiered process with thresholds that ends in a “stopping rule”. At that final stage, approval is to be withheld at first, the model reviewed and the process interrupted. BaFin itself calls these principles “preliminary thoughts” and “guidance”. Switching off has thus been described for five years, yet it is binding in only one part of the bank.

In securities trading an algorithm has shut itself down at fixed thresholds since 2018

Anyone who has run a trading book knows two moves that the debate on AI tends to blur. One pulls open orders out of the market, the other stops the strategy that generates them. Commission Delegated Regulation (EU) 2017/589 under the Markets in Financial Instruments Directive II (MiFID II), known in practice as RTS 6 (Regulatory Technical Standards 6) and applicable since 3 January 2018, treats them separately. The “kill functionality” in Article 12 lets an investment firm “cancel immediately, as an emergency measure, any or all of its unexecuted orders”. It is the tool that the analysis of a kill switch for AI agents described at firm level in July. Shutdown is covered by Article 14(3): the trading algorithm or system must be capable of being shut down as part of the business continuity arrangements “without creating disorderly trading conditions”.

Article 14(3) is not where the regime stops. Article 15(3) automates the stop: “After a pre-determined number of repeated executions, the trading system shall be automatically disabled until re-enabled by a designated staff member.” The regime also includes pre-trade controls that check every single order against price collars and maximum volumes (Article 15(1)), real-time monitoring (Article 16) and an annual review of the business continuity arrangements (Article 14(4)). In trading, then, the machine checks every order, and a person decides when the machine may carry on. Autonomy is bounded by thresholds that nobody has to negotiate case by case, and by someone who answers for the restart.

What the lack of a stopping procedure can do is shown by the Knight Capital case of 1 August 2012, which the analysis of the examination framework of the Office of the Comptroller of the Currency (OCC) of 9 November covers in detail. According to the order of the Securities and Exchange Commission (SEC), Knight had no procedure for halting its order router when it misbehaved. The router generated millions of orders in roughly 45 minutes, and Knight ultimately realised a loss of USD 460m. The case predates RTS 6 and rests on a US rule (Rule 15c3-5). It illustrates the damage without justifying the European rule.

The Financial Stability Board (FSB) has itself begun to draw the link between trading and artificial intelligence. In its consultation of 10 June 2026 it describes “kill switch” mechanisms for AI as a spectrum “ranging from complete shutdown (which can be used in disabling algorithmic trading) to dynamic or graduated degradation where systems transition from autonomous to human operation”. In its examination framework of 4 November, the OCC lists “circuit breakers, rate limits, least-privilege permissions, and kill switches” as possible means of containing an agent that acts incorrectly, and places real-time credit decisioning in its illustrative top tier. Neither document is binding, and the OCC itself states that its framework does not establish supervisory expectations for banks. Nor is the supervisory briefing of the European Securities and Markets Authority (ESMA) of 26 February 2026 binding. It records that “neither MiFID II nor RTS 6 specifically address AI”.

The analogy has a clear limit. The kill functionality cancels orders that have not yet been executed, and a loan that has been paid out cannot be cancelled. A faulty trading algorithm hurts the firm’s own book within minutes, and the profit and loss account gives the signal at the same time. A faulty credit model first hurts the applicant it wrongly turns down, and the firm’s own book only months later, when defaults on wrongly approved loans arrive. The individual’s rights under Article 22 GDPR and Section 30(6) BDSG remain in any case, so a rule on switching off would sit alongside them and would not replace individual review. And RTS 6 was written for deterministic algorithms. What the trading rules can offer lending is therefore an architecture, not a ready-made tool.

The safe state of a credit process is the open question

For its stop button, the AI Act uses the term “safe state”. It sounds familiar from trading, but it still has to be given meaning in lending. For a trading algorithm it is easy to describe, with open orders cancelled, positions back with the trader and the market carrying on. A credit process that decides largely on its own has no equivalent default. Once the model is halted, applications pile up and somebody has to work through them. For a bank that, like the Canadian example, takes more than 90% of decisions in its retail portfolios on an AI basis with little human involvement, the obvious question is whether its manual capacity would cope with a stop. The study does not say.

In the study, Marilin Pikaro calls for a contingency plan for cases in which more than one system is affected. Anyone using AI, she says, must work out what the firm should do “if there are multi-system failures, if it’s not just one system that is being impacted”. In trading, RTS 6 requires an annual review of the business continuity arrangements. The legal texts analysed for this piece contain no annual test tailored to automated credit decisions. How many applications a bank could decide by hand on the day after a shutdown, with which staff and under which rules, is therefore a question each institution has to answer for itself.

The triggers also have to differ from those in trading. Where trading relies on the number of repeated executions or a loss limit, lending is better served by measures that show a fault before it surfaces as a default: the approval and rejection rate per segment, the distribution of input data against the development sample, and the rate of overrides that credit analysts set against the model. This selection is an inference from trading practice, not a requirement of any authority. BaFin names none of these measures, but in 2021 it described the principle of tiered thresholds to which they can be applied.

In ALM the question may arise sooner than the study suggests. If an agent generates hedging trades and itself determines whether an order is placed, and at what time, price and size, that comes close to the definition of algorithmic trading in Section 80(2) of the German Securities Trading Act (WpHG), which refers to orders processed “with limited or no human intervention”. Whether such an agent falls within that definition depends on the individual case, and that is an interpretation offered here, not a view of the supervisor. If it did, RTS 6 would apply directly, and the least regulated area of risk management would acquire the strictest shutdown rule in banking law.

For banks expecting autonomous decisions within five years, what matters before the date is whether they can name, for each automated process, who stops it, when it stops itself, who restarts it and what happens to the applications in the meantime. As an inference from trading practice, anyone who wants machines to decide will find it hard to avoid this architecture. Trading does not show that it works for learning models, though, and according to ESMA, RTS 6 expressly does not deal with AI. In securities trading the answer has been written into law since 2018. In lending it exists only in guidance from 2021.

Recommendations for credit risk, ALM and treasury

1. Define shutdown triggers for every automated decision process

Alongside the products, segments and limits that EBA/GL/2020/06 (paragraph 38(g)) and MaRisk BTO 1.2 already require for automated credit decisions, banks can set thresholds at which the process stops: an unusual approval or rejection rate in a segment, a shift in input data, a rising override rate. They only work if they take effect without case-by-case negotiation, as with the automatic shutdown in Article 15(3) RTS 6.

2. Put a number on fallback capacity for each credit segment

It should be clear for each automated credit process which applications continue by hand after a stop, how many the team can decide per day and under which rules, and what happens to applications left waiting. The capacity follows from the application volume per segment and the share the model currently decides alone, so it can be stated in applications per day. How to test shutdown and the fallback route together is set out in the analysis of the OCC examination framework of 9 November. For credit scoring of natural persons, the system must from 2 December 2027 have a stop button under Article 14(4)(e) of the AI Act that brings it to a safe state. What happens to the applications left waiting is not something the provision addresses.

3. Tie the restart to a named function

Whoever releases a halted model should be designated in advance and work independently of the business unit that uses the model, for instance in risk control or model validation. In trading, Article 15(3) RTS 6 requires a designated staff member for the restart, and independence from the business is a sensible addition. Article 26(2) of the AI Act requires people with the necessary competence, training and authority for high-risk systems. The release after a stop is the moment at which a bank shows who is accountable for the system.

4. Check agents in ALM and treasury against the rules on algorithmic trading

Before an agent generates hedging trades with order parameters it sets itself, compliance and legal should assess whether the project falls under Section 80(2) WpHG. If it does, the kill functionality, the shutdown capability and the pre-trade controls under RTS 6 apply from day one, and ESMA’s briefing of February 2026 treats every recalibration of learning components as a change that must be tested.

Timeline: from the shutdown rule in trading to the stop button in lending
Rules, cases and surveys on the autonomy of machine decisions in banks
1 August 2012
Knight Capital
A router without a shutdown procedure generates millions of orders in about 45 minutes, ending in a loss of USD 460m.
3 January 2018
RTS 6 applies
Kill functionality, emergency shutdown and automatic shutdown after repeated executions become mandatory for algorithmic trading.
15 June 2021
BaFin principles for algorithms
Tiered thresholds ending in a “stopping rule”, published as guidance.
30 June 2021
EBA/GL/2020/06 applies
The credit risk policy must define products, segments and limits for automated credit decisions.
7 December 2023
CJEU judgment on SCHUFA
A score that plays a determining role in granting credit is itself an automated decision under Article 22 GDPR.
10 June 2026
FSB consultation
A kill switch for AI as a spectrum from complete shutdown to graduated fallback to human operation.
22 October 2026
Study by FT Longitude, SAS and Microsoft
76% of respondents expect autonomous credit or risk decisions within five years.
26 October 2026
Bank of England and FCA survey
No more than 9% of participants expect agents within 12 months that decide on customer outcomes, execute trades, initiate payments or change limits.
20 November 2026
Section 30(6) BDSG takes effect
Borrowers taking out a general consumer loan with an automated creditworthiness assessment can demand the intervention of a person.
2 December 2027
AI Act for credit scoring
High-risk obligations, including the stop button, for creditworthiness assessments of natural persons.

Glossary

Kill functionality: obligation under Article 12 RTS 6 to cancel immediately any or all unexecuted orders. It takes orders out of the market but does not switch off the strategy.

Shutdown: stopping the trading algorithm or system itself, in an emergency under Article 14(3) RTS 6 and automatically under Article 15(3) after a set number of repeated executions, with a designated person switching it back on.

Safe state: term from Article 14(4)(e) of the EU AI Act, meaning the condition in which a halted high-risk system comes to a stop without causing new harm.

IRB approach: method under which a bank calculates the capital requirements for credit risk with its own rating models, approved by the supervisor.

Override: a credit analyst’s deliberate departure from the model’s result. A rising override rate shows that people contradict the model more often.

ALM: asset and liability management, the steering of interest rate and liquidity risk arising from the interplay of the assets and liabilities on the balance sheet.

Christian Schablitzki

Christian Schablitzki – the agentic banker

Strategy & Management Consultant · Agentic AI expert for financial institutions

More than 20 years in investment banking and derivatives trading, followed by over 10 years advising financial institutions. Certified as Claude Partner – Claude Code (Anthropic), as well as in Google AI, Generative AI Leader (Google Cloud) and IBM RAG and Agentic AI.

LinkedIn profile → · agenticbanker.ai →
newsletter
the agentic banker

Keep reading – every 14 days in your inbox.

Capital-markets insights, regulatory updates and AI trends. Concise, well-founded, free.

GDPR-compliant. Unsubscribe at any time.

← Back to overview