Financial firms moved AI into operating plans as US model-risk guidance stepped back

Between April and June, banks attached headcount targets to AI and restricted staff access to a model provider after a U.S. export-control directive, while payment networks piloted checkout for software agents. Risk, operations and technology leaders made those calls as U.S. and U.K. supervisors changed how they oversee AI.

Editorial snapshot: 31 July 2026

The central finding

Financial firms now set more of their own AI standards: U.S. model-risk guidance leaves out generative and agentic AI, a government order took two Anthropic models offline, and banks made their own calls on whether AI savings go to margins, capacity or customers.

01Key findings

What changed in financial services AI in the second quarter

Four developments defined the quarter. Together they moved AI from counting use cases toward operating decisions about model supply, capital, jobs and oversight.

Capability

Model supply became a bank risk

Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened the largest banks in April over the cyber implications of Anthropic's Mythos Preview model. After a June 12 export-control directive, Anthropic disabled Claude Fable 5, which was generally available, and the more restricted Mythos 5 for all customers, and Evident reported that JPMorgan and Goldman Sachs cut Hong Kong-based staff off from Anthropic models.

Capital

Fintech funding rose on fewer deals

Global fintech venture funding reached $28.6 billion in the first half of 2026, up 22.7% from a year earlier, on 1,605 deals, down 25.7%, Crunchbase reported. Supervisors in Washington and London flagged stretched AI valuations and AI-linked credit as channels for financial stress.

Adoption and economics

Banks split on where AI savings land

Standard Chartered plans to cut about 7,800 corporate-function roles by 2030, and HSBC's chief executive told staff not to resist AI. Of eight U.S. banks reporting second-quarter results by July 15, half grew headcount over the past year, Evident found. On JPMorgan's July 14 results call, CEO Jamie Dimon said the benefit of AI goes to customers because rivals use the same tools.

Regulation

US model-risk guidance excludes generative and agentic AI

Guidance issued April 17 by the Fed, the Office of the Comptroller of the Currency and the FDIC replaced SR 11-7, placed generative and agentic AI outside its scope and said noncompliance will not draw supervisory criticism. A promised request for information on model risk, with attention to generative and agentic AI, had not been issued by June 30.

Operating question

With no supervisory template for generative and agentic AI, what standard will we apply to validating, sourcing and paying for those systems, and who signs off on it?

02Capability

Frontier models became a cyber, access and cost question for banks

For financial firms, the quarter's model releases raised three practical questions: what cyber-capable models mean for defense, who may use a model, and what it costs. Firms answered with routing and fallback engineering, and the Bank of England found little evidence of advanced AI in core credit, underwriting or trading decisions.

Cyber

Cyber-capable models reached supervisors

Anthropic released Claude Mythos Preview on April 7 only to a restricted group of partners for finding and fixing software vulnerabilities, and Treasury and Fed leaders met the largest banks that month on its cyber implications. The European Supervisory Authorities' first report on major incidents under the Digital Operational Resilience Act said highly capable AI-driven tools should push firms to strengthen cybersecurity.

Access

A government order took two models offline

On June 12, Anthropic disabled Claude Fable 5, which was generally available, and Mythos 5, which had limited distribution, for all customers after a U.S. export-control directive barring access by foreign nationals; new sessions fell back to Opus 4.8 or the user's default model. JPMorgan and Goldman Sachs cut Hong Kong-based staff off from Anthropic models, Evident reported, and Greenberg Traurig advised checking model aliases, routing, fallbacks and deemed-export exposure.

Cost

Flagship model prices rose

OpenAI released GPT-5.5 on April 23, and OpenRouter measured the effective increase across its own traffic. On the July 14 results call, JPMorgan Chief Financial Officer Jeremy Barnum said the latest, most expensive model is unnecessary for summarizing an analyst report.

2xGPT-5.5 list price vs. GPT-5.4

49% to 92%Higher effective cost on OpenRouter traffic

Redundancy

Banks engineer for more than one provider

Across the 50 banks Evident tracks, public use cases that reference OpenAI or Anthropic rose more than 70% over the past year. CIBC routes tasks to models with a system it built, and Santander's AI Lab open-sourced software that lets teams switch large language model (LLM) providers.

Core decisions

Little advanced AI in core credit and trading

The Bank of England's Financial Policy Committee (FPC) found little evidence that firms use advanced AI for core financial decisions such as credit and insurance underwriting, or in core trading, mainly because firms judge that gaps in interpretability and predictability make the risks outweigh the gains. Fed Governor Lisa Cook described AI in use for compliance, call centers, back offices, coding and cyber scanning.

Cost and consequence

Barnum said on JPMorgan's July 14 call that token expense was trivial in the first half and that he expects meaningful acceleration in the second. Routing, provider redundancy and open-source fallback limit exposure to price rises and withdrawals; each is also a standing engineering cost.

03Capital

Fintech funding concentrated as banks financed the AI build-out

Venture money into fintech grew while the number of deals fell. For banks, AI also arrived on the lending side, where financing the data-center build-out became a supervisory concern.

$28.6B

Global fintech venture funding, first half of 2026

Up 22.7% from a year earlier and down 17.3% from the second half of 2025.

1,605

Fintech venture deals, first half of 2026

Down 25.7% from 2,161 a year earlier.

Largest rounds

The quarter's largest named fintech rounds

Crunchbase named Ramp's early-June raise and Taktile's June Series C, led by Goldman Sachs Alternatives, among the largest; Flutterwave raised an undisclosed Series E at a $3.2 billion valuation. GV's Elena Sakach described money flowing either to brand-new companies or to a few large incumbents.

$750MRamp, at a $44B valuation

$110MTaktile Series C

Global context

AI funding concentrated in large rounds

CB Insights' count shows AI money concentrated in the largest rounds, the same pattern as fintech's rising total on falling deal numbers.

$149.5BAI equity funding worldwide, second quarter

89%Share raised in rounds of $100M or more

Bank lending

Banks finance the AI build-out

Goldman Sachs CEO David Solomon said on the July 14 call that AI's needs for infrastructure, energy and data centers are creating structuring and financing business, at an early stage and with setbacks possible. Dimon's April letter put hyperscaler AI spending at about $450 billion in 2025 and about $725 billion in 2026.

$225BHyperscaler-related bonds issued through midyear

About 15%Direct-lending fund portfolios in AI and IT, BIS estimate

Supervisors

AI debt named as a stress channel

The FPC said U.S. AI valuations remained particularly stretched and that a sharp repricing could tighten U.K. funding conditions, and Bank of England Deputy Governor Sarah Breeden said AI-related debt could amplify a fall. Cook said fears of AI disruption had widened speculative-grade technology spreads and pressured business development companies, though leverage remains below its peaks before the global financial crisis.

Bank budgets

AI spending sits inside technology budgets

Chief Operating Officer Jennifer Piepszak's April letter gives JPMorgan's total technology budget and says gains will be reinvested, with fewer staff not the goal. HSBC's figure, reported by Evident, is a target for value delivered, not spending.

$19.8BJPMorgan technology budget, 2026

$100MHSBC-Google Cloud value target over two years

What it signals

For banks, AI now sits on both sides of the balance sheet: as technology spending inside total budgets, and as credit extended to the build-out. The second is a lending decision, made under credit limits and now watched by U.S. and U.K. supervisors as a stability question.

04Products and delivery

Payment rails for software agents, mostly in pilot

Product activity concentrated in payment infrastructure that lets software agents buy for a customer, and in assistants and in-house models at large banks. None of the agent-payment launches disclosed transaction volumes, and most were pilots or announcements; Stripe's agent wallets and Klarna's ChatGPT app were available at launch.

Payments

Agent payments stay on card authorization

Visa offers credentials, tokenization and risk controls as the trust layer for agent payments, and Stripe's agent wallet issues a one-time card for each task, with the user approving every payment. Both designs keep card authorization, and its fee, inside transactions that software agents initiate.

Integration

Catalog servers and protocol gateways

Klarna's ChatGPT app searches more than 100 million products and 400 million merchant listings in 13 markets through a Klarna server built on the Model Context Protocol (MCP), which connects assistants to data and tools. Visa's Intelligent Commerce Connect puts four agent protocols (TAP, MPP, ACP and UCP) behind one integration.

Inside banks

Assistants and in-house models at large banks

Piepszak's April letter says JPMorgan recently rolled out Employee Assistant, a personalized AI agent for staff, alongside its in-house LLM Suite. HSBC's Google Cloud partnership uses Google's forward-deployed engineers, and Nubank developed its own AI model, NuFormer, to extend credit in Brazil, Evident reported.

Status of agent-commerce launches in the quarter

Status as described in each company's announcement.

Status of agent-commerce launches in the quarter
Launch AnnouncedStatus at the end of JuneWhat it does
Visa Intelligent Commerce Connect Announced April 8 Status at the end of June Pilot; wider rollout planned for 2026 Partners include Aldar, AWS, Diddo, Highnote, Mesh, Payabli and Sumvin What it does One integration across networks and four agent protocols
Stripe Link wallets for agents Announced April 29 Status at the end of June Live from launch day What it does One-time card per task; user approves each payment
Stripe Agentic Commerce Suite additions Announced April 29 Status at the end of June Availability not stated What it does Adds Google, Wix, BigCommerce and WooCommerce
Klarna Shopping Search in ChatGPT Announced May 20 Status at the end of June Available; purchase completes on merchant site What it does Product search with sponsored placements
Visa and OpenAI agentic commerce Announced June 10 Status at the end of June Future tense; no date or markets What it does Visa payments on OpenAI's platform with user-set limits

For integration teams

A bank or merchant connecting to agent commerce now faces several protocols, spending limits, tokenized credentials and catalog exposure through MCP, plus open questions on how payments rules apply to agent-initiated payments: the FPC asked for further work on agentic AI in payments, and HM Treasury plans to consult.

05Adoption and economics

AI use by segment, and where the savings are expected to land

Large banks now report breadth: staff on tools, prompts per day and use cases by stage. Community banks report near-universal use that is mostly vendor-embedded or in pilot, and the latest insurer surveys date from 2022 to 2025. Each row comes from a different publisher and population, so the rows are not comparable with one another.

What each segment reports about AI use

Bank figures come from second-quarter results calls in mid-July and coverage of them; survey figures are as published.

What each segment reports about AI use
Segment and source PopulationReported useDepth of deployment
JPMorgan Population More than 300,000 employees Reported use About 150,000 weekly users of its internal LLM Depth of deployment Almost 1,000 use cases; status not given
Bank of America Population — No comparable figure Reported use More than 200,000 staff; 400,000+ prompts a day Erica: more than 24 million active users Depth of deployment 300+ approved; 114 live generative; 34 fully implemented
Citigroup Population — No comparable figure Reported use Nearly 9 in 10 staff use its AI tools Depth of deployment — No comparable figure
Community and regional banks (Wolf & Co.) Population 20 respondents at 17 banks, $750M-$25B in assets Fielded Jan. 31 to Feb. 28; published May 1 Reported use 100% use AI; 30% only ad hoc or vendor-embedded Depth of deployment 5% of active adopters run a scaled, governed program 40% in pilots; 25% with proofs of concept in production
U.S. insurers (NAIC surveys) Population Auto, home, life and health insurers Reported use Auto 88%, home 70%, life 58%, health 92% Surveys from 2022 (auto), 2023 (home, life) and May 2025 (health) Depth of deployment Counts use, planned use or planned exploration
U.S. businesses (Visa and Morning Consult) Population 512 U.S. business decision-makers Reported use 53% would let agents negotiate with other agents Depth of deployment Stated willingness, not use

— No comparable figure.

Large banks

Earnings-call language moved to the present tense

Across eight U.S. banks reporting by July 15, Evident found that present-tense statements about AI nearly doubled from a year earlier, while the future-tense share fell from about half to a fifth. Its “agentflation” gap, between agent mentions on calls and the agentic share of use cases, roughly doubled over the year.

Community and regional banks

Smaller banks get AI through processors

Dimon said on the July 14 call that AI tools will reach smaller competitors through Fiserv, FIS and other fintechs, and 30% of Wolf & Co.'s respondents use AI only ad hoc or embedded in vendor products. Respondents ranked security and data leakage (55%), expertise (50%), vendor AI risk (40%) and data quality (40%) ahead of unclear return (30%) as barriers.

Insurers

Insurer use data date from 2022 to 2025

Twelve states were piloting the National Association of Insurance Commissioners' AI Systems Evaluation Tool as of March, with adoption anticipated at the 2026 Fall National Meeting. The FPC found little evidence of advanced AI in core insurance underwriting.

Wealth and capital markets

Advisers first; trading evidence is supervisory

Bank of America's chief financial officer credited adviser productivity largely to AI, and the U.K. Financial Conduct Authority (FCA) is testing AI-enabled targeted investment support. For trading, the quarter's evidence came from supervisors: the FPC found little evidence of advanced AI in core trading, and Cook listed correlated trading, model collusion, manipulation and concentration among algorithmic-trading risks.

How banks described AI's effect on jobs and costs

Positions as stated at investor days, to staff, in analyst research and on second-quarter results calls. Headcount changes for U.S. banks are Evident's count of eight banks reporting by July 15.

How banks described AI's effect on jobs and costs
Bank or forecaster Stated positionFigureDate
Standard Chartered Stated position Fewer corporate-function roles, higher income per employee Figure About 7,800 roles by 2030 More than 15% of corporate-function roles; income per employee up about 20% by 2028 Date May 19
HSBC Stated position Staff told not to resist; cuts framed as productivity-led Figure About 20,000 cuts committed Cuts as reported by TNW; HSBC has more than 211,000 staff, per Reuters Date May 20
Morgan Stanley analysts Stated position Forecast for European banks Figure Up to 20% of jobs, about 400,000, by 2030 Doubled from about 10% in January Date May
JPMorgan Stated position Capacity and redeployment; savings go to customers Figure 30% to 40% fewer jobs in discrete areas; headcount up 1% Headcount over the past year, per Evident Date July 14
Goldman Sachs Stated position AI raising productivity; no structural headcount rework now Figure — No comparable figure Date July 14
Bank of America Stated position Praised headcount discipline over six quarters Figure — No comparable figure Date July 14
Wells Fargo Stated position — No comparable figure Figure Workforce down 7% Over the past year, per Evident Date Second quarter

— No comparable figure.

Economics

Dimon said on the July 14 call that because rivals deploy the same tools, the benefit of AI goes to customers, and that margins would be 80% if technology savings stayed with banks. Standard Chartered's income-per-employee target assumes the bank keeps more of the gain. A budget built on AI savings should state which of the two it expects.

06Outcomes

Reported results: efficiency, revenue per employee and fraud defense

The quarter's reported results came from companies, an industry body and a Fed governor. None is an independent evaluation, and most do not isolate AI's share of the result.

Bank of America

Efficiency ratio of 59%

Chief Financial Officer Alastair Borthwick said new AI capabilities used by more than 200,000 employees helped contribute to the ratio; on the call he tied the improvement to revenue growing faster than expenses. CEO Brian Moynihan said the bank's developers code more efficiently.

59%Efficiency ratio, second quarter

63%A year earlier

Klarna

Revenue per employee near $1.4 million

Klarna reported first-quarter revenue of $1 billion, up 44%, and adjusted operating profit of $68 million. The release does not name AI as the cause and makes no statement on AI staffing.

Nearly $1.4MRevenue per employee, first quarter

4xThe 2022 level

Federal Reserve

A Fed classifier cut computing costs

Cook said a small text classifier built by Fed staff matched a larger model's accuracy at up to 80% lower computation cost; she did not give the period or workload.

Up to 80%Lower computation cost than a larger model

UK banks

Unauthorized fraud prevented in 2025

UK Finance reported the prevention figure across all methods, without attributing it to AI.

£1.68BUnauthorized fraud prevented

70pPrevented per £1 attempted

Visa

Device-token fraud declined

Visa compared the period with a year earlier and credited authentication and network controls.

9.6%Fall in device-token fraud, July to December 2025

Stripe

Sign-up screening at AI businesses

Stripe said on April 29 that Radar blocked the sign-ups at eight AI businesses in one month, and that free-trial abuse at AI services on its platform more than doubled in six months. The figures are Stripe's own.

3.3MRisky sign-ups blocked in a month

1 in 6AI-service sign-up attempts from bad actors

Reading the results

Each figure blends AI with pricing, hiring and other controls over the same period, so none sets a benchmark for a specific process. The fraud figures share a direction: defenses against unauthorized fraud improved while losses moved to scams that customers authorize.

07Corrections

Four widely repeated claims, corrected

Each claim circulated in coverage of the quarter. The corrected version changes what a reader should plan for: the base of a job cut, the scope of a rule, the unit of a reduction and the status of a product.

Headcount · Wrong base

Standard Chartered's target is about 7,800 corporate-function roles

Claim under review: Standard Chartered will cut 15% of its workforce by 2030.

On the bank's own figures, reported by TNW and Reuters, more than 15% of corporate-function roles, about 7,800 jobs, will be cut by 2030. That is roughly 9% of a workforce of about 83,000. Semafor's headline applied the percentage to the whole workforce.

Model risk · Overstated

The guidance excludes generative and agentic AI and is nonbinding

Claim under review: The new U.S. model-risk guidance governs banks' use of AI.

SR 26-2 and OCC Bulletin 2026-13 place generative and agentic AI outside their scope, expect it to be relevant mainly to banks with more than $30 billion in assets, and state that noncompliance will not draw supervisory criticism. Findings that a practice is unsafe or unsound remain possible.

Jobs · Missing scope

The reductions were in discrete areas, and firm headcount rose

Claim under review: AI has eliminated 30% to 40% of JPMorgan's jobs.

Dimon told analysts on July 14 that jobs fell 30% to 40% in discrete areas and that most affected staff were offered roles elsewhere. Evident's tally of second-quarter results put JPMorgan's headcount up 1% over the past year.

Agent payments · Premature

The Visa-OpenAI collaboration was announced in the future tense

Claim under review: Visa and OpenAI have launched agent payments.

Visa's June 10 release describes payments for agentic commerce across OpenAI's platform, with user-set limits and tokenized credentials, but gives no date or markets. Visa's Intelligent Commerce Connect, announced April 8, is in pilot.

08Failures and accountability

Authorized scams, a withdrawn model and thin governance

The quarter's best-documented failures came from outside banks' own models: scams that persuade customers to pay, a supplier's models withdrawn by government order, and governance capacity that trails deployment. Fraud data and agent-security research describe one mechanism: untrusted input persuading an authorized actor (a customer or a software agent) to act.

Fraud

UK scam losses rose as unauthorized fraud fell

UK Finance counted £576.4 million of authorized push-payment (APP) losses across 248,070 cases, 66% of them starting online, and £703.4 million of unauthorized losses. A foreword by fraud vendors BioCatch and Nasdaq Verafin says criminals are “deploying deepfakes during account openings.” The report calls agentic AI likely the next force multiplier for criminals.

19%Rise in authorized push-payment losses, 2025

5%Fall in unauthorized fraud losses, 2025

Liability

Most in-scope scam losses were reimbursed

UK Finance reported the share for the first year of the Payment Systems Regulator's reimbursement scheme. Visa counted nearly $1 billion of scam activity, its largest consumer payment-fraud category, from July to December 2025.

88%In-scope APP losses reimbursed in the scheme's first year

Supply

Vendor-model contingency lost its explicit expectation

Sullivan & Cromwell noted that the revised U.S. guidance no longer explicitly asks for contingency plans for vendor models. In the same quarter, a government order removed two models, including the generally available Fable 5, from every Anthropic customer, leaving continuity planning for third-party models to each bank's judgment.

Governance

Governance capacity trails deployment

In Bank Director's survey released March 31 and sponsored by Baker Tilly, a third of bank leaders said they do not understand agentic AI at all, while 84% were concerned about AI-enabled fraud against customers. In a speech that did not address AI, Fed Governor Michael Barr said matters requiring attention at the largest banks had fallen by the end of 2025 to about half their 2024 level.

Agent security

Agent tooling became an attack surface

The Cloud Security Alliance documented a command-execution default in the official MCP software development kits, with an estimated 200,000 vulnerable instances, and “agentjacking,” in which instructions planted in error-tracker events hijacked coding agents in 85% of Tenet Security's tests. Epoch AI counted about 1,500 high or critical vulnerabilities disclosed in June. Breeden called for fast patching and recovery options for core systems.

Who is accountable

In each case the firm answers for the outcome: a customer it must protect authorized the scam, a supplier it chose withdrew the model, and the guidance that might have set agent controls leaves them to the firm. The controls that matter most sit where a customer or a software agent authorizes an action.

09Regulation

US supervisors narrowed model-risk guidance as UK authorities moved to testing

U.S. bank supervisors withdrew prescriptive model-risk expectations and left generative and agentic AI to general governance pending a request for information. U.K. authorities kept technology-agnostic rules but added simulations, live testing and doubt that human review of every action can govern agents. In the EU, both legislators approved moving the high-risk date for credit scoring and for life and health insurance pricing to Dec. 2, 2027, but Aug. 2, 2026, remained the legal date until publication.

United States · Issued April 17

SR 26-2 replaces SR 11-7

The Fed, OCC and FDIC guidance supersedes SR 11-7, OCC 2011-12 and the 2021 BSA/AML model statement, and the OCC rescinded its 1997 bulletin on credit-scoring models. Footnote 3 sets generative and agentic AI outside its scope: “Generative AI and agentic AI models are novel and rapidly evolving.” Sullivan & Cromwell noted it drops the expectation of at least annual validation and annual board approval and was adopted without notice and comment.

United States · Promised April 17; not issued by June 30

AI model-risk request for information pending

The agencies said they plan to issue in the near future a request for information on model risk management that considers banks' use of AI, including generative and agentic AI. Fed Vice Chair for Supervision Michelle Bowman said supervision should prioritize matters that lead to bank failure and that third-party risk guidance is being simplified; she expected a Financial Stability Board consultation on AI sound practices in the third quarter.

Colorado · Repealed May 14

Colorado repealed its AI Act before it started

The law, which covered consequential decisions including lending, was due to take effect June 30; the Justice Department had joined xAI's lawsuit against it in April. A narrower automated-decision law, whose covered domains include lending and insurance, starts Jan. 1, 2027, with enforcement by the attorney general only, and federal preemption remained a discussion draft.

United Kingdom

Bank of England simulates; FCA tests with firms

Breeden's April 1 letter to the Treasury Committee set out scenario analysis, herding studies, a fourth Bank of England-FCA AI survey and work building on the SS1/23 model-risk principles, and said no AI provider had yet been designated a critical third party. The FCA's second live-testing cohort, including Barclays, Lloyds, UBS and Experian, tests agentic payments, anti-money-laundering and know-your-customer uses through 2026; no AI-specific rules are planned.

United Kingdom · Speech, June 30

Doubt that people can approve every agent action

Breeden said technology-agnostic frameworks may not suffice, that a human in the loop for every action is unlikely to be realistic, and that circuit breakers or kill switches are under consideration. HM Treasury plans to consult on payments regulation for AI agents.

European Union · Aug. 2, 2026, remained the legal date

AI Act high-risk date for credit scoring

Annex III covers creditworthiness assessment and credit scoring of individuals, except systems used to detect financial fraud, and risk assessment and pricing in life and health insurance. Parliament adopted the omnibus that moves these obligations to Dec. 2, 2027, on June 16, and the Council gave final approval at the end of June; the change takes effect three days after publication in the Official Journal, which was still pending at quarter end.

Compliance calendar at the end of June

Scheduled, expected and pending items published by June 30.

Compliance calendar at the end of June
Item DateJurisdictionStatus
AI Act high-risk obligations, including credit and insurance pricing Date Aug. 2, 2026 Jurisdiction EU Status Legal date; delay to Dec. 2, 2027, awaits publication
FSB consultation on AI sound practices Date Third quarter of 2026 Jurisdiction International Status Expected
NAIC AI Systems Evaluation Tool Date 2026 Fall National Meeting Jurisdiction U.S. states Status Adoption anticipated after a 12-state pilot
FCA AI Live Testing, second cohort Date End of 2026 Jurisdiction U.K. Status Evaluation report in the first quarter of 2027
Colorado automated-decision law Date Jan. 1, 2027 Jurisdiction Colorado Status Enacted May 14
Request for information on model risk, including AI Date Not dated Jurisdiction U.S. Status Promised April 17; not issued by June 30
Simplified third-party risk guidance Date Not dated Jurisdiction U.S. Status In preparation
HM Treasury consultation on payments rules for AI agents Date Not dated Jurisdiction U.K. Status Forthcoming
Fourth Bank of England-FCA AI survey Date 2026 Jurisdiction U.K. Status Announced
FCA report on good and poor AI practice Date Later in 2026 Jurisdiction U.K. Status Planned

Practical effect

For a U.S. model-risk team, the change means more discretion and less cover: nothing in the guidance says how to validate an agent, and findings that a practice is unsafe or unsound remain possible. For EU lenders and life and health insurers, Aug. 2, 2026, remained the legal date at the end of the quarter, with an approved move to Dec. 2, 2027, awaiting publication.

10Voices

Supervisors, bank leaders and technology buyers on pace, savings and oversight

Positions converged on three points: back-office, risk and compliance work changes first; fraud is the most immediate harm from AI; and governing agents remains an open problem for supervisors and banks. They diverged on the pace of job cuts, on who keeps the savings and on how supervisors should respond. Bank executives, Visa and Evident speak from commercial positions.

US supervisors

Fed officials differed on the direction of supervision around the new guidance.

Michelle W. Bowman

Vice Chair for Supervision, Federal Reserve Board

“The revised guidance now applies narrowly to traditional models and basic AI applications.”

She argued that guidance should not stand in the way of new tools and that smaller banks need flexibility to adopt AI.

Regulator · speech, April 27

Michael S. Barr

Governor, Federal Reserve Board

The wider retreat in supervision, including grade inflation in large-bank ratings, fewer matters requiring attention and cuts to staffing and horizontal reviews, weakens supervisors' ability to detect risk.

His speech did not address AI. It dissents from the broader deregulatory direction of which the model-risk change is part.

Regulator · speech, June 6

Lisa D. Cook

Governor, Federal Reserve Board

“alongside experimentation, strong governance and risk management must be our foundation”

She cited a Fed experiment in which AI agents followed market trends less than humans did, possibly meaning fewer bubbles driven by animal spirits.

Regulator · speech at Stanford, May 27

UK supervisors

Sarah Breeden

Deputy Governor, Bank of England

“AI is reshaping finance at speed.”

She doubted that a human can approve every agent action, warned that agents responding alike to the same triggers could amplify volatility in stress, and said the Bank is experimenting with simulation methods to see which agent designs drive herding.

Central bank · ECB Forum on Central Banking, June 30

Bank leaders

Bank leaders agree AI will change jobs. They differ on how fast, and on who keeps the gain.

Jamie Dimon

Chairman and CEO, JPMorganChase

“AI will definitely eliminate some jobs, while it enhances others.”

On the July 14 results call he said the benefit of AI goes to customers because rivals adopt the same tools, and he declined to project its effect on expenses.

Bank leader · shareholder letter, April 6

Bill Winters

Chief Executive, Standard Chartered

Winters presented the plan as replacing lower-value roles with machines and investment, alongside a target to raise income per employee by about 20% by 2028.

Standard Chartered is one of the European global banks with cost-to-income targets that attached headcount numbers to AI.

Bank leader · investor day, May 19, as reported by TNW and PYMNTS

Georges Elhedery

Chief Executive, HSBC

Elhedery told staff not to resist AI, saying generative AI will destroy some jobs and create new ones.

Bank leader · remarks reported by Reuters, May 20

Denis Coleman

Chief Financial Officer, Goldman Sachs

AI is raising productivity, and Goldman is not structurally reworking its headcount now.

His account of added capacity is the U.S. counterpoint to the European banks' headcount targets.

Bank leader · results call, July 14

Technology buyers inside banks

The people who buy models describe redundancy and cost as design requirements.

Teresa Heitsenrether

Chief Data and Analytics Officer, JPMorganChase

JPMorgan cannot over-rely on a single model provider.

JPMorgan was one of two banks that cut Hong Kong-based staff off from Anthropic models after the June directive.

Technology buyer · as reported by Evident, June 25

Mike Pizzi

Global Head of Technology and Operations, Morgan Stanley

Build on the best model first, then move tasks to the most efficient one, keeping the architecture open while accepting some lock-in where it speeds delivery.

Technology buyer · as reported by Evident, June 25

Payment network

Michael Jabbara

Senior Vice President, Payment Ecosystem Risk and Control, Visa

“The rapid adoption of AI has fundamentally lowered the barrier to entry for fraud.”

Visa also sells the credentials, tokenization and risk controls it offers as the trust layer for agent payments.

Vendor · Visa Threats Report release, May 20

Independent benchmark

Evident

Banking AI benchmark and newsletter

“Banks' actions still haven't caught up with their words”

Its newsletter compared eight U.S. banks' second-quarter calls with the agentic share of their use cases.

Commercial benchmark · Banking Brief, July 16

11Decisions

Three decisions for risk, technology and finance leaders

Chief risk officer

Set the standard for AI the guidance leaves out

SR 26-2 leaves generative and agentic systems to general governance and drops explicit validation cadences and vendor-contingency expectations, while unsafe-or-unsound findings remain possible. Banks under $30 billion in assets face guidance expected to matter less to them, yet receive much of their AI inside processor and vendor products.

Which generative and agentic systems would we have to defend in an examination, and what validation and fallback evidence do we hold for each?

Chief technology officer

Decide how much provider redundancy to fund

A government order took two Anthropic models offline, one of them generally available, two large banks cut Hong Kong staff off from one provider, and flagship prices rose. Routing, multiple providers and switching software reduce that exposure, and each adds standing engineering cost.

If our main model provider became unavailable to some or all staff tomorrow, which workflows would stop, and how quickly could we move them?

Chief financial officer

Decide where AI savings are expected to land

Standard Chartered plans around fewer roles and higher income per employee; JPMorgan and Goldman Sachs describe added capacity; Dimon expects competition to hand savings to customers. JPMorgan expects token expense, a new variable cost, to accelerate in the second half.

Does our plan count AI gains as margin, capacity or lower prices, and what in our competitive position supports that choice?

12Actions

Four actions for the coming quarter

  1. 1

    Split the AI inventory by guidance scope

    List statistical models inside SR 26-2 separately from generative and agentic systems outside it, including AI embedded in processor and vendor products, and record an owner, a validation approach and a fallback for each. The result is a register an examiner or board can read.

    Model risk

  2. 2

    Run a provider-withdrawal drill

    Simulate losing one model provider: test model aliases, routing and fallbacks, as Greenberg Traurig advised after the June directive, along with access controls by staff location. The result is a measured failover time and a list of workflows with no fallback.

    Technology

  3. 3

    Separate authorized-scam losses and untrusted agent inputs

    Report losses on payments the customer authorized separately from unauthorized fraud, by channel, and treat tickets, logs and other inputs to software agents as untrusted. The result is a clear view of the loss category that grew 19% in the U.K. and of the injection paths documented this quarter.

    Fraud and payments

  4. 4

    Give token cost its own budget line

    Report model and token spend by workflow, with the routing rules that send simple tasks to cheaper models. The result is a cost per completed task in hand before second-half token growth, of the kind JPMorgan expects, reaches the budget.

    Finance

13References

References

Covers U.S., U.K. and EU developments from April 1 to June 30, 2026, including second-quarter results calls held in mid-July. Market-data and research-platform vendor launches are not covered in this edition.

Supervision and regulation

Models, cost and security

Bank results and statements

Capital and surveys

Thoughtful engineering for what’s next.

Have a question about what this means for your team? Let’s talk it through.

Start a conversation

Have a project in mind, or just exploring an idea? Let’s talk.
Fill out the form or give us a call.

413-570-3299

What can we help you with?