Large financial firms moved AI agents into back-office operations in Q1 2026

Banks, insurers and card networks spent the first quarter assigning AI agents to trade accounting, client onboarding, claims intake and payments, mostly in testing or narrow deployments. For operations, risk and technology leaders, the open questions concern what those agents can touch, how supervisors will see their work and where the savings go.

Editorial snapshot: 30 April 2026

The central finding

Large banks and insurers reported AI gains as capacity, slower hiring and efficiency targets rather than lower prices, while U.K. and EU supervisors began shifting oversight from validating individual models toward monitoring agents and the few vendors behind them.

01Key findings

What changed in financial services AI in Q1 2026

Agents became practical for narrow back-office tasks, incumbents paid for them from technology budgets, the gains appeared as capacity and attrition, and supervisors began asking how such systems are monitored and stopped.

Capability

Agents moved into back-office work

Goldman Sachs confirmed six months of work with embedded Anthropic engineers on agents for trade accounting and client onboarding, still in testing. BNY runs more than 130 "digital employees," and Travelers launched a generative AI claims voice agent in January.

Capital

Fintech deals fell as incumbents bought

CB Insights counted 762 global fintech deals in the first quarter, a multiyear low, while Capital One's $5.15 billion Brex deal led fintech M&A. JPMorgan raised its 2026 technology budget to about $19.8 billion in February.

Adoption and economics

Gains were booked as capacity and attrition

Bank of America's efficiency ratio improved to 61% in the first quarter as headcount fell by about 1,070 through attrition. None of the banks or insurers cited here that quantified AI gains described passing them on to customers as lower prices or fees.

Accountability

Supervisors turned to monitoring agents and vendors

Bank of England roundtables recorded firms' view that traditional model validation is unsustainable for generative AI and agents. The ECB told banks that accountability for AI decisions stays with them, and U.K. lawmakers called for AI stress tests.

Operating question

For each process where an agent now acts, can the firm show a supervisor what the agent did, who approved it, how it would be stopped and where the time it saved went?

02Capability

Agents for accounting, onboarding, payments and claims

Frontier releases in the quarter packaged agents as products: Anthropic's Claude Opus 4.6, with agent teams in Claude Code as a research preview, and OpenAI's Frontier agent-management platform on Feb. 5, then OpenAI's GPT-5.4 with native computer use on March 5. Financial firms applied the capability to narrow, rule-heavy steps. Goldman Sachs built its agents alongside a lab's engineers and BNY on its own multi-model platform, instead of buying finished products.

Banking operations · Testing

Goldman Sachs' accounting and onboarding agents

Goldman confirmed a Feb. 6 CNBC report of six months' work with embedded Anthropic engineers on agents for trade and transaction accounting, client due diligence and onboarding. CIO Marco Argenti said the agents are in early development and should cut processing time significantly. No launch date was given.

Custody and payments · Deployed in 2025

BNY's "digital employees"

BNY's agents, which it calls digital employees, validate payment details and remediate code vulnerabilities. They run on Eliza, the bank's model-agnostic platform, which now integrates Google's Gemini Enterprise.

130+Digital employees deployed in 2025

Insurance · Launched in January

Travelers' claims and underwriting agents

Travelers launched a generative AI voice agent for first notice of loss in mid-January, alongside agents that mine internal and external data for commercial underwriting. Underwriting preparation and claims service are also where it reports measurable results.

Capital markets · Vendor data

Automated trading in one lab's API traffic

Anthropic's March 24 Economic Index report found that automated-trading workflows at least doubled their share of its API traffic, in a sample taken in early February. The figure describes one vendor's customers and says nothing about trading volumes.

Benchmark

Hours-long professional tasks still mostly fail

In APEX-Agents, a benchmark built by expert-data vendor Mercor from investment banking, consulting and corporate law tasks that take professionals hours, frontier agents completed fewer than 25% of tasks. The result is consistent with firms confining agents to narrow, checkable steps.

<25%Tasks completed by frontier agents

Cost and consequence

The chosen tasks have clear rules and checkable outputs, which suits agents that still fail most open-ended work. The larger cost is integration: agents need governed access to ledgers, policy, claims and know-your-customer systems. Firms report AI costs inside wider technology budgets, so the cost of a completed task is not visible from outside.

03Capital

Fintech deal counts fell while incumbents bought and spent

Global venture funding set a record in the first quarter: Crunchbase counted about $300 billion, with AI companies taking about 80% and four rounds, led by OpenAI's $122 billion, taking 65%. Fintech moved the other way, and for financial firms the larger capital story sat in three other places: incumbents' technology budgets and acquisitions, the repricing of data vendors and lenders exposed to software, and the financing of AI's build-out.

Disclosed fintech transactions

Largest disclosed fintech deals in Q1 2026

Values in USD millions. The acquisitions are the two largest fintech M&A deals CB Insights counted in the first quarter; the funding rounds are the three largest raises in its banking category. Fintech M&A deals fell 26% from the fourth quarter to 199, a six-quarter low.

Capital One–Brex

Acquisition — $5.15B

Pending; announced Jan. 22

Mastercard–BVNK

Acquisition — $1.8B

Pending

Uala

Funding round — $195M

$3.2B valuation

Allica Bank

Funding round — $150M

$1.2B valuation

Anchorage Digital

Funding round — $100M

$4.2B valuation

CB Insights, State of Fintech Q1'26 public summary; Capital One release of Jan. 22. Neither acquisition had closed by quarter end; Capital One expected the Brex deal to close in mid-2026.

762

Global fintech deals in Q1 2026

A multiyear low; deal counts fell in seven of the last eight quarters.

$932M

Banking-fintech funding across 34 deals

Down from $1.8 billion in the first quarter of 2025. Funding for core-banking and onboarding providers fell.

What large firms disclosed about technology and AI spending

Each firm disclosed a different measure, so the figures are not comparable with one another.

What large firms disclosed about technology and AI spending
Firm What was disclosedAmountPeriod
JPMorgan Chase What was disclosed 2026 technology budget Amount About $19.8 billion Up almost $2 billion; total 2026 expense guidance about $105 billion Period Company update, February 2026
Wells Fargo What was disclosed Incremental technology expense Amount About $1.1 billion Period 2026 guidance, Jan. 14
Bank of America What was disclosed AI spending Amount Several hundred million dollars Period 2026, stated Jan. 15
Travelers What was disclosed AI and other technology initiatives Amount More than $1.5 billion Period 2025
BNY What was disclosed Cyber resiliency compared with AI Amount Cyber spending exceeds AI spending Period Stated in January

Public markets

Data vendors repriced on Feb. 3

After Anthropic released its Cowork plug-ins, investors marked down market-data and analytics companies. The London Stock Exchange Group lost between 6% and 13%, depending on the report, and MSCI about 7% in one session.

−10.5%FactSet, Feb. 3

−9%Morningstar, Feb. 3

Private credit

Alternative managers marked down on software lending

Apollo, Ares, Blackstone, Blue Owl, Carlyle and KKR fell 3% to 11% as Oppenheimer analysts linked software weakness to credit problems, then recovered 0.2% to 5%. On March 27 the Bank of England's Financial Policy Committee flagged private-credit exposure to software firms that AI could disrupt.

AI financing

Bond markets and bank channels are financing the build-out

Alphabet sold $20 billion of notes in February, and Morgan Stanley estimated $400 billion of hyperscaler bond issuance in 2026. OpenAI's round of $122 billion in committed capital, which closed March 31, included more than $3 billion from individuals through bank channels. Fed Governor Michael Barr cited estimates of about $1 trillion of new AI debt over five years.

What it signals

In financial services, AI spending sat mainly inside incumbents' technology budgets, while startup deal counts fell and incumbents made the largest fintech acquisitions. Financial firms also carry AI exposure as private-credit lenders to software companies and as channels for AI-lab equity. On March 26, Fed Governor Lisa Cook used an AI disappointment as an example financial-stability scenario.

04Products

Four ways AI is reaching financial firms

The quarter's deployments arrived through four delivery models with different integration demands. The first three need governed access to ledgers, policy, claims and customer-identity systems; agent payments need issuers and merchants enrolled in network tokens.

Bank-owned platforms

Several models behind the bank's own platform

BNY runs its agents on Eliza, a platform built to use models from several labs, with Gemini Enterprise now integrated. Citi deployed proprietary AI tools in 84 countries and is re-engineering just over 50 of its largest processes.

Co-development

Lab engineers embedded in the bank

Goldman Sachs is co-developing agents with Anthropic engineers for specific accounting, due diligence and onboarding steps. The agents reportedly run on Claude Opus 4.6 and remain in testing. The approach gives Goldman agents designed around its own processes and makes the lab's engineers part of its development team.

Insurance workflows

Generative AI inside claims and underwriting steps

Travelers put generative AI into first notice of loss, submission intake, renewal summaries and commercial underwriting research. State Farm was named among early adopters of OpenAI Frontier, a platform for deploying and managing agents that launched in limited availability on Feb. 5.

Payments

Agent payments on tokenized card rails

Visa said Intelligent Commerce was live in the U.S. and its Central Europe, Middle East and Africa region and that it was working with more than 100 partners to enable it. Mastercard opened Agent Pay to U.S. issuers. Both build agent payments on network tokens, and nearly 40% of Mastercard's transactions were already tokenized.

Selected launches and deployment status, Q1 2026

Selected launches and deployment status, Q1 2026
Product or deployment DateStatusWhat it covers
Travelers first-notice-of-loss voice agent Date Mid-January Status Launched What it covers Initial claim reports by voice
Bank of America Microsoft 365 Copilot Date Stated Jan. 15 Status Deployed to about 200,000 employees What it covers Workforce productivity tools
Visa Intelligent Commerce Date Stated Jan. 29 Status Live in U.S. and CEMEA; pilots in Asia-Pacific, Europe What it covers Agent-initiated payments; 30+ partners in sandbox
Mastercard Agent Pay Date Stated Jan. 29 Status Open to U.S. issuers; global issuers targeted for end of Q1 What it covers Agent-initiated payments on network tokens
Lloyds customer AI assistant Date Announced Jan. 29 Status Planned for 2026 What it covers Savings, borrowing and investing
Klarna on Google's Universal Commerce Protocol Date Feb. 2 Status Announced What it covers Agent checkout across platforms
Claude Opus 4.6 Date Feb. 5 Status Released What it covers Agent teams in Claude Code (research preview)
OpenAI Frontier Date Feb. 5 Status Limited availability What it covers Agent management; State Farm an early adopter
Goldman Sachs–Anthropic agents Date Reported Feb. 6 Status Testing; no launch date What it covers Accounting, due diligence, onboarding
GPT-5.4 Date March 5 Status Released What it covers Native computer use

Statuses are as the companies described them during the quarter.

05Adoption

AI use at large firms and where the savings went

Surveys measure AI use under loose definitions, in different populations and years, so their figures cannot be pooled. Among large regulated firms use is close to universal, which makes the more useful comparison how firms describe the gains. They describe capacity, attrition and efficiency targets; none of the firms below tied AI savings to lower prices or fees.

How AI use was measured

Each row uses its own population, definition and date.

How AI use was measured
Population MeasureValuePeriod
U.K. financial firms (118 surveyed by the Bank of England and FCA) Measure Firms using AI Value 75% Insurers 95%; international banks 94% Period Published November 2024
Large euro-area banks supervised by the ECB Measure Banks using AI in some form Value More than 85% Period Latest annual ECB data collection
All U.S. employer firms (Census survey) Measure Firms using AI Value About 18% Period December 2025
U.S. businesses with more than 250 employees Measure Businesses using AI Value About 30% Period Cited by Fed Governor Barr, Feb. 17
Citi employees Measure Adoption of proprietary AI tools Value Above 70% More than 21 million interactions Period To January 2026
Bank of America employees Measure Daily AI access Value All 200,000 90 AI installations Period Q1 2026

Efficiency targets, headcount and how firms credit AI

Company statements on January earnings calls, Bank of America's first-quarter call and the HSBC and Klarna results releases in February.

Efficiency targets, headcount and how firms credit AI
Firm Efficiency measureTargetHeadcountWhat AI is credited with
Bank of America Efficiency measure Efficiency ratio 61% in Q1, from 63% Target 55% to 59% over the medium term Headcount Down about 1,070 in Q1, through attrition What AI is credited with Coding capacity worth about 2,000 people in 2025
Citi Efficiency measure Efficiency ratio 63.0% in 2025, adjusted Target About 60% in 2026 Headcount 226,000; reductions to continue What AI is credited with An additional source of efficiency
Wells Fargo Efficiency measure $15 billion of expenses cut over its program Target About $2.4 billion of gross cuts in 2026 Headcount 22 straight quarterly cuts, down more than 25% since Q2 2020 What AI is credited with More tools than ever, especially AI
BNY Efficiency measure Pretax margin 35% in 2025 Target 38%, raised from 33% Headcount Slight decline What AI is credited with Staff capacity; targets may not fully reflect it
HSBC Efficiency measure Cost growth 3% in 2025, target basis Target About 1% in 2026 Headcount — No comparable figure What AI is credited with Results release does not mention AI
Travelers Efficiency measure Expense ratio 28.5% in 2025 Target About 28.5% in 2026 Headcount Claim call-center staff down a third What AI is credited with Gains may go to profit or be reinvested
Klarna Efficiency measure Revenue up 104%, operating costs down 8%, since Q4 2022 Target — No comparable figure Headcount Down 49% since Q4 2022 What AI is credited with AI-enabled productivity, the company says

— No comparable figure.

Large and global banks

Headcount is the cost lever

Headcount is 60% to 70% of Bank of America's cost base, and the bank is reducing it through attrition while adding relationship managers. Citi paid about $800 million of severance in 2025 and Wells Fargo booked $612 million in the fourth quarter for 2026 actions; both present AI as one source of efficiency among several.

Community and regional banks

Adoption through vendors and voluntary controls

Treasury aimed its February AI resources at smaller institutions, and PNC CEO William Demchak, an executive member of the initiative, said institutions regardless of size are now positioned to use AI. Members of the Bank of England's AI Consortium expect smaller firms to adopt quickly, "fail fast" and rely more on third parties, which puts the weight on vendors' documentation.

Insurers

Heaviest AI users, with an exam tool in pilot

U.K. insurers were the heaviest AI users in the 2024 Bank of England and FCA survey, at 95%. In the U.S., 25 jurisdictions had adopted the NAIC model bulletin by April 1 and 12 states were piloting an AI evaluation tool, so governance files for third-party data and models are becoming exam material.

Payments and fintech

Agent payments kept inside card economics

Visa and Mastercard build agent-initiated payments as tokenized card transactions, which keeps them on existing rails and fee structures. Neither network gave volumes for agent-initiated payments, and Mastercard CEO Michael Miebach said the timing of scale is hard to predict.

Asset managers and capital markets

Software lending and broker-dealer agents

For managers that lend to software companies, underwriting now carries an AI-substitution question. Broker-dealers deploying agents remain under FINRA's supervision rule, and FINRA's 2026 oversight report, issued in December 2025, lists agents acting without human validation or beyond their intended scope among the risks.

Economics

So far the gains show up in efficiency ratios, margin targets and slower hiring, and payment networks are designing agent commerce to keep card-transaction economics intact. For employees the effect arrives as roles not backfilled and call centers consolidated, rarely as layoffs attributed to AI, which makes it hard to read from headline job cuts.

06Outcomes

Company-reported results in coding, underwriting and service

Each result below came from the company or vendor that deployed the tool, and none came from a controlled study. The most specific figures concern software delivery, underwriting preparation and service operations, where a unit of work can be timed. Citi's widely cited 100,000 hours a week of developer capacity was first reported in October 2025.

Software delivery · Bank of America

AI tools took 30% out of the coding stage

CEO Brian Moynihan described the saving in the coding stage of the bank's software change process and expressed it as a people-equivalent, which may mean capacity rather than job exits.

30%Taken out of the coding stage

About 2,000People-equivalent saved in 2025

Underwriting · Travelers

Renewal summaries and submission intake

Early results show generative AI summaries cutting average handle time in personal-lines renewal underwriting, and AI intake moved bond and specialty submissions from hours to minutes. Travelers gave no baseline period.

More than 30%Lower average handle time on renewals

Claims service · Travelers

Fewer claim call centers

Digital tools and straight-through processing reduced claim call-center staff. More than half of claims are eligible for straight-through processing, and customers take it up about two-thirds of the time. The result mixes AI with other automation.

One-thirdFewer claim call-center staff

4 to 2Claim call centers, 2026 plan

Whole company · Klarna

Revenue per employee since 2022

Klarna credits AI-enabled productivity; since the fourth quarter of 2022 its revenue rose 104% while operating costs fell 8%. The release does not separate AI's contribution from other causes.

$1.24MRevenue per employee, 3.6 times the 2022 level

−49%Headcount since Q4 2022

Fraud · Visa

Account-attack scoring in the U.S.

Visa said its Account Attack Intelligence scored and flagged transactions over 12 months. Separately, it reported more than $10 billion of fraud prevented in Latin America and the Caribbean in six months. The figures are vendor-reported and their definitions were not given.

60B+U.S. transactions scored

Nearly 600MFlagged as suspicious

Generative AI value · Lloyds

Lloyds' value estimate and 2026 target

Lloyds attributed the value to more than 50 generative AI solutions deployed in 2025 and did not define how value was measured.

About £50MValue in 2025

More than £100MExpected in 2026

What the results show

The reported gains are time and capacity on tasks with a countable unit: a code change, a renewal file, a claims call. None of these figures measures error rates, losses or customer outcomes, the measures a risk committee would need before agents take on more of each process.

07Corrections

Four widely repeated claims, corrected

Each of these claims circulated during the quarter. The distinctions change what a reader should conclude about adoption, jobs and market risk.

Adoption · Out of date

The 75% figure comes from a 2024 survey of 118 firms

Claim under review: Three-quarters of U.K. financial firms now use AI.

Reuters' coverage of the Treasury Committee's January report said about three-quarters of U.K. financial firms now use AI. The figure matches the Bank of England and FCA survey published Nov. 21, 2024: 75% of 118 responding firms, with insurers at 95% and international banks at 94%. It is not a 2026 measurement, and it should not be set against the roughly 18% U.S. all-firm rate, which covers a different population.

Jobs · Overstated

Goldman's agents are in testing for specific back-office steps, with no headcount figure

Claim under review: Goldman Sachs is using Anthropic agents to automate accounting and compliance roles.

Reporting on the Feb. 6 disclosure describes agents for trade and transaction accounting, client due diligence and onboarding, in early development with no launch date. Neither the bank nor the reporting gave a headcount figure, and CIO Marco Argenti reportedly said the effort is not about replacing workers, at least not at this stage.

Market data · Misattributed

The plug-ins behind the sell-off covered legal, sales, marketing and data analysis

Claim under review: Finance-specific AI tools caused the Feb. 3 fall in data-vendor shares.

Anthropic's Jan. 30 Cowork plug-ins covered legal, sales, marketing and data analysis, and Reuters named no finance plug-in. Investors marked down FactSet, Morningstar and LSEG on what the plug-ins implied for information products. JPMorgan software analyst Mark Murphy called the idea that a plug-in could replace every layer of mission-critical enterprise software an illogical leap.

Jobs · Unsupported

Neither Morgan Stanley nor the reporting attributed the cuts to AI

Claim under review: Morgan Stanley cut 2,500 jobs because of AI.

The roughly 2,500 cuts, about 3% of some 83,000 employees across all divisions with financial advisers excluded, were reported in March alongside a wider AI trend. Neither the firm nor the reporting attributed them to AI.

08Accountability

How AI deployments stall or fail in regulated firms

The quarter's most specific account of what goes wrong came from supervisors and their industry forums. They named mechanisms: validation that cannot keep up, vendor changes made at short notice, controls that can cause harm of their own and connectors that widen the attack surface. Standard U.S. liability forms, meanwhile, gained optional generative AI exclusions from January 2026.

Model validation

Traditional validation is unsustainable for agents

A Bank of England summary, published Feb. 16, of roundtables held late in 2025 recorded firms' view that traditional validation is unsustainable for generative AI and agents. Deployment was slowed by second-line caution, skills gaps, data protection impact assessments, data location rules and vendors' uneven familiarity with compliance requirements; insurers were limited by thin customer data.

Third-party risk

Vendor changes at short notice are a shared risk

At the Feb. 9 meeting of the AI Consortium run by the Bank of England and the FCA, members flagged changes that third-party vendors introduce at short notice as a possible source of risk correlated across firms. They treated prompts and retrieval layers as within the PRA's model risk principles, SS1/23, and said human-in-the-loop review may become strained as agentic AI spreads.

Operational resilience

Kill switches can cause their own disruption

Consortium members said agentic workflows can act across systems and compress decision time, which strains traditional escalation controls, and that a kill switch could stop an agent while also disrupting payments. Members questioned whether AI-driven trading poses risks beyond those of algorithmic trading, while noting AI's nondeterminism and scale.

Security

Connectors give agents an attack surface

In January, The Register reported a Claude Code prompt-injection bug that carried over to Anthropic's Cowork and allowed data exfiltration, and flaws in Anthropic's Git MCP server, fixed in December 2025, that allowed remote code execution. Agents granted connector permissions to a firm's systems can be turned against them.

Insurance cover

Liability forms can now exclude generative AI

Verisk's ISO forms added optional general liability endorsements excluding generative AI exposures, effective January 2026. A firm deploying agents may find that liability cover it relied on no longer extends to AI outputs.

Who is accountable

The firm that deploys an agent keeps responsibility for its decisions and may find them excluded from its liability cover. Supervisors and their forums point to evidence that sits outside traditional model validation: records of agent actions, second-line review before material uses, tested contingency options for critical cloud services and shutdown procedures that do not disrupt payments.

09Regulation

UK and EU expectations, US voluntary tools and state insurance rules

The U.K. moved furthest, by argument rather than rule: a House of Commons committee asked for stress tests and accountability guidance, and the FCA opened a long-horizon review without new rules. The ECB set supervisory expectations. U.S. federal action consisted of voluntary tools and speeches, and Fed Vice Chair for Supervision Michelle Bowman did not address AI in her Jan. 7 remarks or Feb. 26 Senate testimony. State insurance regulation was the exception.

UK · Parliament · Jan. 20

Treasury Committee calls for AI stress tests

The committee said regulators' wait-and-see approach risks serious harm, citing opaque credit decisions, exclusion, fraud, unregulated chatbot advice, reliance on a few U.S. technology firms and AI-driven herding. It asked for AI stress tests, FCA guidance by the end of 2026 on consumer rules and senior-manager accountability, and designation of major AI and cloud providers as critical third parties. No provider had been designated by quarter end.

UK · FCA · Launched Jan. 27

Mills Review of AI in retail financial services

The review covers agentic AI, competition across the AI value chain, consumers and how the regulator adapts, looking to 2030 and beyond. The FCA proposed no new prescriptive rules. The call for input closed Feb. 24, and a report to the FCA board is scheduled for the summer.

UK · Bank of England · March 27

Financial Policy Committee asks for agentic AI work

The committee saw little systemic use of agentic AI yet but judged that the risk could rise quickly, and asked the Bank and the FCA to examine agentic AI in payments and markets. It judged valuations of AI-focused U.S. technology companies to be particularly stretched.

EU · ECB · Feb. 24

ECB expectations for AI governance in banks

Pedro Machado said banks need clear accountability for AI-driven decisions and should include AI in risk appetite, review material uses before implementation and through the second line, monitor them and keep them explainable enough to challenge. On generative AI and cloud providers, he asked banks to reduce lock-in, keep tested contingency options for cloud services supporting critical functions and account for sub-outsourcing risks, linking these to the Digital Operational Resilience Act.

US · Treasury · Voluntary

A 230-control framework for financial firms

Treasury announced on Feb. 18 the conclusion of a public-private initiative and said it would release six voluntary resources in stages through February, aimed particularly at smaller institutions. They include an AI Lexicon and the Financial Services AI Risk Management Framework, published Feb. 19, which adapts NIST's AI framework into 230 control objectives with an adoption-stage questionnaire, developed with more than 100 institutions.

US · Insurance · As of April 1

NAIC bulletin in 25 jurisdictions

Twenty-four states and the District of Columbia had adopted the NAIC model bulletin on insurers' use of AI, and California, Colorado, New York and Texas have separate AI-specific insurance rules or guidance. Twelve states were piloting the NAIC's AI Systems Evaluation Tool as of March.

Policy and compliance dates known at quarter end

Policy and compliance dates known at quarter end
Item DateStatus at quarter end
White House national AI legislative recommendationsAsks Congress to preempt state AI laws that impose undue burdens, focusing on AI development, with carve-outs for general laws on fraud, consumer protection and child safety. Date March 20, 2026 Status at quarter end Proposal
Colorado AI Act (SB 24-205) takes effectThe working-group draft would require notice, explanation, correction and human review for "life-altering decisions." It was not enacted. Date June 30, 2026 Status at quarter end Scheduled; a March 17 draft would replace it
Mills Review report to the FCA board Date Summer 2026 Status at quarter end Scheduled
EU AI Act high-risk obligations Date Aug. 2, 2026 Status at quarter end In law
FCA guidance and designations of critical third partiesHM Treasury was gathering evidence on designations. Date End of 2026 Status at quarter end Requested by the Treasury Committee; none designated
EU AI Act stand-alone (Annex III) high-risk obligations, proposed dateTrilogue negotiations had not concluded. Date Dec. 2, 2027 Status at quarter end Backed by Council (March 13) and Parliament (March 26)

What changes for compliance teams

Beyond insurance, where the NAIC bulletin already applies in 25 jurisdictions, the dated obligations were Colorado's June 30 effective date and the EU AI Act's Aug. 2 date, both subject to proposals to change them. Elsewhere the quarter produced expectations and voluntary controls. Law firm Lowenstein Sandler expects examiners to ask firms using the Treasury framework for control mapping, logs and machine-readable vendor documentation.

10Voices

Where regulators, executives, vendors and investors agree and differ

The voices converge on three points: large regulated firms already use AI widely; governance, data and integration are the binding constraints; and dependence on a few AI and cloud providers is a risk. They differ on four questions. U.K. lawmakers want AI stress tests and guidance, while roundtable firms said AI-specific rules are not yet needed and the FCA proposed none. Bank of America, Travelers and Klarna describe headcount falling through attrition and restructuring, while Wells Fargo and BNY executives treat AI as one efficiency tool among several. Investors who sold data-vendor shares saw substitution, while a JPMorgan software analyst called wholesale replacement an illogical leap and an Ares executive called its exposure small. Mastercard says the timing of agent payments at scale is hard to predict and the Financial Policy Committee saw little advanced AI use yet, while Goldman, BNY and Travelers describe narrow deployments now.

Regulators and supervisors

Pedro Machado

ECB Banking Supervision

“AI does not dilute responsibility. If anything, it raises the bar.”

He set expectations for second-line review, monitoring and tested contingency options for cloud services behind critical functions.

Supervisor · speech, Feb. 24

Dame Meg Hillier

Chair, House of Commons Treasury Committee

She said she was not confident the financial system is prepared for a major AI-related incident.

Her committee's call for AI stress tests is the main point of disagreement with firms and the FCA, neither of which sought new AI-specific rules.

Committee chair · reported statement, Jan. 20

Michael Barr

Governor, Federal Reserve Board

“Society will need to be nimble and bold”

He said the evidence fits gradual adoption and that if AI disappoints, the balance of risks shifts from the labor market to the financial sector.

Central bank governor · speech, Feb. 17

Cory Wilson

Deputy Assistant Secretary, U.S. Treasury

“These resources are designed to help institutions, particularly small and mid-sized institutions.”

Treasury said six resources would be released in stages through February; they include the 230-control Financial Services AI Risk Management Framework.

Treasury official · press release, Feb. 18

Bank and insurer executives

Brian Moynihan

Chief executive, Bank of America

AI coding tools have taken 30% out of the coding stage of software work, saving the equivalent of about 2,000 people, and headcount will come down in 2026.

Headcount fell by about 1,070 through attrition in the first quarter.

Bank CEO · earnings call, Jan. 15

Jane Fraser

Chief executive, Citi

“new sources of efficiency that, three, four years ago, we couldn't have imagined”

At Davos in January she said AI will take some jobs away while creating others.

Bank CEO · earnings call, Jan. 14

Jamie Dimon

Chief executive, JPMorgan Chase

JPMorgan will stay out front and keep spending on technology and AI to compete with fintechs and Schwab, and will not cap spending to hit an expense target.

JPMorgan raised its 2026 technology budget to about $19.8 billion at its February company update.

Bank CEO · reported remarks, Jan. 13

Alan Schnitzer

Chief executive, Travelers

AI is embedded in Travelers' business operations today; he gave no headcount projection beyond claims.

Travelers is the clearest insurer example of agents in claims intake and underwriting.

Insurer CEO · earnings call, Jan. 21

Executives who treat AI as one lever among several

Mike Santomassimo

Chief financial officer, Wells Fargo

Wells Fargo's 2026 efficiency work has no new silver bullet; it rests on automation, lower real estate costs and reduced third-party spending.

CEO Charlie Scharf said on the same call that the bank has more efficiency tools than ever, especially AI. In December 2025 he said it had not reduced the number of people coding despite 30% to 35% efficiency gains.

Bank CFO · earnings call, Jan. 14

Dermot McDonogh

Chief financial officer, BNY

BNY continues to spend more on cyber resiliency than on AI.

The comparison puts AI spending in proportion at a bank running more than 130 agents.

Bank CFO · earnings call, January

Buyers of AI builds

Marco Argenti

Chief information officer, Goldman Sachs

The agents built with embedded Anthropic engineers are in early development and should cut processing time significantly; he reportedly described them as digital co-workers for many of the firm's professions.

Goldman chose co-development for specific back-office steps over a finished product.

Buyer · reported remarks, February

Robin Vince

Chief executive, BNY

BNY buys from several labs on its own model-agnostic platform and treats AI as staff capacity that financial targets may not fully reflect; adoption and integration could limit the gains.

His multilab platform differs from Goldman's co-development with Anthropic.

Buyer · earnings call, January

Payment networks and platforms

Michael Miebach

Chief executive, Mastercard

Agentic commerce is early but Mastercard is ready; it is one more route to payment choice, and the timing of scale is hard to predict.

Mastercard targeted enabling its global issuer base for Agent Pay by the end of the first quarter.

Vendor · earnings call, Jan. 29

Ashish Gupta

Google, quoted in Klarna's announcement

“Open standards like UCP are essential to making AI-powered commerce practical at scale.”

Klarna joined the protocol on Feb. 2, and Visa is building interoperability with it.

Vendor · press release, Feb. 2

Investors and analysts

Mark Murphy

Head of U.S. enterprise software research, JPMorgan

Suggesting that a new plug-in could replace every layer of mission-critical enterprise software is an illogical leap.

In the same Reuters report, Ocean Park Asset Management's chief investment officer said these companies' moats look much narrower.

Sell-side analyst · reported comment, February

Kort Schnabel

Executive, Ares

Only a very small number of Ares Capital's portfolio companies could be disrupted by AI, and the risk is not in core enterprise software.

Alternative managers fell 3% to 11% in early February on fears about software lending before partly recovering.

Lender · reported comment, February

11Decisions

Three decisions for operations, risk and technology leaders

Chief operating officer

Which processes get agents, and where the capacity goes

The quarter's deployments targeted narrow steps with clear rules: trade accounting, onboarding checks, claims intake and payment validation. Firms then booked the gains as capacity, attrition and efficiency targets. Decide for each process, before an agent goes live, whether recovered time cuts cost, absorbs volume growth or moves staff to client work, and who signs off on that choice.

For each agent in production, can we name the unit of work, the time saved and where that capacity went?

Chief risk officer

Whether model risk management covers agents as systems

In Bank of England forums, firms treated prompts and retrieval layers as within SS1/23 and said traditional validation is unsustainable for agents. The ECB expects second-line review before material uses and tested contingency options for critical cloud services, and FINRA flags agents acting beyond their scope. Decide whether oversight moves to outcome testing, action logs and shutdown procedures, and who owns them.

Could we show a supervisor what one of our agents did last week, who approved it and how we would stop it without disrupting payments?

Chief technology officer

Co-develop with a lab, run several models or license a product

Goldman co-developed agents with Anthropic engineers, BNY runs several labs' models on its own platform, and Bank of America rolled out Microsoft 365 Copilot to about 200,000 employees. U.K. firms expect switching to get harder as agents embed, and consortium members flagged short-notice vendor changes. The choice sets how exposed the firm is to one vendor.

If our main model provider changed or withdrew a model at short notice, how long would it take to move the affected agents, and have we tested it?

12Actions

Four actions for the second quarter

  1. 1

    Build an agent register with control evidence

    List every agent that acts across systems, with its permissions, logs, human checkpoints and second-line review status. The result is a record that can be shown to a supervisor applying SS1/23 or the ECB's expectations.

    Head of model risk

  2. 2

    Test a model change and a shutdown

    Run one exercise in which the model behind a production agent changes without notice and one in which the agent is stopped mid-process, and record what breaks downstream, including payments. The result is a tested contingency and shutdown path of the kind the ECB and the Bank of England's consortium described.

    Chief technology officer

  3. 3

    Use the 230 control objectives in vendor reviews

    Map core-banking and fintech vendors' AI documentation to the Treasury framework's control objectives and adoption-stage questionnaire; insurers can extend the review to the third-party data and models an NAIC working group is building a framework to oversee. The result is a vendor gap list before examiners or counterparties ask for one.

    Third-party risk lead

  4. 4

    Report AI gains by where they land

    Separate capacity absorbed, roles not backfilled, cost removed and any price changes in management reporting, so efficiency claims to investors, staff and supervisors rest on the same figures. The result is one account of AI savings that can withstand questions like those raised about Morgan Stanley's cuts and Goldman's agents.

    Chief financial officer

13References

References

Coverage runs from Jan. 1 through March 31, 2026. Sources published in the following month are cited only for what they report about that period.

Models, products and deployments

Company disclosures

Capital and markets

Regulation and supervision

Research, surveys and risk

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?