AI pressure on financial firms shifted to customer accounts and cyber defenses

Banks, insurers and asset managers now run AI agents for code, operations and adviser support under tightening permissions. In the third quarter, new pressure came from outside those programs: consumer agents that read customers’ bank data and buy on their behalf, and frontier models that find software flaws faster than firms fix them.

Editorial snapshot: 31 October 2026

The central finding

Banks responded with access controls and supervisors with existing cyber and third-party powers rather than new AI rules, while the return on firms’ internal AI programs stayed mostly self-reported.

01Key findings

What changed in financial services AI in the third quarter

AI kept spreading inside financial firms under tighter controls. The quarter’s new pressures came from outside: frontier models that speed up cyberattacks, and consumer agents that investors feared could erode the customer inertia behind deposits and insurance renewals. No jurisdiction issued a binding AI rule for financial firms.

Capability

Frontier models find flaws faster than firms fix them

A Bank of England report cites GPT-5.5 finishing one reverse-engineering task in about 10 minutes, against about 12 hours for a human expert. Firms told the Financial Conduct Authority that discovery is outpacing their remediation, and the Federal Reserve warned that attackers could gain such capabilities before the sector patches.

Capital

Fintech rounds fell to the fewest since 2018

Fintech companies raised $10.6 billion across 687 rounds as AI took 64% of global venture funding. Strategic buyers paid for distribution: Stripe and Advent offered about $53 billion for PayPal, and Stripe agreed to buy model router OpenRouter.

Adoption

Agents are common; autonomy and measured savings are not

Of 196 U.K. financial firms, 80% have AI agents deployed, but 2% give them broad autonomy. In a D.A. Davidson survey of 73 U.S. banks published in July, 11% said AI cost savings were measurable.

Accountability

The largest U.S. bank is keeping outside agents out of accounts

Bank and insurer shares fell on Sept. 22 on fears that consumer agents such as Meta’s Muse would erode customer inertia. JPMorgan said on its third-quarter call that it blocks outside agents from acting in retail accounts, and the U.S. open-banking rule that would shape such access is enjoined.

Operating question

Which agents — your own, your suppliers’ and your customers’ — can act on your systems, data and accounts, under what permissions and audit trail, and what does each one demonstrably return?

02Capability

What frontier models changed for security, systems and cost

For financial firms, the quarter’s capability changes sat at two edges. At the attack surface, supervisors on both sides of the Atlantic treated frontier models’ cyber skills as a near-term operational risk. At the customer interface, general-purpose agents began acting for consumers. Inside firms, choosing a model became a cost decision.

Cyber

Frontier models can now run multistage attacks

The Bank of England’s July Financial Stability Report cites U.K. AI Security Institute findings that the latest models are a clear step up and can carry out multistage attacks on vulnerable systems. The Financial Policy Committee says firms and authorities “should not assume those limitations will persist” and expects patching to be needed more often in every scenario.

About 10 minutesGPT-5.5, one reverse-engineering task

About 12 hoursHuman expert, same task

Remediation

Vulnerability discovery is outpacing repair

The Fed’s September cybersecurity report, citing researchers, says newer models can find unknown vulnerabilities, build exploits and link them into chains that run at machine speed. In the FCA’s Sept. 2 review, firms described moving from severity ratings to exploitability-based priorities, because models chain lower-rated flaws into attack paths.

Computer use

Agents can operate software that has no API

OpenAI’s GPT-6 Astra, announced Sept. 3, operates applications without APIs, with administrator controls over which sites and apps it may use. For banks with legacy cores and terminal-based back-office systems, that reaches processes no vendor has wrapped in an API.

Consumer agents

Personal agents began acting for consumers

Meta launched Muse in the U.S. in early September. It reads balances and transactions from accounts users link through Plaid, pays through Stripe’s Link with a one-time card after the user approves, and topped app-download charts in September. On Sept. 28, Instinct, a personal agent that connects to email, messaging, screen and location and can be authorized to make purchases and accept agreements, raised $1 billion at a $10 billion valuation with more than 100,000 invitation-only users.

Unit cost

Model routing became a cost decision

JPMorgan is pursuing model agnosticism to limit lock-in, and Jamie Dimon pointed on the third-quarter call to routing tools that send queries to cheaper, faster models. Token prices fell as volume moved to standard-tier models, but per-task cost still varies: at maximum effort, Claude Opus 5.5 used about 119,000 output tokens per Artificial Analysis Intelligence Index task, against about 27,000 for Astra.

$0.68Effective price per million tokens, early September, Ramp data

Down 41%From the March peak

Cost and consequence

Two budgets moved: patching, which supervisors expect to run more often in every scenario, and agent run costs, where a low token price does not guarantee a cheap completed task. Both depend on architecture choices about which systems agents can reach and which model handles which task.

03Capital

Fintech deal count hit a 2018 low as strategic buyers bought distribution

AI took 64% of the $159 billion in global venture funding Crunchbase counted for the third quarter, and a record 27 companies raised $1 billion or more. Little of that reached finance-specific startups. The money that did move toward financial services bought distribution, data and routing positions, and banks also carried AI exposure through their lending.

$10.6B

Global fintech funding, Q3 2026

Down 28.4% from the second quarter’s $14.9 billion and 21.4% from a year earlier; U.S. companies took 50.5%.

687

Fintech rounds, the fewest in Crunchbase data since Q1 2018

Down 12.9% from the second quarter and 31% from a year earlier.

Disclosed financing

The quarter’s largest fintech rounds, with one consumer-agent round for scale

Disclosed amounts in USD millions, July 1 to Sept. 30, 2026. Crunchbase does not classify Instinct as fintech; it appears for comparison.

Ant International

Series A — $1.2B

$11.2 billion post-money valuation; early-stage funding rose 11% from the second quarter with it and fell 20.4% without it.

Instinct

Series C — $1B

Personal AI agent valued at $10 billion, up from $2.5 billion at its Series B.

Barq

Series A — $329.5M

Qupital

Series C — $300M

Ajaib

Series C — $270M

Tabby

Series F — $233M

Crunchbase fintech data as of Oct. 5; Instinct’s round as reported by PYMNTS.

Concentration

Fewer large rounds drove most of the decline

Rounds of $100 million or more fell 43.3% from the second quarter and accounted for about 93% of the quarter’s dollar decline. Smaller rounds fell only 5.1% from the second quarter but 31.4% from a year earlier, so the thinning reaches beyond the largest deals.

$5.15BRaised in rounds of $100M or more

Payments

Stripe and Advent bid about $53 billion for PayPal

The July 15 offer of $60.50 a share, a 28% premium, came with about $50 billion of financing from JPMorgan and Morgan Stanley. Its stated rationale was PayPal’s consumer network, Venmo and checkout credential rather than AI. Reuters reported the board’s preliminary view that the offer undervalued PayPal.

Model spend

Stripe agreed to buy model router OpenRouter

The Aug. 19 deal, valued at about $7.5 billion by The New York Times (neither company disclosed terms), gives Stripe the routing layer for spending on AI models. Patrick Collison described tokens as the central currency for companies building with AI.

Wealth

Vanguard agreed to buy Altruist

Vanguard’s Aug. 26 agreement to buy Altruist, a custody and technology platform for registered investment advisers described as AI-forward, rests on demand for advice exceeding adviser capacity. Terms were not disclosed, and the deal is expected to close later in 2026.

Lending

Banks are financing the AI build-out

The Bank of England called the pace of AI investment historically unprecedented, said AI companies’ use of credit markets had accelerated rapidly and flagged the mismatch between long-dated debt and short-lived chips. Private credit’s share of AI investment financing rose from 9% in 2024 to 34% in 2025, according to the OECD.

More than 15%Hyperscalers’ share of 2026 U.S. investment-grade issuance to early May, against 3% of debt outstanding at the end of 2025

41%AI issuers’ share of nonrefinancing U.S. high-yield issuance in 2026 to date

What it signals

The positions strategic buyers paid for sit where an agent meets a customer, a payment or a model, each a bid to keep a per-transaction fee as agents choose merchants and models. Banks also carry AI exposure through their loan books and capital-markets businesses, as well as their technology budgets.

04Products

Market data moved into agent platforms, and card networks built agent checks

The quarter’s product changes were mostly infrastructure. Data vendors put licensed data inside other companies’ agent platforms, card networks built ways to recognize purchases an agent initiated, and the largest banks built agent tooling in-house while buying models and data.

Market data

Licensed data now reaches agents through connectors

LSEG said on July 16 that its licensed data will be available in Model ML’s agent platform through an MCP connector, a standard way for agents to query outside data and tools; it already counts Claude and ChatGPT among its AI partners. FactSet’s integration went live in Gemini Enterprise for Financial Services on Aug. 25. Neither vendor published connector pricing.

Payments

Card networks placed themselves between agents and issuers

Mastercard’s first agent-trust service, introduced for testing in the U.S., gives issuers a probability score that an agent initiated a purchase, with Cloudflare and Skyfire as partners. Visa’s research found that 23% of U.S. adults trust generative AI to handle payments, while 61% would trust Visa with agentic transactions.

Large banks

Large banks built agent capability in-house

About 20,000 Bank of America developers use coding agents, and about 95% of staff have AI tools. Citi has 14,000 engineers on Devin and is re-engineering 100 manual processes with AI; Jane Fraser attributed a 30% year-over-year fall in technology vendor spend to in-sourcing.

Agent-related releases from data vendors and networks

2026 announcements; status as described at announcement.

Agent-related releases from data vendors and networks
Product AnnouncedStatusWhat it covers
LSEG data in Model ML Announced July 16 Status Announced; no availability date What it covers Pricing, reference, fundamentals, ownership, ESG and news Also estimates and macro data
FactSet in Gemini Enterprise for Financial Services Announced Aug. 25 Status Live; expanding What it covers Portfolio analytics, fixed income, derivatives, scenarios
Mastercard Agent Pay trust and intelligence services Announced Sept. 30 Status Probability score in U.S. testing What it covers Agent-initiation score; identity, intent and fraud signals
Bloomberg ASKB Announced Feb. 23, before the quarter Status In beta What it covers Agents over Bloomberg data, news and 800+ research providers

Visa published consumer research on Sept. 9 but no operating figures for Visa Intelligent Commerce.

Delivery

For research and trading teams, data entitlements and connector terms become part of choosing a model platform. For issuers, whether an agent initiated a purchase becomes a new input to fraud, authorization and dispute decisions.

05Adoption

Agent use is broad; autonomy, integration and measured savings are not

Every survey here found high AI use, but each counts a different population with a different question, so none can be pooled. The differences that matter are how much agents may do on their own, whether new tools connect to core systems and whether anyone can measure the savings. Consumer agents added a revenue-side question.

What each survey or count measured

Each row is a separate survey or count with its own population, question and date. Figures are not comparable across rows.

What each survey or count measured
Survey and population AI in useAgents or autonomyValue measured
Bank of England and FCA: 196 U.K. financial firmsPublished Oct. 26 without fieldwork dates; half the sample is investment and capital-markets firms. AI in use 86% Use AI, up from 75% in 2024 Agents or autonomy 80% Have agents deployed Value measured 2.95 of 5 Current efficiency benefit; 4.34 expected in three years
Bank of England and FCA: 15 U.K. insurers AI in use 93% Use AI Agents or autonomy — No comparable figure Value measured — No comparable figure
KPMG: 100 U.S. banking leaders at $1B+ revenue firmsFieldwork Aug. 10 to Sept. 14. AI in use — No comparable figure Agents or autonomy 66% Workforce adoption of agents, up from 56% Value measured 64% Report measurable value, including strategic value
Bank Director: U.S. community bank leadersReleased Sept. 15; sponsored by Jack Henry; sample size not published in the release. AI in use 72% Generative AI implemented Agents or autonomy 30% Use agentic AI Value measured — No comparable figure
D.A. Davidson: 73 U.S. banksReport dated July 8; 35% have quantified use cases. AI in use 42% Individual user tools Agents or autonomy — No comparable figure Value measured 11% Say cost savings are measurable
FactSet: seven U.S. insurers’ Q2 callsTravelers, Chubb, MetLife, AFG, Selective, AIG and Hamilton. AI in use — No comparable figure Agents or autonomy — No comparable figure Value measured None Of seven insurers discussing AI, none quantified its effect
Mercer: 131 asset managers, before the quarterFieldwork February and March; published May 22. AI in use 55% AI in at least one strategy’s investment process Agents or autonomy 5% Autonomous or semi-autonomous recommendations or trades Value measured — No comparable figure

— No comparable figure.

Autonomy

Agents assist; few decide

In the Bank of England and FCA survey, 80% of firms use agents for decision support, 46% for constrained execution and 22% with conditional autonomy. Few firms expect agents to take decisions on customers’ money.

2%U.K. firms giving agents broad autonomy

9% or lessExpect agents to approve outcomes, trade, pay or change limits

Community banks

Integration with core systems is the limit

Bank Director’s respondents named connecting new technology to existing systems as a top concern, and more than half of banks with a failed technology project blamed integration. Fintechs outranked big banks as a competitive threat for the first time in the survey, behind local banks and credit unions (61%).

69%Name integration a top concern

58% vs. 48%See fintechs vs. big banks as a threat

Asset managers and trading

Investment AI stays mostly in research

Algorithmic execution, market making and hedging is a current AI use for 14% of U.K. respondents, and liquidity and treasury decisions for 10%, against 86% for enterprise productivity. In wealth management, the gain is framed as adviser capacity: Vanguard cites demand for advice exceeding it, and UBS reports adviser hours saved.

Consumers

Some consumers are ready to delegate

About 20% of U.K. adults, roughly 11 million people, say they are likely to use AI that acts autonomously within goals they set, according to an April survey for the FCA. In Visa’s May poll, 72% of U.S. adults had used an AI assistant.

Deposits and renewals

Investors priced a threat to customer inertia

The S&P 500 Financials index fell on Sept. 22 while the broader market was roughly flat, with JPMorgan and Wells Fargo down more than 3%; Bloomberg tied the drop to Meta’s Muse. The exposure is in deposit spreads and renewal pricing, which depend on customers not shopping.

Nearly 2%S&P 500 Financials fall, Sept. 22

5.5%Allstate’s fall the same day

Spend

Budgets are rising ahead of measured returns

Bank of America spends about $4 billion a year on new technology initiatives and expects to roughly double its AI expense in 2027. D.A. Davidson’s 73 banks expect AI to cut expenses by 4.5% to 5% and lift revenue by 2.5% to 3% within two to three years.

Economics

The efficiency case still rests mostly on expectations, while the new risk sits on the revenue side, in deposit spreads and renewal pricing. Both turn on the same control: how much an agent, the firm’s or the customer’s, may do without a person.

06Outcomes

Reported results were self-measured, mostly in hours and speed

Large banks reported AI results at conferences, on earnings calls and through Evident’s index. Every figure here was reported by the firm itself, without audit, and most measure time or throughput rather than cost. None comes from an independent study of a financial workflow.

Bank of America · About 140 uses

A cost and benefit figure, given at a conference

Brian Moynihan gave the figures at a mid-September investor conference rather than on an earnings call. The bank has not disclosed how it defines the benefit.

$400MCost

$800MSelf-reported benefit

Bank of America · Efficiency

A better efficiency ratio, with no AI share attached

The efficiency ratio measures costs per dollar of revenue, so lower is better. The bank’s CFO framed the second-quarter gain as AI-enabled productivity, but the bank does not attribute a share to AI; Consumer Banking posted a 51% ratio.

59%Q2 2026 efficiency ratio

63%Q2 2025

Citi · Product delivery

Citi reports faster product launches

Jane Fraser said on the third-quarter call that AI makes product launches faster at lower cost. Citi did not say how the figure was measured or against what baseline.

30% to 40%Faster launches, bank estimate

Morgan Stanley and UBS · Time saved

Savings counted in hours

Morgan Stanley reinvested the developer time freed by DevGen.AI rather than cutting IT staff. UBS reports the time its STAAT Insights tool saves advisers each week. Both figures were reported through Evident.

More than 1 millionDeveloper hours freed, Morgan Stanley

About 1,200Adviser hours saved weekly, UBS

Lloyds · Generative AI

A value figure and a projection

The 2025 figure is value Lloyds reports from generative AI; the 2026 figure is the bank’s projection. Both were reported through Evident.

About £50MValue in 2025

More than £100MExpected in 2026

Insurers · Q2 calls

One jointly attributed insurer figure

Travelers confirmed that AI and innovation together improved its business-insurance underlying loss ratio, without separating AI’s share. Chubb’s Evan Greenberg said Chubb tracks AI gains in hard dollars and that token costs are a small fraction of them, but gave no figure.

0.5 pointTravelers loss-ratio improvement, AI and innovation combined

What the results show

These figures show capacity at the largest firms. They become a return when a firm shows where the freed hours went — fewer hires, more revenue or lower unit costs — and Bank of America is the only firm here to set a cost beside its benefit.

07Corrections

Four claims that need a narrower reading

Four widely repeated claims about AI in financial services overstate or misdate what happened. The distinctions change how a board should read returns, fraud losses, jobs and EU deadlines.

Returns · Partly true

A growing minority disclose figures, mostly outside earnings calls

Claim under review: Banks cannot show what AI saves them.

Evident counts 12 of 50 large banks reporting a realized or projected AI return, up from eight. Bank of America gave its figure at an investor conference; Milos Maricic found that none of the 21 banks he tracks had put one on an earnings call, and BNY’s Robin Vince gave none when pressed on the third-quarter call. Disclosed figures are unaudited.

Fraud · Projection, not a measure

The $40 billion figure is a 2024 projection for 2027

Claim under review: Generative AI fraud is a $40 billion problem.

Deloitte projected in June 2024 that generative AI could enable U.S. fraud losses of $40 billion by 2027, up from $12.3 billion in 2023. The FBI’s first count with an AI descriptor, for 2025, recorded 22,364 complaints and $893 million in losses, covering only cases victims identified as AI-related. One projects all AI-enabled fraud; the other is a floor of self-identified cases.

Jobs · Incomplete

Bank totals are falling through attrition while AI roles grow

Claim under review: AI is cutting bank jobs.

Bank of America’s headcount fell from about 213,000 to 209,000 during 2026 through attrition, with no layoffs, Brian Moynihan said, and executives said on the third-quarter call that automation means less hiring to refill roles. Evident counts software implementation roles up 4.3% across 50 banks in a year, and Goldman Sachs CFO Denis Coleman said on July 14 that the firm was not pursuing a structural rework of headcount. Total staff and AI roles are different counts.

EU deadlines · Out of date

The AI Act omnibus moved them to Dec. 2, 2027

Claim under review: EU AI Act duties for credit scoring and insurance pricing apply from August 2026.

The omnibus, in force since July 27, moved the high-risk obligations covering creditworthiness assessment and life and health insurance pricing from Aug. 2, 2026, to Dec. 2, 2027. Enforcement of general-purpose AI rules and Article 50 transparency duties did begin on Aug. 2, 2026.

08Accountability

Failures stayed small as firms narrowed what agents may do

The quarter’s failures in financial services were small and customer-facing, or measured in benchmarks. The larger risks under discussion, frontier-AI attacks and agents moving deposits, had no incident attached in the quarter. Firms responded by narrowing agent permissions, which limits today’s failures and also caps the autonomy the efficiency case assumes.

Incident

A bank chatbot insulted a customer

Around Sept. 18, Libra Internet Bank’s customer chatbot, Lia, called a client an idiot. The Romanian bank apologized, blamed negative patterns the bot had picked up from human input and said it would add stricter controls. No regulator action was reported.

Evaluation

General chatbots failed most money questions

Saturn asked 17 consumer models 121 money questions five times each, passing an answer only if it included every figure, rule, warning and deadline. Most errors were incomplete answers or missing figures; 6.1% were fabricated rules. Saturn sells AI to about 750 advice firms and did not name its graders. These are the assistants consumers are starting to delegate to.

43%Average accuracy

12%Hard multipart tax questions

61%Best model, Claude Opus 5 with reasoning

Security

Agent tooling was itself attacked

In August, a malicious MCP server campaign steered agents toward SSH keys and cloud credentials, attackers kept exploiting Langflow flaws, and researchers reported prompt injection achieving code execution in a coding agent’s auto mode. These findings bear directly on banks that give coding agents to 14,000 to 20,000 developers.

Response

Large banks added human oversight

The share of KPMG’s large U.S. bank respondents who keep agents away from sensitive data without human oversight rose sharply in one quarter. Firms in the FCA’s cyber review described limits on model permissions and human approval for higher-risk actions as guardrails.

75%Require human oversight for agent access to sensitive data

51%A quarter earlier

Response

JPMorgan keeps outside agents out of retail accounts

JPMorgan said after its Oct. 13 call that it currently blocks external agents from acting for retail customers; agents can still read balances through connectors such as Plaid but cannot send money or pay bills. The bank warned that customers who authorize outside agents may take on liability, and who bears the loss remains unsettled while the CFPB’s open-banking rule is enjoined.

Coverage

Insurers are writing AI out of commercial policies

ISO made generative-AI exclusions for commercial general liability available in January 2026, and carriers are filing them; W.R. Berkley’s absolute AI exclusion covers directors and officers, errors and omissions and fiduciary liability. Brokers are telling clients to check which policies still respond to AI losses. Specialist AI liability products are described as coming to market, but none has been named.

Who is accountable

What firms retain is decided before anything goes wrong: which data and actions each agent can reach, who approves higher-risk steps and which AI losses their own insurance still covers.

09Regulation

Supervisors used existing powers on frontier AI as the EU delayed lending and insurance duties

No jurisdiction issued a binding AI rule for financial firms in the quarter. Supervisors in the U.K., EU and U.S. converged on frontier-AI cyber risk and dependence on a few providers, and used tools they already had. The EU postponed the main AI obligation aimed at lenders and insurers by 16 months, and the FCA published an agenda for agentic finance.

U.K. · FCA · July 6

The Mills Review set an agenda for agentic finance

The review of AI in retail financial services, led by FCA executive director Sheldon Mills, made seven recommendations, including adapting the regulatory perimeter, monitoring the move to autonomous models, building foundations for agentic finance and a public-interest AI financial capability service. The FCA said it would publish AI good and poor practice later in 2026.

U.K. and EU · Resilience · July 31 and Sept. 2

Frontier-AI risk runs through resilience and third-party rules

The EBA, EIOPA and ESMA called for consistent supervision of frontier-AI ICT risk and said their oversight of critical ICT third-party providers under DORA, the EU’s operational resilience law, is addressing it. The FCA’s review “does not introduce new rules, guidance or regulatory expectations.” The Bank of England named reliance on a few frontier-AI providers as a new concentration channel.

54%U.K. firms’ AI relationships with the top three model API providers

79%With the top three cloud providers

U.S. · Banking agencies

No AI model-risk rule, and existing exam tools for AI attacks

April’s interagency model-risk guidance excludes generative and agentic AI and says “non-compliance will not result in supervisory criticism.” The Fed’s September report relies on existing tools such as joint exams of eight global systemically important banks. On Sept. 29, Michelle Bowman invited community banks to say what clearer expectations they need.

U.S. · Open banking

The rule that would shape agent access is on hold

The CFPB’s Section 1033 open-banking rule remains enjoined and under reconsideration; the bureau told the court it considers its own rule unlawful, and no new compliance dates have been set. Until it is settled, banks’ own access policies decide what customers’ agents can do.

U.S. · SEC · Sept. 10

The SEC will not prescribe the models firms use

Chair Paul Atkins flagged hallucinations in disclosures investors rely on and rejected using AI summarization to loosen the materiality principle. He tied the remarks to Regulation NMS reform, after the SEC proposed in June to rescind Rules 611 and 610(e).

U.S. · Insurance

State insurance regulators piloted an AI exam tool

The NAIC’s AI Risk Evaluation Supplement was piloted with 12 states from March through September for market-conduct and financial exams, and at least half of states have adopted the NAIC model bulletin. Colorado’s new AI law, effective Jan. 1, 2027, leaves insurance under the state’s 2021 insurance AI law.

Dates on the AI compliance calendar

Status at the end of the quarter. Later dates are scheduled, not events.

Dates on the AI compliance calendar
Item DateJurisdictionStatus
General-purpose AI enforcement and Article 50 transparency Date Aug. 2, 2026 Jurisdiction EU Status Applies
FCA AI good and poor practice publication Date Later in 2026 Jurisdiction U.K. Status Scheduled in July; nonbinding
Bank of England and PRA consultation on cyber and ICT risk Date No date set Jurisdiction U.K. Status Planned
Interagency RFI on model risk for generative and agentic AI Date No date set Jurisdiction U.S. Status Planned
Colorado SB 26-189, with insurance left under 2021 law Date Jan. 1, 2027 Jurisdiction Colorado Status Enacted
EU AI Act duties for credit scoring and life and health insurance pricing Date Dec. 2, 2027 Jurisdiction EU Status Moved from Aug. 2, 2026
CFPB Section 1033 open-banking rule Date Pending Jurisdiction U.S. Status Enjoined; under reconsideration

Compliance

For U.S. firms, the working limits on agents come from banks’ own access controls, an unsettled open-banking rule and insurers’ exclusions more than from AI law. EU lenders and insurers gained 16 months on high-risk duties but not on DORA’s third-party oversight.

10Voices

Where supervisors, bankers and analysts agree and differ

Supervisors, bank chiefs and card networks broadly agree that agents need identity, permissions and an audit trail before they act on money, and that frontier-AI cyber capability is the near-term risk. They differ on whether AI’s return is measurable now, whether consumer agents will pull deposits and how quickly autonomy should come. On each question the sides often describe different things: a cumulative program against a reporting quarter, a whole system against one bank’s franchise, a regulator’s horizon against firms’ current practice. Most voices here hold commercial positions, noted with each.

Regulators and supervisors

The FCA is preparing for agentic finance; U.S. banking and insurance supervisors are working with existing tools.

Sheldon Mills

Executive director, Financial Conduct Authority

“Artificial intelligence will transform financial services by 2030.”

His review recommends building foundations for agentic finance and monitoring the move to autonomous models. FCA chair Ashley Alder called the approach principles-based and outcomes-focused.

Regulator · FCA review, July 6

Michelle Bowman

Vice chair for supervision, Federal Reserve Board

“AI is becoming a critical component of these security measures”

Speaking to a community-bank cyber workshop, she invited feedback on clearer expectations and announced no AI guidance.

Regulator · speech, Sept. 29

Michael Conway

Colorado insurance commissioner

The NAIC should write an AI model law, because otherwise someone else will fill the gap.

His push for new law contrasts with federal banking supervisors, who are working through existing guidance.

Regulator · NAIC Summer National Meeting, as reported

Bank leaders

Large-bank chiefs agree on control and differ on whether returns show up yet.

Jamie Dimon

Chief executive, JPMorgan Chase

JPMorgan is neither for nor against agents, but it will not let just any outside agent inside the bank, given the risk of large numbers of rogue agents.

This sets the terms for consumer agents at the largest U.S. bank.

Bank leader · Q3 earnings call, as reported

Brian Moynihan

Chief executive, Bank of America

About 140 AI uses cost $400 million and generated $800 million in benefit, and the bank expects to double its AI budget in 2027, but a wrong answer to a client will limit where AI is used.

He disclosed AI cost and benefit at a conference rather than on an earnings call.

Bank leader · investor conferences, September, as reported

Jeremy Barnum

Chief financial officer, JPMorgan Chase

AI’s benefits are not yet observable in JPMorgan’s quarterly earnings, and he is reluctant to spend much effort quantifying them.

A reservation about measurement from a heavy adopter, given in answer to Wells Fargo analyst Mike Mayo. Asked whether agents could speed deposit runs, he said deposit stability comes from running the bank conservatively and from focusing on primary bank relationships rather than yield-seeking money.

Bank leader · Q3 earnings call, Oct. 13

Buyers and investors

Issuers and venture investors want transparency and evidence before they commit.

Todd Kennedy

Executive vice president, payments strategy, Capital One

Greater transparency into how agent transactions are initiated, authorized and executed is needed to meet consumer expectations.

Issuers make the authorization decision that agent-commerce signals are built to inform.

Buyer · Mastercard announcement

Laura Bock

Investor, QED

“The market remains bifurcated — AI vs. not and market leaders vs. not.”

Said of a quarter in which fintech round counts fell to their lowest since 2018.

Investor · Crunchbase News

Kirsten Green

Investor, Forerunner

Investors want stronger performance evidence before writing Series A and B checks, and agent payments are now a question of timing rather than possibility.

Investor · Crunchbase News

Independent analysts

The sharpest skepticism came from analysts, each with a commercial position.

Torsten Slok

Chief economist, Apollo Global Management

If households let agents move cash to higher yields, banks could lose much of the cheap deposit funding behind their lending.

Apollo manages assets that compete with bank deposits. Bloomberg’s report of the note cites no deposit-outflow data.

Economist · Apollo note, Sept. 27, as reported by Bloomberg

Milos Maricic

Founder, SPEC Research and Maximand

Across 919 earnings calls from 60 U.S. financial firms since 2023, AI cost came up on 53% of last season’s calls against 3% in late 2022, yet almost none quantify savings; he argues edges fade and costs rise.

Maximand sells AI token-cost optimization to financial institutions.

Analyst · Fortune op-ed, Oct. 9

Stewart Johnson

Author, FactSet Insight

“more positioning than proof”

His description of insurers’ AI commentary on second-quarter calls. FactSet sells data and AI connectors to financial firms.

Analyst · FactSet Insight, Aug. 14

11Decisions

Three decisions for financial services leaders

Head of retail and digital banking

Set the terms for customers’ agents

Consumer agents can now act for customers, JPMorgan keeps them out of accounts, and the federal open-banking rule is on hold. Each bank will set its own position on read and write access, authentication and liability.

Which actions will we let a customer’s agent take, how will we know an agent initiated them, and who bears the loss when one gets it wrong?

Chief risk officer and CISO

Fund patching at the speed frontier models find flaws

The Bank of England expects patching to become more frequent in every scenario. Firms told the FCA that discovery is outpacing remediation, and the FCA asks whether firms can separate plausible AI-found vulnerabilities from exploitable ones and whether key suppliers are ready.

Can we prioritize by exploitability and patch faster than attackers can chain flaws, and which critical suppliers cannot?

Chief financial officer

Define an AI return before the budget doubles

Bank of America plans to roughly double its AI expense in 2027, yet 11% of banks in one survey can measure cost savings and most disclosed results are hours saved. Disclosure is moving to conferences and annual materials, where it is unaudited.

Which programs will we report in money, against what baseline, and where will freed hours show up: fewer hires, more revenue or lower unit cost?

12Actions

Four actions for the next quarter

  1. 1

    Re-rank the vulnerability backlog by exploitability

    Score open flaws by how they chain into attack paths, as the FCA describes, not by severity alone, and ask critical suppliers for their patch timelines. Result: a patch-cadence target and a list of suppliers that cannot meet it.

    CISO

  2. 2

    Treat coding agents as privileged users

    Given August’s malicious MCP server campaign and prompt-injection findings, give each coding agent its own identity, an allow-list of tools and MCP servers, and logged actions. Result: agent activity that can be traced and revoked like a privileged account.

    Head of engineering

  3. 3

    Map every agent path into customer accounts

    List each aggregator connection, internal agent and partner integration that reads or moves money, with its authentication and logging. Result: the factual basis for setting agent-access terms, including whether to use network signals such as Mastercard’s agent score.

    Head of digital banking

  4. 4

    Track cost per completed task for three production uses

    For three AI uses in production, record the model, tokens, review time and where freed hours went. Result: a cost per completed task that reflects routing choices, and figures that can support a disclosure decision.

    Chief financial officer

13References

References

Covers July 1 to Sept. 30, 2026, as of Oct. 31, 2026. Reports and earnings calls published in October are cited for what they report about that period.

Supervisors and regulation

Capability and cost

Analysis and commentary

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?