AI agents in banking: which banks run agents, what they do, what regulators require.

What is an AI agent in banking?

An AI agent is a language model given tools, a goal and a stopping rule: it plans the steps, calls the bank's systems, checks its work and either completes the task or hands it to a person. A chatbot answers and a copilot suggests; an agent acts, inside limits the bank sets. 66 of the 140 banks profiled here record an agent on the record, 42 in production; the wider generative-AI count is 120.

The AI in banking pillar describes three generations: predictive models that score, generative assistants that draft and agents that act. The difference between the last two is the loop. An assistant takes a prompt and returns text; an agent takes a goal, decides which of its tools to call (a ledger, a document store, a CRM, a payment rail, a code repository), reads the result, decides again and keeps going until the task is done or a rule says stop. The stopping rule is where banks put their effort: UBS's advisor agents alert and prepare but trade only after the advisor decides; NatWest's complaints agent investigates across data sources and presents a summary to a human handler; Citi's coding agents read only internal repositories and a person approves every change.

The words banks use are not consistent. Bank of America's CTIO describes EricaAssist as a “human-assisted AI agent” and says there is “so much juice to be squeezed right now with assistive agents that are actually working with humans in the loop”; CIBC's CEO says of its agentic workspace “it's not a chatbot, it's a coworker”; BNY counts “digital employees”. This page counts a bank when its record names an agent, an agentic workflow, an agent platform or an agent-initiated payment; chat assistants that answer customers are on the conversational AI page unless they also act. On that definition 66 of the 120 banks with a generative-AI record have an agent record.

Which banks run AI agents?

42 banks run agents in production on the record, led by JPMorgan, Bank of America, Citi, BNY, Morgan Stanley, U.S. Bank, Huntington, TD, Amex, NatWest, UBS, DBS and Santander; 9 are rolling out, 7 are in pilot and 8 have announced plans. The production agents are narrow: coding, reconciliation, servicing exceptions, collections calls, onboarding files, advisor alerts and agent-initiated payments.

One row per bank at its strongest agent record, strongest status first and largest bank first within a status. The date is the earliest sourced event on the bank's timeline for that activity, so it is usually a first disclosure rather than a go-live. 42 rows are in production. Agent-initiated payments count as production where the bank says cardholders are enabled or merchants can accept them, and as pilot where the bank describes controlled transactions.

Three patterns. First, the largest banks built an agent platform before they built agents: JPMorgan's LLM Suite, BNY's Eliza, Citi's Arc, Northern Trust's Byron, Lloyds' Envoy, BBVA's Frame, U.S. Bank's internally developed agentic layer. Second, the regional banks with the most agents let employees build them: Customers Bank (more than 600 deployed, 75% of staff on OpenAI tools), Texas Capital (280 by about 64 employees, duplicates to be consolidated), Live Oak (more than 640 agents and skills). Third, the core providers are now the route for community banks: Fiserv's agentOS launched with 13 agents and six co-developing institutions including SouthState, Bank OZK and First Interstate.

What do the agents actually do?

Eight jobs recur on the record: servicing and complaints (Huntington, NatWest), collections (TD, Valley), onboarding and lending files (NatWest, TD, ING, Customers Bank), finance and reconciliation (Morgan Stanley's FIXR, Huntington), software engineering (Citi, PNC, Goldman, Barclays), research and advisor support (UBS, Deutsche Bank, Stifel), payments and agentic commerce (Amex, U.S. Bank, HSBC, Santander) and employee-built task agents (BNY, Customers Bank, Texas Capital, DBS). In every case the agent prepares or executes a bounded step and a person owns the decision.

The table groups the production and pilot records by job. Each row names the kind of agent, what it does and the banks on the record that run it; the bank page carries the source. The common shape is a step that used to be a person reading several systems and typing the result into another one.

What is missing from the table is as telling as what is in it. No bank on the record lets an agent make a credit decision, move a customer's money without that customer's instruction or answer a regulated complaint without a human check. JPMorgan aims to have a consumer-facing travel agent in pilot before the end of 2026 while expecting customers not to delegate purchasing to agents soon; Wells Fargo plans long-running agents for advisers to change beneficiaries or move funds with guardrails and a human in the loop; SoFi plans agentic actions inside Coach. These are announcements, and the table marks them as such.

What results do banks report from agents?

TD's collections agent reaches 20% to 25% of past-due customers against about 6% to 7% for human callers; its agentic mortgage pre-adjudication takes under three minutes against 15 hours. BNY's contract-review agent cut legal review from four hours to one across more than 3,000 agreements a year. PNC says about 115 agents built its new mobile app and agent-centred development is 5 to 10 times faster. Barclays' agent squad migrated a legacy database in 20 days against three months. Standard Chartered's onboarding agents cut hiring-manager effort 35%. Huntington runs 50 agents in production, Customers Bank more than 600, Santander more than 280 process-automation agents.

The banks' own figures, with date and source. They measure different things (a connect rate, hours per contract, agents built, developer throughput) and are not comparable with each other. Counts of agents built by employees are an adoption measure, not an output measure: Texas Capital says its 280 agents include duplicates to be consolidated. Vendor-reported figures are labelled.

The strongest numbers are the ones with a before-state and a denominator: TD's connect rate (6% to 7% with people, 20% to 25% with the agent), BNY's contract review (four hours to one, more than 3,000 agreements a year), Barclays' migration (20 days against three months, according to Microsoft). Lloyds is the one bank to put a value on the programme in advance: more than £100 million of generative and agentic AI value expected in 2026. Santander reports €35 million of AI business value in Q1 2026 with more than 280 process-automation agents in production.

How do banks control their agents?

Human approval before the action, scoped access and action limits set at build time, no outside agents on customer credentials, audit trails and kill switches, autonomy earned against a trust score, identities and access reviews for agents as for employees, and deterministic paths for regulated journeys. Each has a bank describing it on the record; the agentic banking control plane page sets out the full design.

The CEOs' language is cautious: Jamie Dimon says “you don't want rogue agents, and there could be thousands of them”; Brian Moynihan says the risk “largely is around agents just left to operate. We just don't do that”; PNC's president says making sure agents do what they are supposed to do “is a top priority”; Northern Trust's CFO asks whether they will “do the right thing at all times”. The table lists the controls banks describe on the record. The design of a full control plane (identity, permissions, policy checks, logging, kill switches, evaluation) is on the agentic banking control plane page and is not repeated here.

What do regulators say about AI agents?

Nothing agent-specific is in force for banks yet. The April 2026 US model-risk guidance leaves generative and agentic AI to broader governance; NIST's CAISI asked for input on agent security in January 2026; EU, Colorado and California rules require disclosure when a customer deals with AI; the Bank of England and FCA's 2026 survey asks about agentic AI explicitly; MAS's October 2026 guidelines and the FSB's consultation treat agents inside AI risk management; the FCA's live-testing track let NatWest test an agentic complaints investigator with the regulator watching.

The tracked documents that mention agents or bind them through disclosure and governance duties. The notes say what each means for an agent programme; the full regulatory treatment, by jurisdiction and with the supervisory speeches, is on the agentic banking regulation page. For the model-risk layer see model risk management; for the governance structure around agents see AI governance.

Two things to notice. The US guidance that matters most, SR 26-2, does not regulate agents; it says they are outside the model-risk framework, which pushed every large bank to describe an enterprise AI governance structure instead. And the first dedicated agent document is from NIST's standards centre, not a bank supervisor: the CAISI request for information on agent security. The supervisors are watching through surveys (the Bank of England and FCA) and risk perspectives (the OCC) rather than rules.

Which vendors supply agent platforms?

38 vendors in this site's directory describe agents in their products. The largest banks build on frontier-model and cloud platforms (OpenAI, Anthropic, Google, Microsoft, AWS); regional banks get agents from their core provider (Fiserv agentOS, nCino, FIS, Q2); specialists such as Covecta, Glia and Nasdaq Verafin sell agents for one job.

The table lists vendors in this site's AI tools directory whose products include agents, with the number of banks in the US set that name them on the record. A bank naming a vendor is a disclosure, not a market share. The largest banks rent models and build the platform; the record shows Fulton on Salesforce Agentforce (first agent set up in an hour, 3,000 employees six weeks later), USAA on IBM watsonx Orchestrate, Wells Fargo on Google Agentspace and SouthState treating Microsoft Copilot Studio as a core capability for citizen development.

The hiring record shows where the work is. Open roles at the banks here include agentic AI platform engineers at Bank of America and Capital One, an enterprise agent orchestration lead at Citi, agentic identity security at Wells Fargo and Citizens, agentic commerce product and fraud roles at Amex and a head of mobile and agentic banking transformation at TD. The AI hiring page tracks the full set.