Agentic AI in banking: what bank 10-Ks and 10-Qs tell the SEC about AI agents, and what the Fed has said

Banks are telling the SEC about agentic AI mostly as a risk and a change in their industry, and only rarely as something they run: of the 40 bank 10-K and 10-Q filings that name agentic AI or AI agents and could be read and coded, 7 describe the bank's own use or evaluation, 17 a risk and 16 an industry trend.¹ The words arrived late in bank filings, with 4 bank 10-Ks and 10-Qs in all of 2025 and 27 in the first quarter of 2026, against 294 filings of every kind that used “agentic” in that one quarter.¹

The Federal Reserve’s most direct statement on the subject came from Governor Christopher J. Waller on 2026-09-29, who described two models of “agentic commerce” and called trust between buyers and sellers “the biggest barrier” to scaling it.² The supervisory guidance on models issued in April 2026 leaves agentic AI out of its scope altogether.¹¹

Where the term first appears in bank filings

EDGAR full-text search finds no bank 10-K or 10-Q using “agentic” or “AI agents” in any quarter of 2024.¹ The first is a 10-K filed by BankUnited, Inc. on 2025-02-28, with one sentence: generative and agentic AI “are emerging technologies that present both opportunities and risks to users.”¹ Axos Financial’s 10-K of 2025-08-21 followed, then WaFd’s and Pathward’s 10-Ks in November 2025, both for fiscal years ending in September.¹

Across all filers the word moved earlier and faster. Full-text search counts 3 10-K and 10-Q filings with “agentic” in 2024-Q4, 28 in 2025-Q1 and 72 in 2025-Q4.¹ Banks were a small part of that: in 2026-Q1, 23 of the 294 filings that used “agentic” came from filers with a bank SIC code, about 7.8%.¹

Filings per quarter that name agentic AI or AI agents

The method is two searches, one for “agentic” and one for “AI agents”, each run on forms 10-K and 10-Q one calendar quarter at a time and paged to the end, with each filing counted once by its accession number. A hit is kept when the filer’s first SIC code on EDGAR is 6021 (national commercial banks), 6022 (state commercial banks), 6035 or 6036 (savings institutions). That leaves out bank holding companies EDGAR files under other codes and every credit union, broker and card network.

The shape is the 10-K calendar. Calendar-year banks file 10-Ks in February and March, which is why 2026-Q1 holds 27 of the 44 bank filings counted from 2025-Q1 to 2026-11-11.¹ The quarters after it fall back to 2, 4 and 7, and most of those are 10-Qs that repeat a risk factor or an item from the year’s 10-K. The 7 so far in 2026-Q4 run only to 2026-11-11, the date of the count.

Context: own use, risk, or industry change

Each of the 44 filings was read on sec.gov and every sentence naming agentic AI or AI agents was kept. Four had no sentence the reader could isolate in the main document, leaving 40 filings from 33 banks.¹ Each filing was then coded once by the first of three rules that any of its sentences matches:

• Own use. The bank, as subject, deploys, integrates, invests in, adopts, implements, partners for or evaluates the technology, with no “may”, “could”, “might” or “monitor” between the subject and the verb.
• Risk. The sentence names risks, fraud, threats, liability or adverse effects, or uses “may”, “could” or “might”.
• Industry change. Anything else: the technology described as a development in banking with no stance.

The risk sentences repeat across banks. One phrasing, that AI models “particularly generative or agentic AI models, may produce outputs or take action that is incorrect”, appears in the 10-Ks of WaFd, Capital City Bank Group, Patriot National Bancorp and Western Alliance.¹ The industry sentences repeat too: “Recently, the financial services industry has experienced rapid developments in artificial intelligence, including agentic artificial intelligence” appears word for word at Simmons First, Burke & Herbert, Atlantic Union and QNB.¹ Shared risk-factor wording is common, and it means the count measures how often a sentence travels between filings, not how many banks have an agentic program.

The larger banks write about agents in narrower terms. Capital One’s 10-K ties them to card losses:

Bank of America’s 10-K lists what could go wrong inside the bank: expanded use of AI, “including emerging third‑party AI services and autonomous AI agents, may result in increased data risk, unpredictable system interactions, inadequate controls or safeguards, AI failure, or produce unintended operations or consequences.”⁴ Ally Financial’s 10-K says the risks “may vary depending on whether our AI use is internal or customer-facing,” on what systems and data the tool touches “and what level of autonomy an AI solution is granted.”¹² Eastern Bankshares names agentic AI as a tool for the other side, allowing “bad actors to more easily use sophisticated and iterative methods of facilitating fraud through the creation of synthetic identities and ‘deepfake’ images and documentation.”¹⁰

Citigroup’s 10-Q for the third quarter of 2026 is the one bank filing in the set that defines the term in its glossary: “Artificial intelligence that autonomously pursues goals by planning actions, using tools, observing results and iterating with minimal human supervision.”⁹

What the Fed said about agents and payments

Governor Waller gave “Payments in the Age of AI Agents” at Sibos 2026 in Miami; the Board’s page dates it 2026-09-29.² The speech is about payments, not bank operations, and it is the clearest description a Fed official has given of what an agent does with money. Agents, Waller said, “can plan and execute multistep processes using LLMs, enabling those agents to transact autonomously.”²

He drew two models. In the agent-assisted model “the buyer remains in control. They make the decisions, and they handle the payment.” In the agent-delegated model “a buyer grants authority to an AI agent to shop and make payments on their behalf.”² He named three trust problems, “authentication, liability, and fraud,” and put the liability question plainly: “Who is on the hook if an agent makes the wrong purchase?”² Capital One’s risk factor asks the same question from the issuer’s side.³

For a finance department the most direct passage is about business purchases. Waller said B2B buying “may be especially well suited for agentic commerce” because purchases “are often recurring and abide by a set of rules, such as approved suppliers and budget limits.”² He added that higher B2B transaction values “amplify the financial exposure from agent errors or unauthorized actions, so robust controls and monitoring systems will be essential,” and that agents would have to pay across “ACH, wire, instant payments, and cards.”² He also said fraud systems “calibrated to human behavior, may not translate well to agents.”²

The model risk guidance the Fed, the FDIC and the OCC issued on 2026-04-17 as SR 26-2 does not cover any of this. Its footnote 3 reads: “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance.”¹¹ It points banks instead to their general risk management and governance practices for “any tools, processes, or systems not covered in this document.”¹¹

Finance-function uses banks name

Seven filings from five banks describe their own use or evaluation, and they are the most useful for a finance leader because they say where the tool sits.¹

Business First Bancshares reported in its 10-Q of 2026-05-01, and repeated in the next two 10-Qs, that “On February 17, 2026, we announced a strategic partnership with Covecta, to deploy agentic AI across the bank's day to day workflows.”⁶ Valley National Bancorp’s 10-K says the bank has made “and expect to continue to make investments to integrate artificial intelligence tools into our solutions,” naming agentic AI among them.⁷ C&F Financial’s 10-K, under efficiency and technology, says the company is “evaluating potential use cases across all subsidiaries and departments for the prudent adoption of Artificial Intelligence (AI) including robotic process automation, generative and agentic AI, and machine learning tools.”⁸

Two banks, Sierra Bancorp and Bank of Marin Bancorp, filed 10-Ks with closely matching language that says where each stood when it filed.

Bank of Marin’s 10-K says the company “continues to evaluate additional AI capabilities, including agentic AI, in limited capacities,” and lists the risks it weighs: errors in work product, privacy, intellectual property, bias and cybersecurity.¹³

None of the 40 filings names a finance-department task (the close, payables, reconciliations, forecasting, regulatory reporting) as a place where the bank runs an agent. The process names that do appear sit in risk factors about AI generally.

What banks say stays with a person

The reader also searched every kept sentence for a human-review statement: human review, oversight, judgment or intervention, “human in the loop”, review by personnel or management, manual review. None of the 40 filings has one in a sentence that names agentic AI or AI agents.¹ That does not mean the banks have no such controls; it means they did not describe them next to the word.

The nearest statement is Citigroup’s definition, which builds the absence of a person into the term itself: “iterating with minimal human supervision.”⁹ Governor Waller’s speech puts the control question on the buyer in the delegated model: the buyer “may specify some constraints and set up guardrails, but the agent operates autonomously.”²

What is not yet in the filings

Four things a reader might expect are absent from the 44 bank filings counted here.¹

In the sentences that name the term, no bank names an agent vendor other than Business First’s partner, and none gives a cost, a headcount effect or a count of agents in production. No bank filing in the set mentions SR 26-2, though the guidance excludes agentic models by name.¹¹ No bank describes agents initiating payments from its own accounts, the business-to-business case the Fed governor singled out.² And no filing in the set reports an incident caused by an AI agent.

The count will move each February. The site’s refresh reruns the two searches daily for recent quarters, and this article’s figures are recomputed from the saved results by the script in the site’s repository (scripts/cfonews-ai/articles/compute.mjs). The bank filings that describe AI in the finance function more broadly are collected in AI in their own finance work, and the Fed speeches on AI are listed in the Federal Reserve tracker.