What mid-size US banks and credit unions disclose about AI

About half of 50 mid-size US banks and 50 credit unions name an AI initiative. Two banks and two credit unions credit a cash figure to AI.

What did the study find?

About half of mid-size US banks and credit unions publicly name a specific AI initiative, and almost none put a cash figure on one. Of 50 banks, 26 named an initiative and two credited a cash figure directly to AI. Of 50 credit unions, 25 named one and two credited a cash figure directly to AI, both of them forecasts.

AI research agents (Anthropic's Claude) did the research: they searched the public record for each institution, recorded every AI initiative they found with a quote and a link, and graded what each quote claims. Every quote was then read again on its source page, by script or by a separate checking agent. Luke Wilcox, Roiva's founder, then reviewed the grade and use-case mapping of all 178 quotes from that pass. A second set of agents later listed the bank earnings calls, conferences and annual meetings in the window, read 310 of the 319 (the rest had no free transcript), and added 22 quotes; those were checked on their pages and graded under the same rules, not reviewed by a person.

How do banks and credit unions compare with other industries?

They name AI about as often as manufacturers and far less often than insurers. The table sets the two financial samples beside two earlier samples run the same way: US insurers (leaving out AI-native insurtechs) and US manufacturers, both SEC filers with $50 million to $2 billion of revenue.

Banks (n = 50) Credit unions (n = 50) Insurance, excl. AI-native insurtechs (n = 42) Manufacturing (n = 49)
Name at least one specific AI initiative 26 (52%) 25 (50%) 36 (86%) 27 (55%)
Give any figure for one (cash, hours or operational) 6 (12%) 12 (24%) 16 (38%) 11 (22%)
Counting only figures the institution or the press published 5 (10%) 4 (8%) 16 (38%) 10 (20%)
Give a cash figure of any kind 3 (6%) 3 (6%) 4 (10%) 7 (14%)
Give a cash figure they credit directly to AI 2 (4%) 2 (4%), both forecasts 1 (2%) 2 (4%), both forecasts
Nothing found at all 18 (36%) 25 (50%) 5 (12%) 12 (24%)

The insurance sample counts AI in products an insurer sells; the other three don't. A bank or credit union's own customer-facing chatbot or AI loan decisions count as its own operations.

Do larger institutions disclose more?

Yes, at banks steeply and at credit unions less so. None of the 6 banks under $2 billion of assets named an initiative, against 11 of the 13 at $10 billion and over.

Bank assets Banks Named an initiative Gave a figure Nothing found
Under $2B 6 0 0 5
$2B-$5B 17 7 1 7
$5B-$10B 14 8 2 5
$10B and over 13 11 3 1
Credit union assets Credit unions Named an initiative Gave a figure Nothing found
Under $1B 8 1 1 7
$1B-$2B 18 9 6 9
$2B-$5B 16 9 2 7
$5B and over 8 6 3 2

The two samples use different asset bands. At $5 billion of assets and over, 6 of 8 credit unions named an initiative, against 19 of 27 banks.

Why do credit unions give figures twice as often as banks?

Because vendors publish them. Credit unions give a figure for an AI initiative twice as often as banks (24% against 12%), but 8 of the 12 that do rest on vendor case studies alone. Counting only what the credit union itself or the press published, 4 of 50 give a figure (8%), against 5 of 50 banks (10%).

Credit unions file nothing with the SEC and hold no earnings calls, so their AI shows up mostly in vendors' customer stories and releases. Vendors wrote 36 of the 67 credit-union named-initiative records and 16 of the 23 figure records, and 8 credit unions are named only on vendor pages. For banks the SEC record matters more: named initiatives turned up in SEC filings for 18 banks and only outside them for 8. Most of that came from annual reports to shareholders, proxy letters and investor decks filed as 8-K exhibits, rather than the body of the 10-K.

Which institutions credit a cash figure to AI?

Two banks and two credit unions credit a cash figure directly to AI, and both credit-union figures are forecasts. Two more give a cash figure with a weaker link to AI. Each figure below is the institution's or its vendor's claim, as published; none was checked against the institution's accounts, and none is a benchmark.

Institution Figure Credit to AI Source
Metropolitan Bank Holding "CoCounsel reduce outside counsel spend by roughly $200,000 a year," of Thomson Reuters' AI legal assistant, which is in use Direct Chief AI Officer at the bank's Investor Day, March 2026
HomeTrust Bancshares Ocrolus says HomeTrust "estimates annual savings of 8,500 hours across loan processing teams ... $90,000" from Ocrolus document automation Direct; an estimate Ocrolus customer story, undated
PenFed Credit Union The CEO "said the program will save $1.6 million annually" Direct; a forecast CX Today, May 2026
Communication Federal Credit Union "$134.5M in incremental Vehicle loan originations and up to 20% reduction in losses" Direct; a forecast Scienaptic go-live release, July 2026
United Community Banks Fraud losses "dropped by 50% over the last 2 years partially because of" its fraud vendors' AI Shared with other causes CEO on the Q1 2026 earnings call
4Front Credit Union "We've reduced our cost with them by 78%," of its after-hours call vendor Not stated; the page doesn't say the voice assistant caused it Posh customer story, September 2026

HomeTrust's figure is Ocrolus's account of HomeTrust's estimate, in Ocrolus's own customer story. PenFed's $1.6 million is CX Today's paraphrase of what James Schenck, PenFed's CEO, said on Salesforce's earnings call, not a direct quote. Communication's figures are Scienaptic's projection at go-live. Metropolitan's is the only measured cash figure a bank in the sample states itself and credits directly to AI. No credit union in the sample reports a realized cash result from AI.

What do the other figures look like?

Most are hours or operational rates, and most come from vendors. Among the few an institution published itself, First Hawaiian's 2024 annual report says Zest AI let the bank "instant decision 95% of qualifying applications received in 2024", and First Internet Bancorp's first-quarter 2026 results release says "our virtual customer service agents resolve 45% of inquiries."

Credit-union figures cluster in two places: automated loan decisioning (approval and automation rates, mostly from Zest AI, Scienaptic and Corridor) and chat or voice containment (the share of conversations an assistant finishes without a person, in case studies from vendors such as Glia, Posh and Talkdesk). One hours figure is Salesforce's: its customer story on PenFed quotes the credit union as saying, "We save 30,000 minutes a day of labor with automated call summaries alone."

What do banks and credit unions use AI for?

Contact centers first. Chat and voice assistants, call summaries and agent assist make up 12 of the banks' initiatives and 21 of the credit unions' 61 specific ones.

Use Banks Credit unions
Contact centers and assistants 12 21
Credit underwriting and decisioning 4 10
Fraud detection 7 3
Enterprise AI programs 3 3

Banks recorded 80 specific initiatives and credit unions 61; the others are general statements about AI, plus one AI product a credit union sells. Banks also reported employee assistants such as Copilot and ChatGPT (8), KYC, AML and sanctions screening (6) and back-office automation agents (5). Three credit unions train staff with AI role-play.

Some institutions say they use AI only inside vendors' tools. BCB Bancorp's 10-K says AI "is employed within select third-party solutions integrated into our operations," and Princeton Bancorp's names them: Sophos, Verafin, the Glia chatbot and Cato Networks. 21 of the 50 banks (42%) mention AI in their SEC filings only as a risk factor.

How was the study done?

Each sample is a seeded random draw of 50, stratified by total assets, from a defined population. The research window covers disclosures dated 2024-01-01 to 2026-11-08.

  • Banks: 50 of the 246 US banks and thrift holding companies filing 10-Ks with 2025 bank revenue of $50 million to $2 billion, where bank revenue is net interest income plus noninterest income. The band covers about $1 billion to $50 billion of assets.
  • Credit unions: 50 of the 541 US credit unions filing NCUA's December 2025 call report with 2025 revenue of $50 million to $2 billion, where revenue is total interest income plus total non-interest income. The band covers about $0.6 billion to $35 billion of assets.
  • Sources searched: for banks, 10-Ks, 10-Qs, annual reports, proxies, 8-K exhibits, earnings calls, press releases, investor decks, bank websites, vendor case studies and trade press. For credit unions, press releases, their own websites and annual reports, vendor case studies and trade press, plus web searches pairing each credit union's name with AI vendors.
  • Who did it: seven research agents (Claude Opus) per sample, 7-8 institutions each, plus one follow-up agent for 18 credit unions whose vendor searches had been cut short.
  • What counts: a specific AI initiative named in a public source. A chatbot or virtual assistant counts only if its source says AI or names an AI vendor. AI in something an institution sells to others is left out.
  • What counts as a figure: a result (a rate, a before-and-after, cash or hours) or work the AI did (a share of applications decided, inquiries resolved). A count of people using a tool, with no result, such as a share of staff using AI, is recorded but isn't counted as a figure.

What are the study's limits?

The samples are small: 50 each, so one institution is 2 points. The two revenue measures also differ. Bank revenue is net of interest expense and credit-union revenue is gross, so the two bands don't hold the same size of institution: the sampled credit unions' median is $1.9 billion of assets, against $5.5 billion for banks.

  • Search gaps. A web-search limit shared by the parallel credit-union agents ran out partway through. A follow-up agent re-ran vendor-name searches for 18 credit unions and found 5 more records, but several were still searched with combined queries. Some of the 25 credit unions with nothing found may be search gaps rather than silence.
  • Earnings calls. A later pass listed 319 bank calls and conference transcripts in the window and read 310; 16 banks hold no calls that could be found, and Arrow's quarterly calls have no free transcript.
  • Blocked or unreadable sources. Some news sites, press-release wires and institutions' own sites refused automated reads, and many credit-union annual reports are flipbooks or image-only PDFs.
  • Dates. 21 of the 72 credit-union quotes are undated, mostly vendor customer stories.
  • Judgment. Mapping and grades were made by research agents under one written rubric, then reviewed by a person, except the 22 quotes from the earnings-call re-check. 23 bank initiatives and 21 credit-union initiatives carry review flags, among them possible duplicates and weak links to AI.

Has this study been corrected?

Yes, twice, both on November 10, 2026.

  • A count of people using an AI tool isn't a figure. Banc of California's only figure, "more than 80% of our developers using AI in their daily workflows," is a usage figure, so it no longer counts as a figure. The same rule moved one manufacturer.
  • The earnings calls were re-read. A re-check of the banks' calls, conferences and annual meetings (310 of the 319 found) added 22 quotes the first pass had missed, among them Metropolitan Bank Holding's measured $200,000 a year from CoCounsel. Banks that give any figure went from 5 (10%) to 6 (12%), cash figures from 2 to 3, and cash credited directly to AI from 1 to 2. The insurance and manufacturing columns reflect the same re-check of their own samples: insurers that give a figure rose from 14 to 16 and those crediting cash directly fell from 2 to 1, and manufacturers that name an initiative rose from 24 to 27.

No credit-union figure changed.

Who ran the study?

Roiva ran this study. Roiva is AI portfolio management software that tracks every AI initiative's budget, approvals, cost and ROI, so whether institutions can show a return on AI is a problem the company sells into. That is a reason to check the work.

The uses this study found most often, and the gaps it turned up, shaped four of Roiva's financial-services templates: Banking Virtual Assistant & Contact Center AI, Commercial Credit Spreading & Memo Drafting, Deposit & Loan Operations Automation Agents, and Personalized Financial Insights & Coaching. They sit beside the existing AI Credit Underwriting & Decisioning and Fraud Detection & Transaction Monitoring templates. None of the figures above went into them.

Questions about the data?

Ask us about the method or the per-institution records.

Get the next one

New field notes by email as they're published. Or subscribe to the RSS feed.