How we did this
What this site covers
This site shows what US banks say about artificial intelligence (AI) in the reports they file with the SEC. It covers 325 banks and 12050 filings, from 1 January 2023 to 6 Oct 2026.
Banks included. US banks and bank holding companies that file reports with the SEC under the industry codes for national and state commercial banks, plus seven large bank holding companies that the SEC files under other codes (Goldman Sachs, Morgan Stanley, Charles Schwab, Raymond James, American Express, Synchrony and Discover). Savings institutions are not included in this version.
Banks not included. Most small US banks do not file reports with the SEC, so they are not here. That is why few banks under $1 billion appear. Credit unions are not included.
Reports included.
- Annual reports (10-K), including amendments that restate the whole report.
- Quarterly reports (10-Q).
- Earnings releases and investor presentations, filed with the SEC as attachments to 8-K filings.
Report year
A report year is the year a report covers, not the year it was filed. An annual report filed in February 2026 about 2025 is a 2025 report. Charts of annual reports use report years 2022 to 2025. A report year appears only when at least 50 annual reports for it have been filed.
How AI mentions are found
Every report is downloaded from the SEC and split into passages: paragraphs, and single slides for presentations. A passage is kept when it contains a term from a published list, for example “artificial intelligence”, “AI” in capitals, “machine learning”, “generative AI”, “large language model”, “agentic”, “AI agents”, “chatbot” or “virtual assistant”. The full list is on the download page. “AI” in capitals is matched only as a word on its own, so words such as AIG are not counted.
Passages inside the forward-looking statements notice that many reports include are marked as legal notices and never counted, because they list AI alongside many other risks without saying anything specific.
We checked this step on 20 reports read by hand. It found every AI mention a person found, with no false matches. A wider check across all annual reports matched the SEC’s own full-text search for “artificial intelligence” in every year.
How passages are labelled
Each passage is labelled by an AI model (Claude Sonnet 4.6 on Amazon Bedrock) and checked by hand on a sample. The model sees the passage, the bank, the kind of report and the section, and answers:
| Label | Meaning |
|---|---|
| Using AI now | The bank says it uses AI now, for a stated purpose |
| Testing or planning AI | Testing, piloting, building, investing in or planning to use AI |
| Explains how AI is controlled | Board or committee oversight, AI policies, model risk management that covers AI, human review, staff training |
| Sees AI as a risk | AI as a risk to the bank: its own use, vendors, competitors, fraudsters, regulation |
| General statement about AI | A belief or intent about AI, or AI in the economy or industry, with no specific use |
| Standard wording or passing mention | Nothing about the bank itself: a glossary line, a list of technologies, a biography |
When a passage fits several labels, the first in this order is used: using now, testing or planning, controls, risk, general statement, passing mention.
Detail level. General: the passage could appear in any bank’s report. Names an area: it says where the bank uses or plans AI, or names its own committee or program. Concrete example: it names a tool or vendor, gives a number, a date or a result.
The model also quotes the words that support its label. The quote must appear exactly in the passage; when it does not, the passage is held back from all counts until checked.
Accuracy. On 8 Oct 2026 the labels were compared with 100 passages labelled by hand (26 by the site’s author and 74 drafted by a second AI model and awaiting the author’s review). The labels agreed on whether a passage is about AI 99% of the time, on what it says 90% of the time, on the exact detail level 82% of the time, and on detail level within one step 98% of the time.
Standard wording
Many banks use almost the same sentences, often from the same law firms. A sentence counts as standard wording when five or more banks use it, almost word for word, in the same report year. A passage is marked “Similar wording appears in N other banks’ reports” when most of its words are in such sentences. Bank names and self-references such as “the Company” and “our” are ignored when comparing.
New this year
A passage is new when most of it does not appear in the bank’s previous report of the same kind: the previous annual report, the previous quarterly report, or the previous quarter’s earnings materials. For earnings materials we also mark passages that are not in the bank’s latest annual report.
Bank size
Size is the total assets of the bank’s FDIC-insured banks at the end of the report year, from FDIC call report data. It is not the same as the holding company’s consolidated assets, which can include other businesses. Each SEC filer is matched to its FDIC holding company by name and state; matches that are not certain are checked by hand. Size groups follow Bankgraph: under $1 billion, $1 billion to $50 billion, $50 billion and above.
Known limits
- The site shows what banks choose to say in their reports, not what they actually do with AI.
- Labels are made by an AI model and can be wrong. Every label links to the passage so you can check it.
- “Lists AI as a risk, but the report does not say how AI is controlled” means no such text was found in the report. It does not mean the bank has no controls.
- Earnings call transcripts are not included.
Corrections
If a label is wrong, use “Report an error” next to the passage, or email the address on the about page. Corrections are checked by hand and applied the next time the site is rebuilt. A corrected label replaces the model’s label everywhere on the site and in the downloads.