How the BankingNewsAI Bank AI Disclosure Index is calculated

What goes into the score, and how much does each part weigh?

The score has eight components, each read from the bank's page on this site. Independent coverage carries 45% of the weight because a page built from a bank's own newsroom and job postings can be long without the outside world having noticed. The other seven components cover what the bank itself has said and done: use cases in production, a named AI platform or programme, its own primary disclosures, dated moves in the last twelve months, disclosed numbers, named leaders and all use cases described. The weighted sum is multiplied by 100, so a bank would need to lead on every component to reach 100.

How are the components scaled?

On seven components a bank receives its percentile rank among the 120 banks: (average rank minus 1) divided by (number of banks minus 1). The lowest value scores 0 and the highest 1, and banks with the same value share the average of their ranks. A named AI platform or programme is not ranked: it is 1 if the page names one and 0 if not.

Percentile ranks replaced fixed caps on 2026-09-24. Once every bank page had been researched in depth, fixed caps saturated (several banks sat at exactly 100) and raw counts moved with research effort. A percentile can only separate banks; it never saturates. The score therefore shows where a bank's record sits against its peers, not whether it cleared a fixed bar.

What happens to a bank that has disclosed no AI programme?

A bank whose page records that it has not disclosed an AI programme, leader, vendor or budget is capped at 30 and kept at least 0.1 below the lowest score of any bank that has disclosed one, so it ranks below all of them. The cap scales those banks' scores rather than clipping them, so they keep their order among themselves. Their pages record what is known instead, usually customer education, fraud guidance or a core-banking conversion.

How are ties broken?

Banks that disclose an AI programme come first. Within each group banks are ordered by score, rounded to one decimal place; equal scores are ordered by the number of independent sources, then by total assets.

How are sources classified?

Each source cited on a bank page is sorted into one of four kinds, by its publisher, address and title. Only independent sources count toward independent coverage. Own primary disclosures are the own sources counted once per distinct title, so a release and its wire copy count once.

• Own: the publisher is the bank (its name, short name or lead bank as a whole phrase, never a single common word), or the page sits on one of the bank's own sites. Earnings-call transcripts and summaries count here too, because they carry the bank's own words, and so does a bare wire copy of a release that leads with the bank's name.
• Vendor: a technology or consulting vendor's site or release, and anything marked sponsored, partner content or "presented by".
• Wire: wire copies of other organisations' releases, third-party filings, job boards and automated aggregator write-ups. Evident's ranking is also counted here: it is a third-party ranking, cited on a page only as an attributed line.
• Independent: everything else, which is reporting by outlets that chose to cover the bank. A story syndicated on Yahoo Finance is judged by the outlet it names, not by the Yahoo address.

What does the score not measure?

The score measures disclosure, not capability. It rewards a bank for saying what it does and for being reported on by others. It does not judge how well a programme works, how much a bank spends on technology, or whether its models are safe. It under-rates banks that build quietly: a bank that runs a large programme and describes few systems publicly scores lower here than its reputation. Use it as a map of where the public record is deep, then read the bank page.

How often is the index updated?

Recomputed on every bank page update; frozen quarterly editions. The live ranking on /banks/ai-index is recalculated from the bank pages whenever one of them changes (last data date Oct 21, 2026). Each quarter the ranking is frozen as a dated edition with its own page and CSV, which never changes after publication, so a citation of an edition always points at the same table.

Where is the data, and how do I cite it?

The live ranking is on /banks/ai-index, with the live data as CSV. Frozen editions: Q4 2026 (CSV, frozen Oct 22, 2026). Each CSV has the columns rank, bank, slug, score, the eight component values, undisclosed, as_of and edition.

Cite an edition when the numbers must stay fixed, for example: BankingNewsAI Bank AI Disclosure Index, Q4 2026 edition, BankingNewsAI, https://www.bankingnewsai.com/banks/ai-index/2026-q4