Banking has more AI in production than any other industry we index.
It also has the widest gap between what is announced and what is running. This is what the register holds on banking: 64 vendors, 1,340 customer claims, and the 407 of those that an organization has confirmed.
Three in ten banking claims survive contact with the customer
Of claims taken to the organization named in them, most confirm a relationship. Far fewer confirm a deployment past one business unit.
Register edition 2026.4 · compiled 18 Oct 2026
Why banking looks further ahead than it is
Banks have been running machine learning in production for two decades, which makes the industry genuinely more advanced than most and also makes it easy to mistake that inheritance for progress on the current wave.
The inheritance is real
Fraud scoring, credit models, surveillance and anti-money-laundering have been model-driven since long before the current cycle. Banks have model risk functions, validation practice and regulators who already know how to ask about a model, which is a starting position most industries do not have.
The new work is concentrated in text
Almost everything deployed since 2024 touches documents, cases or conversations rather than numbers. Credit memos, onboarding packs, complaints, alerts, disputes. The bank's existing model governance was built for scores, not for language, and that mismatch is where most programs slow down.
Deployment stops at the business unit boundary
The most common pattern in the register is a deployment that works in one country, one product line or one operations centre, and has not crossed into the next one eighteen months later. The constraint is rarely the model and almost always the process variation underneath it.
The regulator is not the blocker people claim
Supervisory expectation shapes the sequence, not the ceiling. Banks that moved fastest treated explainability, audit trail and human override as design inputs from the first week rather than as a gate to clear at the end.
Nine categories, and they are not moving at the same speed
Vendor counts, claims captured, and how many of those claims an organization has confirmed. Absence of confirmation is as much a finding as its presence.
| Category | Vendors | Claims | Confirmed | Deepest stage established |
|---|---|---|---|---|
| Document and submission processing | 14 | 312 | 118 | Full deployment |
| Financial crime and AML operations | 11 | 241 | 96 | Full deployment |
| Contact centre and servicing | 13 | 268 | 84 | Full deployment |
| Credit and lending operations | 9 | 164 | 51 | Partial deployment |
| Payments and disputes | 8 | 121 | 29 | Partial deployment |
| Risk, compliance and controls | 10 | 118 | 18 | Partial deployment |
| Wealth and advisory support | 7 | 63 | 9 | Partial deployment |
| Capital markets operations | 6 | 38 | 2 | Yet to start |
| Agentic customer operations | 12 | 15 | 0 | Yet to start |
Read the last two rows carefully. Twelve vendors are selling agentic customer operations into banking and not one claim has yet been confirmed by the organization named in it. That is not a verdict on the category. It is a statement about what exists today.
Two ways to engage us, and two things included with either
Everything rests on the same register. We do not introduce you to vendors, take referral fees, or accept sponsorship from anyone selling into banking.
Expert guidance
A standing line to the people who maintain the register. Bring a shortlist, a category you are about to enter, a proposal on the table, or a program that has stopped moving.
- Inquiry calls and written responses inside 48 hours
- Shortlist reviews and proposal reads before you sign
- Bought as sessions, not seats, with register access included
Executive workshops
Your own decision, worked through by the people who have to own it, against a banking market picture drawn from the register rather than from vendor material.
- A function, the leadership team, or the management layer
- Our maturity assessment, frameworks and canvas, brought to the room
- Moves chosen, owners named and a written record in five working days
Category review
Commissioned. Every vendor in a banking category, every claim across them, and where the working deployments sit. Published like any other determination.
Register access
Included with guidance. Search every banking vendor and claim, filter by stage and currency, export the data, and get alerted when a determination lands or lapses.
What 1,340 banking claims look like when you follow them through
Every number below is a count from the register, not a survey, an estimate or a projection. Quote them with the edition and date attached.
Thirteen per cent of banking claims are confirmed running at full scale
Each step is the subset of the one above it that survived the next layer of evidence.
Confirmation rate varies eightfold across banking categories
Share of captured claims that an organization has confirmed.
Where the fifty use cases actually sit
Stage mix within each of the six groups.
The sequence most banks get backwards
Where each horizon sits in time, and how much confirmed evidence stands behind it today.
Fifty banking AI use cases we track
Six groups, fifty use cases, and against each one the stage the register can currently confirm. Live means a confirmed full deployment exists somewhere. Getting started means partial deployments exist. The third label is our forecast, and we mark it as one.
Counts and stages are drawn from register edition 2026.4, compiled 18 October 2026. A use case marked 2027 to 2028 has no confirmed full deployment anywhere in the register today, and our view is that the remaining constraint is technical rather than organizational.
Where banking actually sits, function by function
Each cell is the number of confirmed deployments the register holds at that stage for that function. Read across a row to see how far a function has got. Read down a column to see how thin the right-hand side still is.
Operations, financial crime and engineering carry weight on the right-hand side. Lending, risk and capital markets are still concentrated on the left, which means claims rather than deployments.
Forty-seven claimed against two at full deployment. More vendors sell into this function than into any other, and almost nothing has crossed the line.
Not one confirmed deployment at full scale anywhere in the register. If a vendor tells you otherwise, ask them to put the claim forward.
What actually stops banking deployments, in the order it stops them
Drawn from the enquiries behind confirmed and unconfirmed claims. Regulation is on this list, but it is not at the top of it.
Process variation across business units
The deployment works where it was built and meets a different version of the same process in the next country or product line. Most stalled banking programs stall here.
Data residency and lineage
Not whether data can be used, but proving where it went and who touched it. Deployments designed without lineage from day one get rebuilt rather than extended.
Model governance built for scores
Validation practice that works for a credit model does not transfer cleanly to a language system, and the committee that must approve it has no precedent to work from.
Ownership after go-live
The program team hands over to an operations function that was not resourced for it. Six months later usage is flat and nobody can say whether it worked.
What we have put out on this industry
Everything below is free to read and draws on the same register. None of it is sponsored, and no vendor saw any of it before publication.
Identity and onboarding agents in retail banking
Nineteen claimed deployments taken to the banks named in them. Eleven confirmed, five overstated, three unanswered.
October 2026 Use-case guideDocument processing in trade finance
What is running, what the operations teams changed to make it run, and the handover step that separates the two.
September 2026 Research briefingModel governance for language systems
How banks that cleared validation did it, condensed for a risk committee.
August 2026Bring us the banking decision in front of you
A shortlist, a category you are about to enter, or a program that stopped moving. We will tell you what the register can answer, what it cannot, and which of our services fits. If none of them does, we will say that.
What banks ask about this page
From the register. Every claim is captured from a vendor's own published material, and the confirmed ones have been taken to the organization named in them and answered under anonymity. Counts are as at the compilation date on the page, not estimates, projections or survey responses.
Because no claim in that category has yet been confirmed by the organization named in it. Twelve vendors are selling it into banking and fifteen customer claims exist, all at Level 1. That may change next quarter. It has not changed yet, and we would rather print a zero than imply otherwise.
No. It means no vendor has published a claim that has been confirmed at that stage. Banks that built it internally are not in the register, because the register is a record of vendor claims rather than a census of the industry. Internal builds are a real gap in this view and we say so rather than pretending otherwise.
It carries an edition number and a compilation date, and it is rebuilt every quarter. Individual determinations underneath it carry their own dates and lapse to unconfirmed after twelve months, which is why the counts move down as well as up.
Either we have not opened a record on them yet or they sit outside the five technology categories we cover. Request them and we will open a record, capture every customer claim on their site and publish it. That is free and usually complete inside ten working days.
Yes, and it is the only part of this page that is. It reflects our judgement about where the remaining constraint is technical rather than organizational, and we revise it each edition. Everything in the first two groups is a record of what has happened, not a view about what will.
Yes, and the register records the region of every organization that answered. Confirmed deployments in the register currently skew to North America and Western Europe, with a growing share in the Gulf, India and Southeast Asia. Region-level breakdowns sit behind register access.
Yes. A shortlist check takes the vendors in front of you and runs every claim each of them makes in your category through the method, with the determinations and the full log returned to you and published to the register like any other.
Before the next banking AI decision goes to committee
Check what the register holds on the vendors in front of you, commission a category review, or bring the decision into a room with your own leadership team.