AI in Financial Services: Giant Rollouts, Proof by the Dropperful

A fully equipped sector, and proof still pending

After insurance last week, financial services push the reversal to its conclusion: here, the adoption question is settled. 92% of European Union banks are deploying AI, and 55% already use generative or agentic AI in customer-facing processes (EBA). In the United Kingdom, 75% of the financial firms surveyed by the regulator use it: a median of nine use cases per firm, twenty-one expected within three years (Bank of England / FCA). The United States tells the same story at census scale: 33.9% of finance and insurance businesses use AI, against 19.8% nationally (US Census Bureau). The Federal Reserve already placed the sector at ~30% by late 2025, against 18% across all sectors (Federal Reserve). In France, the national statistical survey leaves the financial sector outside its scope, but the supervisor observes that the vast majority of institutions have use cases in production; the national average of 18% sets the order of magnitude the industry comfortably exceeds (INSEE; ACPR).

The explanation is structural: finance works a material that is already digital: transactions, documents, conversations, all born as data. Rollouts therefore happen at a scale no other sector displays: 200,000 employees equipped with generative AI at JPMorganChase, mandatory prompting training for some 175,000 staff at Citi, assistants opened to the whole firm at Goldman Sachs (JPMorganChase; Citi; Goldman Sachs).

And yet: of the fifty largest global banks tracked by the sector's reference index, only four publish a realized return on investment for their AI (Deloitte; Evident). Across 250 retail banks in 14 markets, 6% have an enterprise-scale AI roadmap, and 2% of executives regularly track performance indicators for their generative AI (Capgemini). Licenses get deployed; instruments of measurement rarely do. And the warning carries to the next wave: more than 40% of agentic AI projects will be stopped before the end of 2027, for lack of demonstrated value and controls (Gartner).

This staircase structures the whole report, and it has an immediate strategic consequence, whatever the institution's size: since almost no one measures, measuring has become the cheapest competitive advantage on the market. The four banks that publish their figures are not four lucky banks: they are four banks that instrumented their use cases from day one, and therefore know where to reinvest. Their lead already shows: the score of North American banks exceeds the rest of the world by 20%, seven of the top-10 places are theirs, and the gains of the first, about $2 billion of declared value, finance what comes next (Evident; JPMorganChase).

The sector that deploys the most is also the one that proves the least, not because the value does not exist, but because few institutions instrument it.

Fraud, margins, headcount

If finance is accelerating, it is not out of comfort. First front: fraud, where AI now arms both camps. More than $485 billion in global payment-fraud losses in 2024 (Mastercard); €4.2 billion in the European Economic Area (ECB / EBA); a record $15.9 billion in fraud losses reported by American consumers in 2025 (FTC). France illustrates the mutation: €1.241 billion of payment fraud in 2025, and above all "manipulation" fraud (the fake bank adviser, the fake transfer), now 40% of the total, a share that has doubled in four years, while the card fraud rate falls to a historic low (Banque de France, OSMP). Where the technical defense wins, the attack moves to the human.

Second front: costs. Banks spend about $600 billion a year on technology, for a productivity their own industry judges disappointing (McKinsey); global compliance penalties reached $3.8 billion in 2025, with EMEA up 767% and France now the second most punitive jurisdiction (Fenergo). Third front, the most sensitive: headcount. A survey of 93 technology chiefs at major banks anticipates up to 200,000 fewer positions within three to five years because of AI: a projection, not a measurement, but one the sector itself owns (Bloomberg Intelligence); 54% of banking jobs show high automation potential, more than any other sector (Citi GPS); and the first European strategic plan to put numbers on the equation has arrived: €1.6 billion of savings and more than 12,000 departures by 2029 at Intesa Sanpaolo, "with no social cost" (Intesa Sanpaolo). In France, banking headcount is already receding, at 373,600 employees at the end of 2024 (FBF).

In a business whose raw material is information, AI is therefore not one project among others: it is the variable that decides the cost of risk, the cost of production and the size of the teams, all at once. The question is no longer whether to go: everyone has gone. The question is who actually collects the value, and the rest of this report takes it from the angle that settles it: where AI stands relative to the financial decision.

Three decision regimes: the map that matters

Sorting use cases by business line (customer relations, credit, compliance) describes the org chart, not the risk. What actually governs the cost, the timeline and the governance of a financial AI project is the machine's position relative to the decision: paying, declining, advising, lending. The twenty-odd documented deployments in this report fall into three regimes: AI serves, AI prepares, AI decides. And everything an executive cares about (return on investment, timelines, regulatory file, the first indicator to put in place) follows from the regime chosen.

Choosing a use case means choosing a regime and, with it, the return you can expect, the timeline and the scale of the governance file. This framework prohibits nothing: it scales the requirements.

The Paris market publishes its numbers

The French market does not merely follow the movement: it puts numbers on it, with a particularity few markets share: a national model provider at the heart of the rollouts. Four institutions suffice to show it.

The movement extends beyond the big networks. Crédit Agricole has announced about €500 million of AI investment over three years, explicitly favoring European solutions (Crédit Agricole): sovereignty has become a purchasing criterion, and Mistral AI its natural beneficiary. The fintechs, for their part, are moving to the AI-native model: execution and analysis agents at Qonto, an accounting copilot at Pennylane (Qonto; Pennylane). Read against the three regimes, these French numbers say something precise: the market excels in the first regime and is investing in the second. And that is exactly where the question of proof arises, because a published target (€750m, €500m) is not yet a realized return. Publishing a target is a first step; it is the published measurement that moves you up the table.

Four fault lines specific to finance

When an AI project stalls in finance, it does not stall the way it does elsewhere. Crossing the field evidence, the four banks that prove as much as the majorities that deploy without measuring (Deloitte; Capgemini), brings out four fault lines this sector shares with no other. All have one thing in common: the decision that neutralizes them costs nothing, provided it comes before the launch.

Fault line 1: licenses before instruments. The organization deploys thousands of generative-AI licenses, and no instrument of measurement: no starting figure, no tracked indicator, no publishable ROI. It joins the 94% without an enterprise roadmap and the 98% of executives who track no GenAI indicator (Capgemini). The banks that prove do the opposite: JPMorganChase puts its value at ~$2 billion, BNP Paribas publishes a €750 million target and stands by it publicly (Evident; BNP Paribas). What prevents it: instrument every use case from day one: proof is built, it is not observed after the fact.

Fault line 2: GenAI outside the perimeter of model validation. Banking is the only sector where model risk has been a regulated discipline for decades. Yet generative assistants are often deployed beside that apparatus, like ordinary office tools. The day one of them touches a decision (a score, an eligibility, an amount), the project discovers it falls under the high-risk regime and the requirement of specific denial reasons (AI Act; CFPB), and starts over. What prevents it: attach generative AI to the existing model governance from the first deployment: finance already has this muscle, and using it costs less than building a second apparatus.

Fault line 3: the assistant that drifts toward advice. A generated text that compares, recommends or steers toward a product stops being service: it becomes advice, with its duty, its traceability, its liability; a hallucination then becomes a breach. That is the boundary accomplished deployments do not cross: at Morgan Stanley, the assistant prepares the meeting and the notes, the adviser speaks and decides (Morgan Stanley); in the UK market, only 2% of use cases let the machine decide alone (Bank of England / FCA). What prevents it: write down the servicing-advice boundary of the use case before the first line of development: AI prepares, the human advises.

Fault line 4: the fraud arms race. This sector's absolute particularity: the adversary equips itself at the same pace as the defense. While strong authentication drives technical fraud to a historic low (0.048% on cards in France), manipulation fraud doubles its share in four years and reaches 40% of the total (Banque de France, OSMP): the attack goes around the system and targets the human. A single bank's isolated model loses this race. What prevents it: defend as a network (shared signals, payee verification, AI carried by the payment networks, as Visa, Mastercard and HSBC practice it) and train customers at the same pace as models.

Where to start: three questions, four paths

There is no universal first step: the right move depends on where the institution starts. Three questions are enough to place any executive team in the sector, and each of the four possible answers responds by name to one of the fault lines above.

Question 1: Do you publish today a validated value indicator on at least one use case? If the answer is no, the first project is not one more use case: it is instrumenting what already runs. Pick the most advanced deployment, set the starting figure (cost per contact, minutes per file, hours per process), measure the arrival on the same perimeter; a quarter is enough, and it is the direct answer to fault line 1. As long as that indicator does not exist, every new project adds licenses to a portfolio nobody can defend in committee.

Question 2: Does your model-validation apparatus cover generative AI? If not, extending it comes before any committing deployment: an inventory of assistants in production, attachment to model review, a written rule on the servicing-advice boundary: the answers to fault lines 2 and 3, with a single deliverable: a governance review attached to every costing, before signature. Finance is the only sector that already owns this discipline; not using it means paying twice.

Question 3: Does your next use case touch the decision: a credit, a denial, advice? If not: industrialize the second regime, at the scale of the ambition: Citi's mandatory training or BPCE's "one employee in two" target belong in the budget, not in the addendum, and the total cost includes core-system integration and data consolidation. If yes: open the high-risk file at costing time: AI Act classification, documented denial reasons, ACPR audit domains, with the December 2, 2027 deadline written into the project plan, not discovered along the way.

There remains the path nobody chooses but the majority takes: deploying without measuring. It has a price that can be counted: the gap with the leading banks widens by 20% and self-finances on their side (Evident), unblocked fraud is paid for good, and the regulatory deadlines will turn ungoverned cases into liabilities. In a sector where everyone deploys, the ranking is no longer made on ambition or on licenses: it is made on instrumentation: not more licenses, more measurement.