In finance, artificial intelligence is already everywhere, and almost nowhere the client would see it. Nine out of ten French financial market participants use it or will adopt it within the year. Yet most of those use cases never leave the back office.
Tandem works with around thirty companies on AI applied to operations, including several investment funds and asset managers. This article gives the grid we use on assignments. It ranks twelve projects by value, feasibility and regulatory risk. The first one is not the one boardrooms talk about.
Where does AI in French finance actually stand?
The French financial markets authority measured actual adoption. Its study published in February 2026 covers 100 respondents, asset managers, investment service providers, listed companies and audit firms. 90 % of them already use AI or plan to adopt it within twelve months. 54 % have at least one use case in production, and 72 % have formalised an AI governance policy.
Size makes the difference. Among entities already using AI, half are large companies. Among those declaring no use case at all, the vast majority are small structures and micro-companies. In asset management firms, the AI budget generally stays under 1 % of IT spending.
Large banking groups are moving with dated announcements. La Banque Postale signed a three-year partnership with Mistral AI in May 2026 to deploy language models on its own servers, in a sovereign environment. The first tier targets 5,000 employees in 2026. Stéphane Dedeyan, chair of the management board, says he wants to combine economic performance, sovereignty requirements and sustainability.
BNP Paribas renewed its own agreement the same month, also for three years. The announced scope covers corporate and institutional banking as well as commercial banking. Above all it covers KYC, where AI takes over low-value steps while expert validation is kept (Retail Banker International, May 2026). Two groups, two timelines, one shared logic. AI enters through the teams before it enters through the client.
Why finance deploys AI internally before showing it to clients
The imbalance is stark, and it is measured. Across the 106 use cases detailed to the French financial markets authority, 83 % concern internal tools. Writing, summarising, translation, copilot-style assistants, data quality, code generation. Only 17 % touch the client relationship. And 1 % target the provision of an investment service as such.
This is not timidity, it is a perfectly rational risk calculation. An internal assistant that gets it wrong costs a review. An automated recommendation that gets it wrong exposes the firm to its client and to its regulator. The two projects share neither timeline, nor control cost, nor exposure.
Augmentation or automation, two projects and two timelines
Every use case belongs to one or the other, and confusing them is the most common cause of overrun. In augmentation, AI prepares and the analyst decides. The portfolio manager, the compliance officer or the controller stays in charge of the file. In automation, the flow runs and the team only handles exceptions.
| Augmentation | Automation | |
|---|---|---|
| Time to service | 2 to 6 weeks | 3 to 9 months depending on system access |
| Nature of the gain | Time per task, consistency of deliverables | Unit cost of the flow, processing time |
| Ease of measurement | Low, time saved is self-reported | High, volume and rework rate sit in the logs |
| Regulatory exposure | Low, a human remains the author of the decision | High, traceability, client information, human recourse |
| What makes it fail | Adoption, an unused tool produces nothing | The data reference, dirty data blocks the whole chain |
There is a level before automation, and it stays badly under-used. I filmed a whole video on the fact that the AI agent is too often the first reflex, when a configured assistant or a simple automation is enough in most cases. The four levels, from assistant to autonomous agent, are detailed in our article on the levels of automation.
That threshold calculation is exactly what a risk department expects. A success pays, a detected failure costs a rework, an undetected wrong answer costs far more. Setting those three values, then crossing them with the measured accuracy rate, gives the threshold above which the flow is worth it. In finance that threshold sits higher than elsewhere, and rightly so.
Twelve finance AI projects, ranked by value, feasibility and risk
Three scores from 1 to 5 are enough to decide. Value, which adds freed capacity to gains beyond time. Feasibility, which measures data access and the existence of written rules. Regulatory risk, which runs from internal data with no client impact to a decision affecting the client. The index is value times feasibility, divided by risk. The full logic sits in our AI audit method.
The result is surprising, and it holds. The leading project is financial modelling and file reprocessing, a topic nobody presents to a strategy committee. KYC, the project large banks cite most, only comes ninth, because it touches client data and control obligations. Credit scoring closes the list.
The spreadsheet remains a finance team's first project
The first pocket of value in finance is not an agent, it is the calculation file. Models, export reprocessing, reconciliations and reporting all live in a spreadsheet. They are handled by people whose time is expensive, on data that never leaves the company. High value, maximum feasibility, minimal regulatory risk.
What used to take me three hours now takes a few minutes. I summarised the manipulations that work in a cheat sheet published on LinkedIn in June 2026, and detailed the three options in my newsletter edition on the subject. Building a three-statement model, scenario sensitivity analysis, cleaning a chaotic export, explaining an inherited formula nobody dares touch any more.
Three tooling levels answer each other, and the choice depends on volume. The add-in inside the spreadsheet covers daily work. A file environment handles producing a complete file and exporting it to a shared sheet. A coding agent takes the heavy consolidations, fourteen monthly exports with inconsistent headers or a file of tens of thousands of rows. The detail sits in our article on AI applied to Excel and data.
What the regulatory frame actually allows in finance
Four texts frame an AI project in a financial institution, and they are not read in random order. DORA has applied since January 2025 and requires a register of critical providers. The EU AI Act has applied since August 2026, with its article 50 transparency obligations. The omnibus regulation that entered into force in July 2026 postpones Annex III high-risk obligations to 2 December 2027. And the GDPR applies continuously.
Two financial uses are explicitly classified as high risk. Assessing a natural person's creditworthiness for lending, and pricing in life and health insurance. For those systems the market surveillance authority is the French prudential supervisor, not the European Central Bank. Financial fraud detection is explicitly excluded from the category.
- What is open today. Everything handling internal data without deciding in place of a human. Corpus, summarisation, control, file preparation.
- What requires explicit information. Any interaction where a client speaks to a machine. Article 50 requires that the person knows it.
- What requires traced human review. Any decision with a significant effect on a person, under article 22 of the GDPR.
The insurance neighbourhood follows the same logic, with its own thresholds. We detailed it in our article on AI in insurance, and on the accounting side in the one devoted to AI use cases in accounting firms.
Build or buy, the question that has already cost money
The reflex to build your own assistant has a documented history in French banking. Société Générale had deployed its internal tool, SoGPT, to front and back-office teams. The group started decommissioning it in late 2025 in favour of a market solution, as the gap with available tools widened too fast (L'Usine Digitale, January 2026).
The lesson is not that you should never build. It is that standard bricks are no longer built. Generation, transcription, search, conversational assistant, those layers ship new features every other week. An internal team reimplementing them spends its time catching up with the state of the art instead of creating business value.
What gets built internally is what is specific to you. Control rules, the document corpus, connectors to your systems, quality measurement. We detail that split in our review of why most companies see no return on their AI.
What a first AI deployment costs in finance, and how fast it pays for itself
The calculation holds in five variables, all obtainable within a week without a data project. The annual volume of items in the flow, the average time per item measured rather than estimated, the loaded hourly cost of the function, a realistic automation rate net of rework, and the full project cost. Freed capacity is the product of the first four.
- A realistic rate, never a theoretical one. Count only the items that pass quality control without rework. On a heterogeneous document flow, the observed starting point sits around half. On a standardised flow with a clean data reference, it climbs much higher.
- Freed capacity is not cash. Separate the share redeployed onto other tasks, which is real value but invisible in the accounts, from the share in avoided cost. Only the second interests a finance department.
- Run cost is budgeted from day one. Inference, hosting, supervision, exception handling, rule maintenance. That is the line business cases forget most often.
On our assignments in funds and their portfolio companies, the payback on the initial investment usually lands under six months, and sometimes under two. I said as much in a conversation for the La Reprise podcast in May 2026. Those orders of magnitude apply to well-chosen augmentation projects, not to a system overhaul.
One figure from the French financial markets authority study deserves framing. 92 % of surveyed firms acknowledge efficiency gains, but 64 % report having no precise indicator to measure them. An unmeasured gain cannot be defended in a steering committee, and it cannot be reconstructed afterwards. Measure current volume, time and error rate before launching anything. A week of sampling is enough.
What AI cannot do in finance
An honest article on the subject also says where to stop. Three limits keep coming back on our assignments, and none of them is solved by a better model.
- It does not replace a missing data reference. If product, client and contract data do not reconcile, no agent will close the gap. It will propagate it faster.
- It does not sign a regulated decision. A system can investigate, prepare and recommend. Responsibility stays entirely with the institution and the authorised person.
- It does not reach 100 %. A flow covered at 80 % with clean human rework beats a promise of full coverage that will never clear internal control.
A fourth limit is cultural rather than technical. A model that cannot say it does not know has no business near a published figure. Require your system to flag its uncertainty, that is the first selection criterion in finance.
Do you need an AI agency to deploy AI in finance?
Not always, and the answer hinges on one thing. If your processes are written down, if one person owns the topic full time and if your data is accessible, an internal team will do fine on the first projects. Otherwise the learning cost is paid in lost months rather than billed days.
- What an external team really brings. The number of cases already seen elsewhere, and therefore the ability to rule out early the projects that will not clear. That is where most of an audit's value sits.
- What to require in the proposal. A measured starting point, a written exit criterion per phase, and a run cost stated before the build. A proposal without those three is a proposal for a permanent pilot.
- What must stay with you. Ownership of the corpus, the rules and the connectors. An agency keeping its hand on those bricks is selling you a dependency, not a capability.
Tandem works on this perimeter with financial players, from investment funds to asset managers. The offer and the use cases by function sit on our AI agency for finance page, and the full approach on the AI audit page.
Which AI project should finance start with?
Start with the spreadsheet and the document corpus, in that order. Those are the two projects with high value, maximum feasibility and minimal regulatory risk. They produce a visible result within weeks, they need no heavy system access, and they surface the processing rules you will need later.
KYC, fraud detection and the voice agent come next, once teams know how to measure. Credit scoring and automated advice to the end client wait for the Annex III frame, in December 2027. That is not caution, it is the order that delivers the most value after twelve months.
One last rule, valid everywhere in finance. Measure the starting point before you begin. It is the only reason two thirds of the firms that acknowledge gains are unable to quantify them.
