AI in finance: where firms stand
Of 100 financial market participants surveyed by the AMF, 54% have AI use cases in production, mostly among the largest. 83% of the uses identified are internal: productivity, document extraction and summarisation. The risks cited include excessive trust in the tool without sufficient human oversight (AMF, February 2026).
In an organisation of 5 to 300 people, output is almost entirely document-based, and value is lost in four places:
- Assessment. Its fixed cost determines the size of the deals the firm agrees to look at.
- Evidence. Regulators and investors ask for dated records, several years after the fact.
- Review. Every document that commits the firm is reviewed line by line by compliance.
- Memory. Past decisions and the reasons for them are held in minutes and in the heads of two or three partners.
Demonstration scenario
Review the deals too small to be assessed
The company described is fictional. This scenario shows how we approach a problem; it does not report a result achieved at a client.
The situation
A private equity management firm: eighteen people, taking minority stakes in regional SMEs. Assessing a deal ties up two people for three weeks: reading the data room, interviews, AML/CFT file, memo for the committee.
This cost is the same for an €800,000 ticket as for a €5 million one. The firm has set a threshold at €3 million. Deals below it are declined, even though they fit its investment thesis.
What the company already has
The data rooms and memos from past transactions, including those that were turned down. The firm’s due diligence grid. The investment committee minutes and the reasons recorded in them.
What the system produces
The system reads the entire data room. It extracts the grid’s points of attention: off-balance-sheet commitments, change-of-control clauses, customer dependency, disputes, leases. It prepares the AML/CFT file and drafts a first memo in the firm’s template. The team carries out its assessment from this memo, and its interviews focus on the points the system has identified.
What is verified
A second system, separate from the one that drafted the memo, checks it before it reaches the team:
- each statement in the memo refers to a document in the data room;
- grid topics on which the data room contains nothing are listed as points of attention;
- each item in the AML/CFT file carries its source and the date it was obtained.
The record produced on an €800,000 deal is the same as on a €5 million deal. The principle is described in detail on the home page.
What the system does not do
It does not decide on the investment and does not give an opinion to the committee. It does not value the target company.
How the result is measured
The amounts invested in deals that the firm used to turn down because of their size.
AML/CFT files, compliance of marketing materials and complaints: three other uses of AI in finance
| Onboarding (KYC, AML/CFT) | Compliance of marketing materials | Client complaints | |
|---|---|---|---|
| Today | Collecting documents, tracing the ownership chain up to the beneficial owner and building the file takes days per client. Periodic reviews fall behind. | Brochures, reports and investor presentations are reviewed line by line by compliance. The approval time slows down commercial output. | Complaints arrive through every channel, sometimes in an email sent to the fund manager. Classifying them and meeting response deadlines depends on one person and a spreadsheet. |
| The system produces | The file in the expected format, the ownership chain rebuilt from public registers, and what is missing or has changed since the last review. | A check of each document against the internal reference framework: mandatory statements, warnings, consistency of figures with regulatory documents. | The complaint classified and linked to the client’s file, a reply based on the supporting documents and on comparable cases already handled, and deadline tracking. |
| What is verified | Each item carries its source, the date it was obtained and its renewal deadline. | The record of the check is dated, names the reviewer and is kept with the document. Points requiring a decision go up to the compliance officer. | No reply is sent without approval. Mandatory statements and consistency with contractual commitments are checked separately. |
| Data needed | Existing files, the internal risk classification procedure. | The internal compliance reference framework, comments made on past documents. | The history of complaints and their replies, contractual documents. |
What Marsanne does not do in finance
- Investment decisions or client advice. The system prepares and checks; an authorised person decides.
- Invoicing, payment reminders and expense reports. Tools on the market are sufficient.
- Software that already exists. When a management or consolidation tool covers the need, we say so and tell you which one.
To gauge the expected return on a project before launching it, see How to measure the return on investment of AI.
Frequently asked questions
Do we need to connect our management tools or our accounting software?
Not to get started. The system works on existing documents and exports: data rooms, client files, general ledgers. A few people use it for real, and the question of connecting it comes up once the value has been shown.
Can the result be presented to a regulator during an inspection?
Yes. Every item produced refers back to its source document, and every check is dated and kept with the document. The decision remains with an authorised person.
How long before a first result?
Development generally takes a few weeks. Users benefit as soon as they start using it.
What happens to our data?
Once the solution is in service, Marsanne no longer has access to your data. You remain its sole owner.
Will our teams need training?
No. The system is designed around the actual work of the team that uses it, and can be used without training.
Articles — Finance
- Bank covenants: seeing the threshold coming before the close 6 October 2026
- Marketing compliance: checking every document before it is circulated 28 September 2026
In the press
The latest articles from the studio's AI watch on AI in this industry, summarised in English with their source.
- Leaving the "stone age" of financial, geopolitical and strategic analysis: fed with data, super-forecasting AIs are set to transform how decisions are made
Forecasting AIs are rivalling human experts in financial, geopolitical and strategic analysis, according to the article.
BFM Business · 12 Oct 2026 · France - MCP: how AI is entering your business software through the back door
The MCP protocol standardises the connection between AI and business software, opening up opportunities and risks for SMEs and software vendors.
Journal du Net · 12 Oct 2026 · France - AI: finance departments facing spending that still escapes their control
A Najar study shows that 93% of AI use in companies is not covered by any contract, driving up hidden costs.
DAF Magazine · 12 Oct 2026 - He is to AI what the foreman was to the factory, and he earns €13,000 a month: the new job whose demand has jumped 1,700% on Wall Street
Wall Street banks are hiring specialists in "AI agent orchestration", paid up to €13,000 a month.
BFM Business · 10 Oct 2026 · United States - Insurers face AI's big dilemma: who pays when it gets it wrong?
The insurance industry hopes for higher revenues thanks to AI but does not know who should pay when damage occurs.
Business Insider España · 10 Oct 2026 · Spain
Find out where AI can help in your organisation
The executive diagnostic takes ten minutes. It compares your answers with the situations we encounter in finance and gives you practical leads.
Start the diagnostic