AI in a French accounting firm: what to automate, in what order, under professional secrecy

This guide is for partners and managers of French accounting firms (cabinets d’expertise comptable) who wonder what to do with AI, in practice, in a profession bound by professional secrecy. It does not compare software. It answers three questions: which processes to hand to an AI, with which tools, and under what conditions to stay within the rules of the profession.

It is written by a developer, not by a chartered accountant. The rules quoted are quoted at source, with their reference. For a question of professional conduct, your regional council of the Ordre remains the right contact.

What changed on 1 September 2026

France’s electronic invoicing reform has come into force, and it reaches the core of a firm’s production work.

Invoices go through an approved platform and arrive as structured data. The consequence for a firm is mechanical: invoice entry and recognition lose value as your clients switch over.

Two things do not disappear. Everything that is not a business-to-business invoice still has to be handled: bank statements, expense claims, receipts, contracts, leases, payroll documents, letters from the tax authority. And the work around the data remains: the missing document, the reminder, the entry that will not reconcile, the client’s question. The hours move; they do not vanish.

Three levels of AI in a firm

The word “AI” covers three things that are not decided in the same way.

The first level is the one you already pay for. Before any project, check what your vendors do today: rebuilding it would be wasted money. The third level starts where they stop, at the processes that are specific to your firm and that cross from one tool to another.

On the second level, if your firm runs on Microsoft 365, the question “is Copilot enough?” is settled one process at a time: see Copilot or custom AI.

Where the hours go: candidate processes

A good candidate has three features: it is repetitive, it can be measured, and a mistake in it is caught before it leaves the firm.

The second row is the most mature technically. I put a document AI pipeline into production on that principle: the AI extracts, deterministic rules check, a human validates what is doubtful, and corrections improve later extractions. The pattern is the same for a firm, and it rests on a simple rule: the AI proposes, the deterministic layer decides.

Searching the firm’s own material needs one more precaution: a staff member must only find the files they are authorised on. That filtering happens in the query, not in the model, as detailed in the RAG on SQL Server 2025 guide.

I give no percentage gain. It depends on your client base, your tools and your organisation, and an average figure would tell you nothing. It is measured on one of your processes, before deciding.

Professional secrecy sets the architecture

This is what separates a firm from an ordinary SME. Article 21 of the Ordinance of 19 September 1945 binds chartered accountants, their employees and trainees to professional secrecy, under the conditions and penalties of Article 226-13 of the Criminal Code: one year’s imprisonment and a €15,000 fine. The duty lies on individuals, not only on the firm.

The national council of the Ordre has drawn a practical instruction from it in its notice on generative AI: do not load personal, sensitive or confidential data, no client emails, accounting entry files or payroll declarations that are not anonymised, into sites the firm does not control. The CNIL, the French data protection authority, points the same way in its questions and answers of 18 July 2024: for personal data or sensitive documents, an on-premises deployment is generally the safest, and the use of consumer services must exclude confidential information.

These texts name no supplier and set no hosting country. They set one condition: the firm must control where its clients’ data goes. Here is how I translate it.

For the third row my rule is strict: hosting in the European Union, no data reused to train a model, and permissions that follow the firm’s own. It is not the only possible reading of the texts. It is the one that stands without argument in front of a client, an insurer or the Ordre.

One point to check if you use Copilot: Microsoft states that the Anthropic models offered in Copilot are excluded from the EU Data Boundary. The administrator can enable them or not. For a firm, that setting is checked before the tool is opened to the teams.

Lastly, the EU AI Act asks organisations that use AI systems to ensure the AI literacy of their staff. That obligation has applied since 2 February 2025. Training staff is therefore not a comfort option: it is also the first precaution the Ordre lists.

What the AI does not do

It does not sign. The model charter published by the Ordre starts there: the use of generative AI must not replace human decision-making or neglect human expertise and reasoning. It adds that answers must be checked and validated before being shared, internally and with clients.

In practice this gives a three-link chain: the AI prepares, the staff member checks, the chartered accountant signs. A project that removes the second link to go faster is not a productivity gain. It is a transfer of risk to the person who signs.

It also sets the order of work. Start with processes where a mistake is visible and corrected in-house (a badly worded reminder, a misfiled document), not with those where it leaves the firm (a tax return, a payslip).

Seven steps to start

1. Take stock of the uses already there. Staff are already using an assistant, sometimes on a personal account. Better to know than to ban it unseen.
2. Adopt a charter. The Ordre provides a model in its “Parlons data !” file, in the private section of its site. It can be folded into your IT charter.
3. Train the teams. On the limits of the tools, on checking answers, and on the data rule: nothing confidential in a site you do not control.
4. Switch on what you already pay for. The AI features of your production software and your approved platform come before any development.
5. Measure one process. Pick one, document collection for example, and count over a month: the number of reminders, the time between the request and the document, the time spent. Without that figure, no project can be judged.
6. Set the architecture before the tool. Where the model runs, who accesses which file, what is logged, what happens when the system is not sure of itself.
7. Start with a process where a mistake can be caught. Then measure again, with the same counters as in step 5.

Going further

Steps 5 and 6 are the ones that call for an outside view. That is what the 30-day AI diagnostic is for: the firm’s processes mapped, the return costed for each priority case, and for each one the architecture that respects professional secrecy, including when the right answer is a feature of your current tools. What follows, when it is justified, is AI agents and business applications, built and shipped inside your environment.