Regulated professions cannot paste a client file into a consumer AI tool. But they should not give up on AI either: their competitors are not giving up. The question of AI and professional secrecy is not a question of prohibition; it is a question of architecture and written rules. This article sets out the concrete rules we apply with the law firms, notary offices and accounting practices we equip: controlled hosting, compartmentalisation, pseudonymisation and a usage charter.
The real risk behind consumer tools
The risk is not AI: it is copy-paste. A rushed associate pastes three pages of a client file into a consumer tool to "save time". At that instant, data covered by professional secrecy has left the firm, towards a service whose terms of use, server location and retention policy nobody has read.
Depending on the tool and its settings, that content may be retained, reviewed for moderation, sometimes used to improve models. The problem is not that a competitor reads it tomorrow; it is that the firm has lost control, and can no longer answer a client honestly when asked: "where is my data?"
Banning solves nothing: usage goes underground, and therefore invisible. We have seen firms where AI was "forbidden" and half the staff used it on their personal phones. The only serious answer is to provide a framework and tools that make the right move easier than the wrong one.
The one-question test
if a client asks you tomorrow "where did my file go, who read it, and can you prove it?", can you answer for every tool in the firm? If not, this is the priority.
The framework: GDPR, the AI Act and professional duty
Three layers stack up, and they must be distinguished to avoid mixing everything.
| Layer | What it concretely requires | Who supervises |
|---|---|---|
| GDPR | Minimisation, legal basis, a processing agreement (article 28) with every provider that touches the data, a register of processing, an impact assessment when the risk justifies it | Data protection authority |
| European AI Act | Staff AI literacy since February 2025 (eased by the July 2026 Omnibus), transparency on generated content since August 2026, reinforced obligations for high-risk uses from December 2027 | National authorities, European Commission |
| Professional duty | Secrecy applies whatever the channel: email, phone, AI assistant. No "convenient tool" exception | Bar, chamber, professional body |
GDPR first: process only what is necessary, have a clear legal basis, and sign an article 28 processing agreement with every provider that touches the data. A tool whose publisher refuses that agreement is not a tool for a firm.
The European AI Act next. For a firm, the most immediate point is staff AI literacy: even eased by the July 2026 Omnibus, training on the limits of the tools remains the first protection of secrecy. Then comes the ability to list your use cases and qualify their risk.
Professional duty last, and it is the strictest layer: professional secrecy has no "convenient tool" exception. Which is why the question to ask of every tool is not "does it perform?" but "can I demonstrate that secrecy survives it?"
The three layers converge on the same practical requirements: controlled hosting, compartmentalised access, traceability, human review. One architecture satisfies all three.
An architecture that protects client files
What we install in firms rests on five principles.
- Hosting in the European Union, with providers under processing agreements, and models accessed through professional interfaces whose terms exclude training on your data.
- Compartmentalisation by file. An AI agent working on file A never sees file B, and access follows the firm's existing permissions: a staff member sees in the AI only what they already see in the case management system.
- Pseudonymisation before processing whenever the case allows. The AI reasons about "Mr X, employed since 2019", not about civil identities. Identities are reinserted in the final document, on the firm's side.
- Logging. Every processing step is dated, attributed, findable. That is what lets you answer a client or an audit without improvising.
- No client data ever trains a model. This is verified in the contract, not in the brochure.
And one rule across everything: human review before anything leaves the firm. AI prepares, a human signs. Our AI agents are built with this mandatory validation point on every sensitive decision.
What this is not
installing a model on a server inside the firm. In-house hosting is expensive, poorly maintained and no safer than a European host under contract. Security comes from access rules and traceability, not from physical proximity.
Permitted and excluded use cases
| Pass, inside the framework above | Do not pass |
|---|---|
| Summarising large sets of documents to prepare a meeting | Pasting an identifiable client file into a consumer tool, whatever the goal |
| Drafting recurring letters and deeds from the firm's own templates | Letting AI issue an opinion that reaches the client without review |
| Research on public sources (statutes, case law, doctrine) | Wiring an agent to the firm's inbox with no compartmentalisation |
| Sorting and prioritising incoming requests | Using a free tool whose business model is the exploitation of your data |
| Transcribing and summarising internal meetings | Recording a client meeting without information or consent |
| Consistency checks on a deed (dates, amounts, names) before signature | Entrusting AI with a procedural deadline calculation without verification |
In between lies a grey zone every firm must settle in writing: that is the role of the usage charter, a one-page document everyone signs, which replaces case-by-case debates.
The one-page usage charter
A good charter fits on one page and answers six questions.
- Which tools are allowed, and which are forbidden, by name.
- Which data may enter them: nothing identifiable outside compartmentalised tools, never an identity document, never health data outside a framework.
- Who validates what goes out: the rule "AI prepares, a human signs", written down.
- How the client is informed, when necessary, and with which wording.
- What to do in case of an error or a leak: whom to notify, within what delay, which trace to keep.
- How the charter evolves: a review every quarter, because tools change fast.
We provide this template in our one-day AI training, adapted to the firm's practice, with exercises on your own types of files.
The first steps
Start with an honest inventory: who already uses what, officially and unofficially. That diagnostic takes a few days and always surprises. Then set the usage charter, and equip one pilot use case, compartmentalised and measured: often file summarisation or the preparation of one recurring type of deed.
Train teams on their real files rather than generic examples, then measure: hours saved per week, shortened turnaround, rework rate. After a quarter, the numbers decide what comes next, not convictions.
| Step | Indicative duration | Deliverable |
|---|---|---|
| Inventory of uses | A few days | List of tools actually used, risks ranked |
| Usage charter | One week | One-page document signed by the team |
| First compartmentalised agent | Two to three weeks | One process in production, action log, human validation |
| Training on real files | One day | Practice-specific prompt library, written framework |
| Measurement at one quarter | Ongoing | Hours saved, turnaround, rework |
That is exactly the path we follow with the firms we equip, described on our page for regulated professions.
Questions to ask any AI vendor
Before a firm adopts a tool, the vendor should answer these questions in writing. The answers belong in the contract, not in a sales deck.
- Where is the data processed and stored, and can you name the hosting providers?
- Is our data used to train or improve your models, and is that exclusion written in the contract?
- How long are prompts, documents and outputs retained, and can retention be set to zero?
- Will you sign a processing agreement under article 28 of GDPR, and who are your sub-processors?
- Can access be restricted per file or per user, following our existing permissions?
- Is every processing step logged, and can we export those logs?
- Who at your company can read our content, under which conditions, and is that logged too?
- What happens to our data when we terminate, and within what delay is it deleted?
- Have you documented how your product fits the European AI Act, including transparency obligations?
- Can you provide references from firms subject to professional secrecy in our jurisdiction?
A vendor who answers all ten clearly is a vendor a firm can work with. A vendor who answers "it depends on the plan" to question two is not.
Takeaway
do not ban, frame. European hosting under contract, compartmentalisation by file, pseudonymisation, logging, human review, a one-page charter. Then one measured use case before widening.
Frequently asked questions
Sources and references
All our guides on this topic: AI agents
TopicsProfessional secrecyLaw firmsGDPRAI agentsProfessional duty
The author
Author of the Stratedge Consulting blog. Stratedge Consulting designs and builds custom websites, software and dashboards in Paris. More than 250 clients since 2022: law firms and notaries, SMEs, startups, groups. Every article starts from the projects we deliver for our clients.
LinkedInIn practice
A 30-minute video call with a Stratedge expert to look at your situation, then, if the need is confirmed, an Express Diagnostic at €500 excl. VAT, credited against the next step.
Book a video call