Skip to content

Using AI at work: 5 simple rules for your teams

Client data, context, checking, decisions, disclosure: the five rules of an AI use policy, in a 1 min 41 s video, then in detail with the EU rules behind them.

5 min read

Video generated with EducPilot for Brivane Conseil, a fictional 40-person consulting firm. In French, 16:9, 1 min 41 s, two review rounds.

Your teams already use AI assistants, sometimes on personal accounts. A five-rule policy explained in under two minutes is a concrete first step. Here are the rules, why they hold, and what they leave out. The video is in French; its narration is translated below.

Why now

In the EU, Article 4 of the AI Act has applied since 2 February 2025: organisations that use AI systems must take measures so their staff have a sufficient level of AI literacy, given their knowledge, training and context of use. Supervision and enforcement rules apply from 3 August 2026.

What the European Commission's Q&A spells out:

  • an employee drafting with a ChatGPT-type assistant is covered, and should know the specific risks, such as made-up answers;
  • no specific training or certificate is mandatory; internal records of what you did are enough;
  • the effort depends on the systems used and their risks.

Outside the EU, the same five rules are simply good practice.

The 5 rules

1. Never put client data in a personal account

On consumer plans, your chats may be used to improve the models if you allow it in your settings. A client contract pasted into a personal account has left the company.

What to do: use the workspace your company has approved (a business plan with checked terms), and only share what the task needs.

2. Give context

Too vagueWith context
"Summarise this.""Summarise this report in five points for our CFO."

A good prompt answers three questions:

  • Who for? The reader, and what they already know.
  • What for? The decision they need to make.
  • In what form? A page, a table, an email.

3. Check everything

An AI assistant can produce wrong or misleading answers, even when they sound sure. Anthropic says so in its own help pages: Claude should not be relied on as a single source of truth. Think of it as a brilliant intern: fast, but to be checked. Figures, names, quotes: always go back to the source.

4. AI suggests, you decide

The answer is a draft. You sign what goes to the client, and you answer for it.

5. Say when AI helped

Clients should know how their deliverable was prepared. A plain note is enough: "first draft written with an AI assistant, reviewed and approved by the team".

What the video does not cover

The video is about daily use. It does not replace:

  • choosing and configuring the tool: France's cybersecurity agency ANSSI published security recommendations for generative AI systems in April 2024;
  • a data protection review of the processing involved;
  • job-specific training: drafting an audit with AI is not learned like drafting an email.

How this video was made with EducPilot

  • Sources uploaded: the firm's five rules (a fictional document) and two public Claude help pages, on data use and on possible errors.
  • A story with a character: Inès, a consultant, has to send a memo to her client before 6 pm and is about to paste the client's report into her personal Claude account. The video follows her from one rule to the next, until the memo goes out.
  • Transitions that ask the next question: "And you, would you paste it into that account?", "But is its memo accurate?", "Once the sources are checked, who answers for the memo sent to the client?", "So, can Inès send her memo?".
  • The plan: the situation ("Thursday at 4 pm"), one scene per rule, Inès sending her memo after five checks, the firm's AI lead when in doubt, the closing reflex: "Before anything goes out: Brivane workspace, framed request, sources checked, deliverable signed, client informed."
  • Devices: the personal account and the Brivane workspace side by side, the details of a framed request as steps (the reader, the decision, the format), "Claude suggests" facing "you decide and sign".
  • Sources in the film: each claim about data and errors links to the help page that supports it.
  • Review: two rounds of adversarial critique by a model from another family. A note applied at 0:15: "Rule 1, no hedging: 'Never enter client data in a personal Claude account.'"

Claude is used here as an example tool. EducPilot has no partnership with Anthropic.

To do the same with your own policy, see How to turn a PDF into a training video.

The narration, translated

Thursday at 4 pm, Inès, a consultant, has to send a memo to her client before 6 pm. She opens her personal Claude account and is about to paste the client's report into it. And you, would you paste it into that account?

Never enter client data in a personal Claude account. On a personal account, chats may be used to improve the models. Use only the Brivane workspace, approved by the firm.

Inès switches to the Brivane workspace. But a vague request gets a vague answer. She states who will read the memo and what that person already knows. She gives the decision to be made and asks for a one-page memo.

Claude answers Inès with confidence. But is its memo accurate? Always check figures, names and quotes against the source.

Once the sources are checked, who answers for the memo sent to the client? Claude suggests, but you decide, sign and remain responsible for the deliverable.

Even if Inès signs the memo, should she tell the client that AI helped her? Yes: tell the client when AI helped prepare a deliverable.

So, can Inès send her memo? Yes, after five checks: Brivane workspace, framed request, sources checked, memo signed and client informed.

And if, despite these five checks, a doubt remains? Contact the firm's AI lead.

Before anything goes out: Brivane workspace, framed request, sources checked, deliverable signed, client informed.

Sources

Full transcript

Original narration, in French.

Jeudi à 16 heures, la consultante Inès doit envoyer une note à son client avant 18 heures. Elle ouvre son compte Claude personnel et s’apprête à y coller le rapport du client. Et vous, le colleriez-vous dans ce compte ?

Ne saisissez jamais de données client dans un compte Claude personnel. Sur un compte personnel, les échanges peuvent servir à améliorer les modèles. Utilisez uniquement l’espace Brivane, validé par le cabinet.

Inès passe dans l’espace Brivane. Mais une demande vague donne une réponse vague. Elle précise qui lira la note et ce que cette personne sait déjà. Elle indique la décision à prendre et demande une note d’une page.

Claude répond à Inès avec assurance. Mais sa note est-elle exacte ? Vérifiez toujours les chiffres, les noms et les citations à la source.

Une fois les sources vérifiées, qui répond de la note envoyée au client ? Claude propose, mais vous décidez, signez et restez responsable du livrable.

Même si Inès signe la note, doit-elle signaler au client que l’IA l’a aidée ? Oui : informez le client lorsque l’IA a aidé à préparer un livrable.

Alors, Inès peut-elle envoyer sa note ? Oui, après cinq contrôles : espace Brivane, demande cadrée, sources vérifiées, note signée et client informé.

Et si, malgré ces cinq contrôles, un doute subsiste ? Contactez le référent IA du cabinet. Avant tout envoi : espace Brivane, demande cadrée, sources vérifiées, livrable signé, client informé.

Try it with your own documents

Upload a PDF or a procedure, approve the plan, correct the video by annotating it.

Read next