How-to · everyday AI
On November 13, we argued that annual awareness training protects no one: what protects is sound processes and short, frequent exercises. AI is the best place to apply it. Your AI acceptable use policy may run twenty pages. Nobody will reread it before pasting a contract into an assistant. What your team remembers is what it can apply on Monday morning.
Here is a fifteen-minute session to run in a team meeting, with no slides and no vendor. It does not replace the policy or the technical controls. It gives everyone the four reflexes that matter.
Why fifteen minutes is enough to start
Three facts frame the exercise.
- Europe already requires it. Article 4 of the EU AI Act, applicable since February 2, 2025, asks providers and deployers of AI systems to take measures for the AI literacy of their staff. Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on July 24, 2026 and in force since July 27, rewrote the wording: the duty is now to take measures to support the development of that literacy, without having to guarantee a given level for each person. The obligation remains. Status: confirmed, text published in the Official Journal. It concerns you if you deploy AI in your European operations or sell a system there.
- Canada has no AI law in force, but it has reference points. The federal Guide on the use of generative AI, written for federal institutions, recommends not entering personal or sensitive information into public tools and checking what they produce. The Canadian Centre for Cyber Security, in its generative AI fact sheet (ITSAP.00.041), points to the risks of data leakage and inaccurate information. Status: official publications. They do not bind a private company, but they make a good template.
- The risk is mundane, not spectacular. In spring 2023, Samsung engineers pasted source code and meeting notes into ChatGPT, according to the Korean press and Bloomberg (reported). Nobody was hacked. Data simply left, for convenience.
What it changes: the session is not meant to make experts. It is meant to make sure nobody, facing an AI tool, wonders what they are allowed to do.
The agenda, minute by minute
Minutes 0 to 2: why we are talking about it. One concrete example, not a statistic. Samsung's, or better, a case from your sector. One framing sentence: "We want you to use AI. Here is how to do it without putting the company at risk."
Minutes 2 to 5: the approved tools. Show the list on screen: which tools, with which account (the company account, never a personal one), for which uses. Then the route to request another one: who to write to, and how many days until the answer. If that route does not exist, the session will reveal it: that is how shadow AI is born, from an unanswered request.
Minutes 5 to 9: what can go into a prompt. The heart of the session. Three lights, easy to remember.
- Green, anywhere: public information, your own drafts with no client name, general questions.
- Amber, only in an approved tool: non-sensitive internal documents, procedures, code with no secrets.
- Red, never without written approval: personal information about clients or employees (privacy law applies to what you paste), a client's confidential information, passwords, keys and tokens, unpublished financial results, contracts under negotiation.
Minutes 9 to 12: check what comes out. AI writes with confidence, even when it is wrong. Three rules: whoever uses the text answers for it, as if they had written it; every figure, name, date, quote or legal reference is checked at the source; nothing is published, sent or run (an email to a client, code in production) without being reread.
Minutes 12 to 14: report. What: red data sent by mistake, a strange answer, an unknown AI tool installed, a call or video message impersonating an executive. To whom: one name or one address, posted. And the promise that makes it work: nobody is blamed for reporting. A mistake reported within ten minutes gets fixed; a hidden one becomes an incident.
Minutes 14 to 15: the exercise. See below.
The three-light exercise
Put up three prompts, one at a time, and ask the room: green, amber or red?
- "Summarize this competitor's public annual report in five points." Green.
- "Improve the tone of this email", followed by an email that contains a client's name, address and medical file. Red: the content decides, not the task.
- "Find the error in this internal month-end closing procedure", in the approved tool. Amber, and acceptable.
The second prompt always starts a debate. That is the point: the rule you want people to keep is that the colour depends on what you paste, not on what you ask.
The three traps
Running the session once. Fifteen minutes once is the annual module, only shorter. Come back for five minutes a month with a new real case, taken from reports or the news.
An empty approved list. If the only answer is "no", the team will use their phones. Approve at least one company tool before the session.
Recording nothing. Note the date, who attended and the content. That is the evidence a client, an insurer or, in Europe, an authority will ask for, and it fits on one line.
What you should have a week from now
The session held in at least one team, a one-page list of approved tools, the three lights posted where people work, a reporting address everyone knows and one line in the training record. This is not an AI governance program. It is the first thing that program should produce.
And you?
How many pages is your AI policy, and how many people have read it? Which of the three prompts would start a debate in your team? Tell us in the comments.
On the platform side, Forta Compliance links your training evidence and AI policy to your clients' requirements and your frameworks.
Sources: Regulation (EU) 2024/1689 on artificial intelligence, Article 4 · Regulation (EU) 2026/1744 (Digital Omnibus on AI) · Government of Canada, Guide on the use of generative artificial intelligence · Canadian Centre for Cyber Security, Generative artificial intelligence (ITSAP.00.041)