AI at work: effective and safe
Complete plan
Sixteen habits to save time with AI every day, without losing control or taking risks: the twelve habits of the course "Save time with AI at work", with the four safety reflexes at the right moments: check before you sign, keep confidential data in, tidy your shares, call back on a known number.
- 38 min
- 16 modules
- 48-question check
- English · Français
- Attestation
You will be able to
- Sort a task before handing it to AI: hand it over (first drafts, summaries, rewording, ideas, working translations), do it together (analysis, plans, comparisons) or keep it
- Know that generative AI can invent figures, sources and quotes
- Write a request in four parts: context (who it is for, why), task, expected format, limits (tone, sources, what not to do)
- Ask the AI to restate your request and ask its questions before it answers
- Spot the generic sentences in an AI draft with the name-swap test, and cut them
- Recognise shadow AI: unapproved AI tools, free or paid
- Have a reply drafted from a one-line intention (the decision and the tone), then read it before sending it yourself
- Question a long document from the decision you have to make (for example: commitments, deadlines, penalties) rather than asking for a summary
- Prepare a meeting with your approved tool: agenda, likely objections, questions to ask
- Ask for the method or the spreadsheet formula rather than the result, or have an approved tool that runs code do the calculation
- Understand that an AI assistant follows existing access rights, including overly broad sharing
- Tell a one-off request, a saved assistant and an agent that acts with tools apart, and pick the lightest level that does the job
- Compare two models or two modes of your approved tools on a real task, with three test cases: a simple one, a tricky one, a long document
- Write plain-text notes, one per topic, with a clear title, the date, an owner, the sources and links to related notes
- Know that a voice, an image or an email can be generated by AI
- Describe an automation on one page: the trigger, then steps that each have one input and one output
Programme
16 modules · 38 min-
Where AI saves you time (and where it costs you time) 2 min
- Sort a task before handing it to AI: hand it over (first drafts, summaries, rewording, ideas, working translations), do it together (analysis, plans, comparisons) or keep it
- Keep decisions about people, official figures and anything you cannot check for a person
- Measure the gain on one real task, with and without AI, instead of trusting the feeling
-
Plausible isn't true 2 min
- Know that generative AI can invent figures, sources and quotes
- Check an answer before using it: open the source, cross-check the figure, reread
- Know that the author and their company remain responsible for what they publish
-
A clear request: context, task, format, limits 2 min
- Write a request in four parts: context (who it is for, why), task, expected format, limits (tone, sources, what not to do)
- Attach one example of what you expect rather than a long explanation
- Fix a disappointing answer by adding the missing part, instead of repeating the same request
-
Let the AI ask you the questions 2 min
- Ask the AI to restate your request and ask its questions before it answers
- Ask for three options from different angles, then build on the best one
- Refine with short, precise instructions ("shorter", "for a CFO", "keep the second paragraph") rather than starting over
-
Never pass on a raw AI draft 3 min
- Spot the generic sentences in an AI draft with the name-swap test, and cut them
- Rework a draft before passing it on: half as long, the conclusion or the request in the first line, adapted to the reader
- Ask the approved tool for the three weak points of its own text for a given reader, then re-read it yourself
-
What never goes into a public AI 2 min
- Recognise shadow AI: unapproved AI tools, free or paid
- Know what never goes into a public AI tool
- Use the approved tool, and report data already pasted to IT straight away
-
Your emails: reply fast, in the right tone 2 min
- Have a reply drafted from a one-line intention (the decision and the tone), then read it before sending it yourself
- Turn a long email thread into three lines: where things stand, what is still open, who is waiting for what from me
- Have your own email read through the recipient's eyes before sending it, and use only the approved tool for client or supplier emails
-
A long document: the right questions to ask it 3 min
- Question a long document from the decision you have to make (for example: commitments, deadlines, penalties) rather than asking for a summary
- Require answers drawn only from the document, with the page or section, and check each point at its page
- Ask what is missing or unusual, and what changed between two versions
-
Meetings: leave with decisions 3 min
- Prepare a meeting with your approved tool: agenda, likely objections, questions to ask
- Turn notes or a transcript into decisions, actions (who, what, by when), open points and a follow-up e-mail, checked before sending
- Tell participants before you record or transcribe, following your company's rules, and never record or transcribe an HR interview or a meeting about health unless your company has explicitly provided for it
-
Numbers, without the trap 2 min
- Ask for the method or the spreadsheet formula rather than the result, or have an approved tool that runs code do the calculation
- Check one total by hand and ask for a list of the problem rows (empty cells, duplicates, amounts stored as text), to check in the file
- Replace customer or employee names with codes before you paste any data
-
Copilot sees everything you can open 2 min
- Understand that an AI assistant follows existing access rights, including overly broad sharing
- Check who has access to your files and remove unnecessary access
- Share with specific people, remove access once it's no longer needed, and report a document visible by mistake
-
Request, assistant, agent: pick the right level 3 min
- Tell a one-off request, a saved assistant and an agent that acts with tools apart, and pick the lightest level that does the job
- Build a saved assistant: reusable instructions, reference documents from approved sources, shared with the team, and three test cases re-run whenever the tool or its documents change
- Require a human sign-off point before an agent sends, pays, deletes or decides for you
-
Several models: test before you trust 3 min
- Compare two models or two modes of your approved tools on a real task, with three test cases: a simple one, a tricky one, a long document
- Choose on five criteria: quality on your task, speed, how long a document it handles, limits, and where the data goes
- Keep a small test file and rerun it when a new model arrives, never comparing in an unapproved tool
-
Your knowledge base, readable by AI 3 min
- Write plain-text notes, one per topic, with a clear title, the date, an owner, the sources and links to related notes
- Keep the team's base in the company's approved space and remove outdated notes
- Ask AI to answer from the base, citing the notes it used, then check at the source
-
Deepfakes: the call-back still works 2 min
- Know that a voice, an image or an email can be generated by AI
- Stop relying on the old clues: typos, robotic voice
- Apply the call-back on a known number, the verification question and the second approval, including for a change of bank details
-
From repetitive task to automation: the mission sheet 3 min
- Describe an automation on one page: the trigger, then steps that each have one input and one output
- Say at each step whether AI reads, proposes or decides, and spot the steps where a rule is enough, often half of them
- Place a person's approval before anything leaves the company, a payment, a deletion or a decision about a person, plan for failure, and hand the sheet to IT
-
Knowledge check
48 questions, pass mark 80 %. Explanations after each answer; you can try again.
-
Attestation
One per person, with the date, the score and the frameworks covered — anyone can check it online with its code.
Version 1 · updated October 17, 2026