Your team doesn’t need an AI genius. It needs a ladder.

In most teams, one or two people are good with AI. Every question ends up in their inbox. Everyone else uses the tool a little and hopes nobody notices.

That is not a skills shortage. It is a transfer problem. The know-how sits with one or two people, nobody writes it down, and it leaves when they do.

The people who are slower to pick it up are not lazy. Nobody has shown them what good use looks like, step by step. They had a two-hour training session and an instruction to "use AI more". When I read 146 job ads for AI adoption roles, 63 asked the hire to train users. Only 17 said how the result would be measured.

So here is a ladder. 20 small steps in five stages, and a side path for what saves teams real hours, dashboards and automations. Every step has a task and a test that tells you when you have it. Plan on ten weeks at twenty minutes a day.

First, what good use looks like

Using AI well does not take secret formulas. It takes two plain things. Give the tool what it needs, and say exactly what you want back.

Everything below is a variation on those two. A vague instruction gets you text that sounds right and is wrong. That is what puts people off. One bad output, and they decide AI is overhyped. A specific instruction gets you something you can use.

The ladder

Stage 1. Get specific (steps 1 to 4)

1. Stop starting from zero. For any task you do often, keep five things to hand: a good example, the rules, a sample of your writing, a bad example, and a note on what a good answer looks like. Paste them in before you ask. You have it when you no longer explain the task and the first draft is already usable.
2. Write a brief, not a wish. Before the request, four lines. Who reads this, how long it should be, what must be in it, and what is banned. Banned words help: "no hedging, no long intros". You have it when a colleague gets the same shape of answer from your instruction.
3. Say what you don’t want. Most people only describe the target. Describe what to avoid too: a tone, a structure, the jargon. You have it when you can name three problems the "avoid" list fixed.
4. Ask for three versions. Safe, middle, bold. Pick one and write one line on why. You have it when your pick still looks right the next morning.

Stage 2. Build once, reuse (steps 5 to 8)

5. Set up a dedicated workspace. A Project in Claude, a notebook in Microsoft 365 Copilot or a custom GPT in ChatGPT keeps your rules and reference files. Write the instructions as if briefing a new starter. You have it when you open it three days running and never re-brief.
6. Fix your voice. Add three things you actually wrote, with a note on sentence length, formality and the words you never use. You have it when a reader can’t tell whether you or the tool wrote it.
7. Edit instead of regenerating. When a draft is 80% right, fix the other 20%. Regenerating throws away the good part and makes the text drift. You have it when a document reaches final in six small edits, not six rewrites.
8. Keep a version log. One line per version: what changed and why. Two minutes that save an hour of wondering which version you sent. You have it when the log explains every difference between the first and last draft.

Stage 3. Get the work checked (steps 9 to 12)

9. Plan before you produce. For anything longer than a paragraph, ask for a plan first, then for the five ways it could fail. You have it when the plan names a risk you hadn’t thought of.
10. Ask for the case against. "Give me the strongest argument against this, then the sentence you would cut first." You have it when you accept two of the notes and change the draft.
11. Break big jobs into small ones. Five or six steps, each producing something the next one needs. You have it when no step ends with "and then we’ll see".
12. Check the sources, not just the answer. Ask for the strongest credible source that disagrees with you, then open the links. You have it when every claim leads to a page you read yourself.

Stage 4. Go beyond the chat window (steps 13 to 16)

13. Search like a librarian. Short, specific, dated searches. Four sharp searches beat fifteen loose ones. You have it when each conclusion leads to a page you read.
14. Connect a real source. SharePoint, Google Drive, Notion or Slack, read-only at first. Then ask what the tool can now see that it shouldn’t. You have it when you can name two files it must never touch.
15. Connect your own files or a database. Keep it read-only until you trust it. You have it when it got three things right and you wrote down the one it got wrong.
16. If your team writes code, try it on a real project. Write the team’s conventions down first, plan before editing, run the tests after. You have it when the change is small enough that a colleague will actually review it.

Stage 5. Run a method, not just prompts (steps 17 to 20)

17. Split the roles. One pass researches, one critiques the draft, one writes, one tidies. Keep the critic demanding. You have it when the objections show up in the final version, answered.
18. Save your three best methods. As reusable instruction templates, so you stop retyping them. You have it when a one-line instruction runs them correctly.
19. Build a scorecard. Ten tasks you know well, scored out of ten every week, with the misses logged. You have it when the scorecard catches a quality drop caused by a "small improvement".
20. Write the handover document. Your files, your briefs, your saved templates, your scores, and the three mistakes you made. You have it when a colleague finishes a task without you and their questions are about judgement, not mechanics.

The side path: dashboards and automations

It runs alongside stages 2 to 4. This is where the team gets hours back.

1. Describe your data before building anything. The columns, the units, how often it updates, who owns it.
2. Build a dashboard with fake numbers in an hour. A filter bar, four tiles, one trend line. Ugly is fine.
3. Keep one number that matters, the three things that drive it and nine inputs under those. Delete the rest.
4. Swap the fake numbers for a real read-only source, and make the dashboard flag an error clearly if the columns change.
5. Schedule the refresh, and check it can run twice without breaking anything.
6. Add an alert that only fires on real movement, with a short note on what moved and why. That is the difference between a useful dashboard and a wall of numbers.
7. Chain it. A form comes in, a ticket appears with a suggested owner, and duplicates are dropped.
8. Put a human approval in front of anything that matters, and keep a record of who approved what.

A team that ships one reliable dashboard and two supervised automations already has something concrete to show.

Where teams get stuck

• Ten prompts and no scorecard. It feels like progress, with no proof.
• Rewriting from scratch every time, then wondering why quality varies.
• Pasting in everything instead of being specific.
• Automating a process that doesn’t work by hand yet.
• Skipping the critique because it stings.
• Building dashboards nobody owns. They are abandoned within three weeks.

The ten weeks

• Days 1 to 14: steps 1 to 4, twenty minutes a day.
• Days 15 to 28: steps 5 to 8, and the first three steps of the dashboard path.
• Days 29 to 42: steps 9 to 12, and steps 4 to 6 of the path.
• Days 43 to 56: steps 13 to 16, and the last two steps of the path.
• Days 57 to 70: steps 17 to 20, and the first scorecard.

After that, one scorecard run and one handover update a month. Without them, the level drops.

To go faster

On your own. Start with the AI rollout readiness checklist. Ten questions, five minutes, free and with no account. It shows where your team stands before the first step.

With your team. This is my job: change management and tool adoption, from the first instruction to measuring use. Book a 20-minute call. I will tell you honestly whether you need me or just the checklist.

Either way, start with step 1 today. It is the step that makes every other step possible.

Alexandre Gilles spent seven years getting teams to adopt digital tools at Avanade (Accenture × Microsoft) and Crédit Agricole CIB, including 2,500 users of an Enedis business application. He has trained close to 40 professionals one to one and built his own business with AI agents.