The first time I helped a leadership team roll out AI, I made the mistake everyone makes. I started with the tools. Here's Projects, here's how you write a good prompt, here's a cool thing it can do. Everyone nodded. Nobody used it the next week.
The second time, at Politiken, I started somewhere else: not with what AI can do, but with what it must never touch. That's the version that stuck.
Here's the whole setup. It takes an afternoon, and you can run it on your own team without me.
Governance first, tools second
Most AI rollouts fail quietly. The tool is usually fine. People just don't trust it, so they use it for nothing that matters. They ask it to reword an email and go back to their real work.
Trust comes from knowing where the edges are. So before you show anyone a single feature, you draw the fence.
The mental model I give every leader is four words: Claude drafts, never sends. The AI writes, proposes, summarizes, and a human decides. Nothing leaves the building on its own. Once people believe that, they relax, and they start handing it the work that actually eats their week.
Step 1: Name the five tasks that eat your Monday
Sit each person down and ask one question: what are the five recurring things you do every week that you'd hand to a sharp assistant if you had one?
Not aspirational. Real. The weekly status summary. The first draft of the board update. Turning messy notes into a decision memo. Prepping the same kind of stakeholder email over and over. The recurring report nobody enjoys.
Write down five. That's the whole target. You're pointing a tool at five specific leaks, not "adopting AI."
Step 2: Name the red zones
Now the more important list. What must the AI never touch?
At a newspaper, this was source protection, personnel data, and live contract negotiations. At your company it's something else, but you have it. Legal exposure. Customer data you're liable for. The places where a plausible-but-wrong draft could cause real damage.
Have each leader write their own red zones. Not IT's list, theirs. It's the most defensible governance move you can make, because the person who knows the danger is the person doing the work.
Five green tasks, a short red list. That's your map.
Step 3: Build for the five, fence off the rest
Now, and only now, the tools.
For each of the five tasks, set up a Project with the context baked in: the tone, the format, an example or two of a good output. Add custom instructions so it defaults to how this person actually works. If a task repeats in a fixed shape, write it once as a reusable skill so they don't re-explain it every Monday.
That's it. You're building five small, trustworthy machines for five real jobs, inside the fence. Nobody has to learn "prompting."
The exec-grade pattern (a trap to avoid)
Someone will ask to connect the AI to the live dashboards. Power BI, the data warehouse, the analytics stack. Resist the developer-grade version of this.
The exec-grade pattern is boring and robust: export the five to ten numbers a leader actually cares about to a place the AI can read (a shared sheet, SharePoint), and let it read from there. No preview tooling, no fragile middleware, and it scales across the whole team trivially. The analyst who lives in the data can have the deeper integration. The leader needs the five numbers, explained.
Step 4: Set the failure rules out loud
Two more things, said plainly to the room.
Treat every draft as a draft. An email from an unknown sender can carry instructions that try to steer the AI. So read the recipients, read the content, and never let it act automatically on something external. It's the same instinct you already have about a suspicious link.
And decide the escalation path now. When the AI drafts something embarrassingly wrong (it will, eventually), who do they call? For the first month it's usually me. After that it should be your own IT or DPO. Name the person before you need them.
Measure the five, not the tokens
Six weeks later, don't ask how much AI everyone used. Ask about the five.
Did the board update go from three hours to twenty minutes? Did the memos get sharper? Did anyone get a piece of their week back? Those are the numbers that tell you whether this worked. The token counter tells you nothing. I wrote a whole piece on why: Measure AI for revenue, not activity.
The part I left out
This is the entire method, and I'm giving it away because the method was never the hard part.
The hard part is the afternoon itself. Pulling the real red zones out of a room that's a little defensive. Spotting that someone's "five tasks" are actually one task in a trench coat. Getting the setup to fit how a specific person genuinely works, so it survives past week one. That's the job, and it's the only piece I charge for.
Run the steps yourself. If it sticks, great, that's the point. If you want someone in the room for the afternoon it doesn't, that's where I come in.