Your AI licences are live. The work hasn’t changed.
Copilot, ChatGPT or Claude has been bought and handed out. The chat box gets used. The board pack, the mailbox and the month-end still run the way they did last year. We work department by department on the jobs themselves: which parts an assistant should do, which parts it must not, and who checks the result.
If this sounds familiar, the licences are not the problem.
- Seats were assigned at renewal, one tier for everyone, before anyone knew who would use them.
- There was a training session. It showed features. On Monday people went back to the same spreadsheet.
- You can say how many people logged in. You cannot say which piece of work is now done differently.
Real work does not live in a chat box.
The job, not the chat, is the unit of change.
Take the monthly board pack. Last month’s pack, three source sheets, updates from six people, figures that have to reconcile, commentary that needs judgement, and a board folder that has to be ready by Thursday.
A prompt can draft the commentary. It cannot find the evidence, reconcile the total, record who approved it, or put the pack in the folder. So the useful question is not “how do we prompt better”. It is:
| Part of the job | Who or what should do it |
|---|---|
| Drafting the commentary from the updates | Assistant |
| Pulling the source sheets together each month | Automated |
| The figures that must reconcile | Calculated outside the assistant |
| Deciding what the board is told | A named person checks |
Every job in an organisation splits this way. Nobody has usually written the split down.
Much of it needs the licence you already hold. Some of it needs no AI at all.
Mailbox rules, templates, a proper place to keep documents, a habit of checking the answer. These do more for most departments than another feature walkthrough, and they cost nothing extra. We say so, including when it means you buy less.
Generic AI courses are written once for everybody. That is why they are cheap and why they cannot tell a claims handler from a finance analyst. What actually blocks use is mostly people and process: who is allowed to do what, where the documents live, and who signs off. A course does not touch any of that.
Four steps, done with each department.
Organisation, then department, then the people who carry it afterwards.
Start with your own evidence
Who holds which licence, where the work actually lives, what people already do with the tools.
Name the jobs before the seats
Three named jobs per department. The seat decision follows the jobs, not seniority.
Follow the whole job
From the file it starts in to the place it lands, not just the prompt in the middle.
Leave a working method
Templates, checking rules and internal leads who can carry it once we have gone.
If a job needs more than a working method, you will know that too, before buying seats or commissioning a build. How it works in detail
A tidy table, and nothing about it looked wrong.
Tested in public. Same file, same prompts, answers documented before either tool opens it.
In a published Excel test, Copilot and Claude were given the same workbook and the same prompts. Copilot returned a neat, sorted regional table.
The regional total was understated by that much, because the workbook spelt London three different ways and the assistant treated them as three places.
An assistant working in a spreadsheet will produce a plausible number. Where a figure has to reconcile, the deterministic tool is the safer instrument. Your data is messier than a test workbook. We also publish the rounds that go the other way. What the Excel test showed
Better seat decisions are a result of this. They are not the offer.
Nobody decides to over-licence. Giving everyone the same tier looks fair, and the decision is made before anyone has evidence about who uses what. Once the jobs are named, a paid seat is justified by how much a job depends on information held inside your systems, not by who is senior.
We take no reseller margin, no partner tier and no referral fee from any vendor whose tools we discuss. Try it on three jobs of your own.
We have no interest in maximising prompts or paid seats. We decide which jobs are worth touching, which parts should use AI, which parts must not, and what has to be true before the result leaves the room.
Ger Perdisatt
Former COO of Microsoft Western Europe. Non-executive director of two state bodies. ISO/IEC 42001 Lead Auditor. About AI at Work
Write to hello@acuityai.co with the jobs you have in mind and what is licensed today.