This morning, I typed two letters: “GO”. In return, I got my day, my delays, my messages to sort and the list of what was waiting for me. I didn't write a single instruction.
Since the end of August, I have been testing an executive copilot on myself. I use it every day, and I hardly prompt anymore. It's not that I have become better at phrasing my requests: it's that the work of phrasing has been done once and for all, in the system.
The individual prompt eventually hits a ceiling
In early November, I analysed the Copilot usage report of an industrial group I support. More than 9 people out of 10 were active. Fewer than 1 in 4 used it almost every day.
Microsoft counts a user as active as soon as they have triggered a Copilot feature during the period (Microsoft Learn, Microsoft 365 Copilot usage report). A single request in 28 days is enough.
The gap between these two figures tells a familiar story. The curious try it. A few become very good. The others fall back into their habits, because every request costs them an effort: remembering the context, choosing the right documents, specifying the format, reviewing.
For an executive, the effort is even more visible. Their time is limited. They have neither the desire nor the availability to look for the perfect wording, and they are right.
Training people to prompt remains the first step
I train teams on Copilot. They learn how to ask well: a role, a task, a context, a format. It is essential. Without these basics, nobody knows what the tool can do.
In these training sessions, one moment always gets a reaction. Everyone generates their profile and analyses their own writing style, then pastes the result into Copilot's settings. The same email, requested before and after, no longer looks anything like machine-written text.
At that moment, the participants are not prompting better. They have set a framework once, and the tool uses it with every request.
That is the next step. It changes the nature of the work.
A framework set once
An executive copilot takes this principle all the way. The human no longer writes every instruction. They define a framework, and the system reuses it with every request.
This framework usually contains:
- the person's role, priorities and routines;
- the trusted sources and the ones that are excluded;
- the expected formats for each deliverable;
- the rules of caution: confidentiality, review, validation before anything is sent.
The executive expresses a short need. The system fetches the context, applies the rules and produces a working draft. Quality no longer depends on the wording of the day.
The testers of the Executive Copilot put it in their own words. One of them, after three weeks: “less procrastination”. Another: planning that is “better, more regular”. The same person adds that the tool still takes “quite a lot of time every day”. That is the right signal: the framework saves time on phrasing, and it is now the start-up routine that needs to be lightened.
Governance lives in the system
When everyone prompts on their own, the rules depend on everyone's memory. In an international group or a regulated sector, that is not sustainable.
A framework designed once makes it possible to build AI governance into the tool itself: which sources, which access rights, which validations before distribution. In my own copilot, one rule never changes: nothing goes to a third party until I have reviewed it. The system prepares a draft, I decide.
This changes the role of the legal, compliance and IT departments. They no longer arrive at the end of a project to block it. They write part of the framework from the start. It is often the missing piece in a deployment that is stalling.
What needs to change in how the programme is run
Three shifts seem necessary to me.
- Move from training people to prompt to designing frameworks and workflows, role by role.
- Give each use case a business owner, in addition to the sponsor.
- Measure each level separately: access, training, usage, adoption, impact. An assigned licence says nothing about adoption, and an active user says nothing about impact.
AI change management means sparing people effort where it adds nothing, and training them where their judgement matters: reviewing, arbitrating, deciding.
Starting without a big bang
- Choose a volunteer executive and three recurring tasks: preparing a committee meeting, sorting the latest news, tracking decisions.
- Describe the work as it is done today, step by step.
- Write the framework: role, sources, formats, rules of caution.
- Have it reviewed by legal, compliance and IT.
- Test for four weeks, then adjust with the person.
- Track simple indicators: days of use per week, tasks covered, the person's feedback.
The goal is a first visible case in production. If it works for one role, the method can be replicated.
Build the framework, then let it work
Training people to prompt opens the door. Building the framework brings AI into everyday work.
Do you have an AI programme moving more slowly than planned? Let's talk to identify what is blocking it and get adoption moving again.