Start from where you want to end up, then work back
Most AI plans start from today. Which of our processes could use AI? Which tool fits which department? The answers are always reasonable, and they always keep the company as it is, a little faster. The existing job titles, approval chains and reporting lines quietly decide what is allowed to change.
Plan backwards
The alternative is to start from the end. Describe the company as it would look if you built it today, from scratch, knowing what software can now do. How many people would you hire, and for what? Which decisions would be made by a system within limits you set, and which would always stay with a person? Where would the data live? Then work back: what has to be true a year before that, and a year before that, until you reach Monday.
This is not a vision statement. It is a working document, specific enough that a manager can check a new project against it and say whether it moves the company closer or not.
What the document should contain
What the company is really for: not the activities it performs today, but the value it creates and for whom. Which tasks are done only to pass information from one person to another. The target picture: who does what, people and systems, and how the people whose work changes are looked after. The first process to rebuild. And the signals that will tell you when to speed up or change course.
It is worth one or two full days with the leadership team, away from daily work. The rule for those days: what looks realistic today may decide how fast you move, never which direction.
Why this comes first
When an AI project fails although the technology worked, the cause is usually that nobody had agreed where the company was going. Each team optimised its own corner, and the corners did not add up. Agreeing on the destination first is cheap insurance against that.
Write the destination down, sign it, and check every project against it.