Start with roles, not tools.
Choose a tool first and you design around a vendor's offering, not around the work of the people doing it. That is the most expensive mistake in adopting AI.
The order is wrong.
Most efforts start with a tool. A licence bought, a platform rolled out, and only then the question of which work fits it. You know how that ends: expensive licences left unused, because no one looked first at where the work stalls. The tool became the starting point. The workflow should have been.
The role map.
Before a tool comes into view, I map for each role which tasks suit automation, augmentation, an agent, or deliberate human work. Not everything has to move to AI, and not everything can. The map shows where the gain sits within a role, and where the human stays essential. The attention is on the people doing the work, not on a vendor's brochure.
The zoom-in.
For one role I lay the working week side by side. The hours as they run now, and the hours with AI as the director. No promise about productivity in general, but a visible difference in that one role, with the building blocks that land first. What works there is the proof for the next.
What it gives you.
Start with roles and the choice of tool becomes a consequence instead of a gamble. You know what you need, because you know which work it has to carry. The organisation recognises itself in the plan, because it is about their work. And the chance of a tool gathering dust gets small. Start with the people. The technology follows.