Do we need to pick an AI tool before defining our processes?

No. Define the work first. A clearly written process will run on any of the major AI tools, and writing it down often tells you which tool fits.

How long does it take to define one process?

Hours, not weeks: one session watching the person who does it best, and one test where someone else follows the page literally.

What kind of work should we start with?

Something that happens every week, that one person does well, and that has a clear finished result, like a weekly client report or a follow-up email.

AI can only do work that's been defined

When a team tells me AI "didn't work" for them, I ask to see the instructions they gave it. Usually there aren't any. There's a chat window, a vague request, and a result that was almost right and therefore useless.

The tool is rarely the problem. The problem is that the work was never written down. People do it from memory and judgement, and AI has neither until you give it some.

Why rollouts stall

Most teams start with the tool. They buy seats, run a lunch-and-learn, and wait. A few curious people get real value. Everyone else tries it twice, gets something generic, and goes back to the old way.

The curious few succeed because they're doing the hard part without noticing: they explain the job in detail, correct the output, and try again. They are defining the process as they go. It just lives in their heads and their chat history, where nobody else can use it.

What "defined" means

A process is defined when someone who has never done it could follow it and get an acceptable result. For AI, that means five things are written down:

The trigger. What starts this work? A new signup, the first Monday of the month, a client email.
The inputs. Where the information lives, and which source wins when two disagree.
The steps. In order, including the boring ones people forget they do.
The judgement calls. This is the part that's usually missing. "If the numbers dropped, say so in the first line." "Never promise a date." "If the client sounds annoyed, flag it and stop."
What good looks like. Two or three real examples of finished work that you were happy with.

How to write a process down

Don't ask people to document their work. They'll write what they think they do. Sit with the person who does it best and watch them do it once, on a real piece of work. Ask "why" every time they make a choice. Write down what they say in their words.

Then hand that page to someone else on the team and ask them to follow it literally. Every place they get stuck is a gap AI would have fallen into too. Fix the page, not the person.

This is hours of work per process, not weeks. And the page is useful even if you never automate anything. It's how you train the next hire.

Then the AI part is easy

Once the process exists on paper, turning it into something AI runs is mostly plumbing: connect it to the tools where the inputs live, give it the page, and test it on one real project before it goes anywhere near a second one.

Which AI you use matters much less than people expect. A clear process works on ChatGPT, Claude, Gemini or Copilot. An unclear one fails on all of them.

I saw this at a 25-person marketing agency. Once there was a simple way to write a process down and share it, the team wrote most of their shared processes themselves. They knew the work. They just needed somewhere to put it.

Start with one

Pick a piece of work that happens every week, that one person is good at, and that everybody else avoids. Define that. Build that. Get it running with the team reviewing the output. Then do the next one.

One process, properly defined and actually used, beats a company-wide AI strategy that nobody can point to.