Foundations
AI cannot fix a process you have never written down
You bought the subscription, you watched the videos, and three weeks later almost nothing has changed. That is the normal outcome, and it is very rarely because the tool was wrong.
The pattern is consistent enough to be predictable. Someone adopts AI with real enthusiasm, gets two or three genuinely impressive results, and then quietly stops using it for anything that matters. The tool did not break. It ran into the fact that nobody could tell it what the job actually was.
The instruction is the product
An AI tool does exactly one thing: it turns an instruction into an output. Everything you get out of it is downstream of how well the instruction describes the work.
So when the work only exists in somebody's head — when the quoting process is "I look at it and I know", when the tone of your customer emails is "you just get a feel for it" — there is nothing to hand over. The AI produces something generic, because generic is the only thing an unspecified request can produce.
Write the process down first
This is the step everyone skips because it is dull and produces nothing you can show anyone. It is also the step that decides whether any of the rest works.
Take one job you repeat — answering a common customer question, writing a quote, drafting a listing — and write down how you actually do it. Not how it should be done. How it is done, by you, on a normal day.
- What starts it. An email arrives, a form is submitted, it is Monday.
- What you need before you begin. The customer's history, the price list, last month's numbers.
- The steps, in order. Including the judgement calls, written as rules rather than instincts.
- What good looks like. An actual example of a finished one.
- What would make it wrong. The mistakes you would send back.
- Where it ends up. Sent, filed, published.
Thirty minutes per job, roughly. You will find at least one step you had never articulated, and usually one you do for no reason anybody can remember.
Why this doubles as the prompt
Here is the part that makes the tedium worthwhile: the document you just wrote is the prompt. Not a summary of it — the thing itself.
Context at the top, steps in the middle, a worked example, the failure conditions, the output format. That is what a good prompt looks like, and you have written one without thinking about prompting at all.
This is also why prompt libraries disappoint on their own. A prompt written for a generic business describes a generic process. Yours is worth more because it contains the specifics only you know.
Start with the boring jobs
The instinct is to point AI at the most interesting work — strategy, the big proposal, the thing you find hardest. That is exactly backwards.
The best first candidates are the jobs that are repetitive, low-stakes, and already have a right answer you would recognise instantly:
- Turning notes into a tidy summary
- Answering the same five customer questions in your own words
- Reformatting data from one shape into another
- First drafts of routine descriptions, listings or updates
These are unglamorous and they are where the hours actually go. They are also safe places to be wrong while you learn what good instructions look like.
The test that tells you it worked
Not "was the output impressive". The real test is whether you could hand the same instruction to someone else and get a usable result without being in the room.
If yes, you have documented a process and you own something reusable. If it only works when you are there to nudge it, the knowledge is still in your head and you have automated nothing.
What this actually buys you
The written process outlives whichever AI tool you are using this year. Tools change constantly; a clear description of how your business does a thing does not.
It is also the thing you need for hiring, for delegating, for going on holiday without your phone. People treat that as a side effect of AI adoption. It is closer to the main event — the AI is just what finally made it worth doing.