When a company decides to “start with AI”, the chosen process usually comes out of the leadership meeting, and it is usually the most visible one: the one with the biggest budget, the most complaints, or the most airtime at the committee.
Visibility is a terrible criterion. These five work better.
The five criteria
1. Volume and repetition. How many times a month does it happen? Below a few hundred repetitions, the effort of building and maintaining doesn’t pay back. Good candidates are boring and constant.
2. A verifiable correct answer exists. This one is decisive. “Extract the total amount from an invoice” has a correct answer; “write a compelling proposal” does not. Without ground truth you cannot measure, and without measurement you can neither improve nor defend the project.
3. Low or reversible cost of error. What happens if the system gets it wrong? If the answer is “someone corrects it at review”, excellent. If it is “it goes to the customer and we lose the account”, that is not your first project.
4. The data already exists and is reachable. Data quality is the ceiling. If the process depends on information nobody recorded, or that lives in one person’s head, don’t start there.
5. There is an owner with skin in the game. Someone whose job measurably improves if this works. Without that person, the project dies at the first friction.
The score
Score each process 1 to 5 on the five criteria and add them up. Our rule:
| Score | What it means |
|---|---|
| 21–25 | Start here. Seriously. |
| 16–20 | Good candidate for the second wave |
| 11–15 | Needs prior work, usually on data |
| Under 11 | Not an AI problem, or not yet |
The value of the exercise is not the score. It is that it forces an honest conversation with the teams involved, and usually reveals that leadership’s favourite process scores a 9.
The usual suspects that score high
At the mid-sized companies we work with, these come up again and again:
- Classifying and routing whatever arrives via a shared inbox or form.
- Extracting data from documents — invoices, delivery notes, orders, policies.
- Reconciling two systems that should match and don’t.
- Drafting recurring reports somebody rebuilds every week.
- Searching internal documentation nobody can find because it lives in four places.
They are not impressive. That is why they work.
The one that almost always scores low
“An assistant that answers any question about the company.” Diffuse volume, no verifiable answer, variable cost of error, scattered data and no clear owner.
It is the most frequently proposed project and the one that most often ends up in the 95% that produces no measurable return.
A warning about sequence
The first project should not be chosen to maximise value. It should be chosen to maximise the probability of success.
You need a small, fast, measurable win that teaches the organisation what this looks like: what it asks for, what it produces, how it gets reviewed. With that reference, the second project can be ambitious. Without it, there is no second project.