When not to use AI
Most problems described to us as AI problems are not. A practical test for whether a model is the right tool, from a company with AI in its name.
We have AI in our name and we talk clients out of it regularly. Not out of modesty — out of experience with what happens six months after a model ships into a process that did not need one.
Here is the test we actually apply.
Is the rule writable?
If a domain expert can state the logic in a paragraph, write the paragraph as
code. A rule engine is auditable, testable, instant, free to run, and it
explains itself when someone asks why a decision went the way it did. A model
that reproduces a writable rule is a more expensive, less reliable version of
an if statement.
Machine learning earns its place when the rule exists but nobody can articulate it — when it lives in a thousand examples rather than in a policy document.
Is the data there, and is it honest?
Models learn the data you have, including its mistakes. If your historical labels were produced by a process that was itself biased or inconsistent, the model will reproduce that faithfully and at scale. We have walked away from projects where the training data encoded a decision process the client was actively trying to move away from.
Before any modelling, we ask: how was this data produced, by whom, and what were they optimising for at the time?
What happens when it’s wrong?
Every model is wrong sometimes. The question is what that costs.
Wrong product recommendation: a slightly worse browsing session. Wrong fraud flag: an angry customer and a manual review. Wrong clinical suggestion acted on without review: something considerably worse.
The higher that cost, the more of the budget goes to evaluation, monitoring and human review paths — and the more likely the honest answer is that the process should stay human with software assisting, rather than the reverse.
Who maintains it?
A model is not a deliverable, it is a commitment. Data drifts, behaviour shifts, and accuracy degrades quietly — a model does not throw an error when it starts being wrong. Somebody has to own retraining and monitoring.
If nobody will own that after we leave, we are building something that will be subtly broken within the year and nobody will notice until it matters.
When the answer is yes
None of this is an argument against AI. When the pattern is real but unarticulable, the data is honest, the failure cost is survivable and somebody owns it — machine learning does things nothing else can, and the results are not marginal.
We would just rather get to that conversation by elimination than by enthusiasm.
This is a seeded example post, included to show the blog’s formatting and set the editorial tone. Replace or delete it before launch.
