When not to use AI

Placeholder article. Replace with your own editorial before publishing.

We are asked for AI more often than we are asked for a solution. The two are not always the same thing, and saying so has occasionally cost us the engagement.

The questions we ask first

Is the rule actually deterministic? If a human follows a written policy to reach the answer, encode the policy. A model that reproduces a rulebook 94% of the time is worse than the rulebook.

Is the data good enough? A model trained on inconsistent historical labels learns the inconsistency. Fix the data collection first; often that alone resolves the problem that prompted the request.

Can you tolerate being wrong? Recommendation and triage tolerate error. Statutory eligibility and payment authorisation do not. Where a wrong answer has legal consequences, the model can assist a human decision, not replace it.

What is the baseline? If nobody has measured how well the current manual process performs, there is no way to know whether the model improved anything.

Where the answer is yes

Unstructured input at volume — scanned documents, free-text complaints, images. Forecasting where the pattern is real but too complex to write down. Detection problems where the interesting cases are rare and humans miss them at scale.

That is a genuinely useful list. It is just shorter than the pitch deck suggests.