Five Times an AI App Is the Wrong Answer

Turning down AI work is part of the job

We build AI applications and we decline them regularly. The projects that should not have been built are the ones that damage a business's appetite for the ones that should.

Check yours against these five before commissioning anything from anyone.

1. The rules are knowable

If you can write the logic down — and for pricing, eligibility, routing and thresholds you usually can — rules are faster, cheaper, perfectly explainable and impossible to talk into doing something strange.

Try writing them first. Succeeding solves the problem for a fraction of the cost; failing teaches you why it is hard.

2. The volume is small

The build, the evaluation and the monitoring are largely fixed costs regardless of volume. Below a few hundred instances a month, a person doing it carefully is usually better and cheaper.

3. The decision must be explainable

Where you must explain to a customer, a regulator or a tribunal exactly why a decision went the way it did, probabilistic output is a liability. Use rules for the decision and AI at most for surfacing information.

4. Nobody can check the output

If verifying an answer takes as long as producing it, automation has moved effort rather than removed it. The four shapes that work — extraction, classification, retrieval-based answering, first drafts — all allow a human to check in seconds.

5. A product already does it

• Transcription, translation, general document extraction, meeting notes — mature markets
• Standard support deflection for common products
• General-purpose writing assistance

Building a bespoke version of a solved problem is how budgets disappear with nothing distinctive to show.

What to do instead

Write the rules. Buy the product. Fix the data capture so a future project is possible. Or wait six months and see whether it is still a problem — a surprising number are not.

Frequently asked questions