How to Pick an AI Use Case That Actually Pays

Most failed AI projects were doomed at selection

By the time an AI project is visibly failing, it is usually eight weeks and a chunk of budget in. But the decision that doomed it was made in an afternoon at the start, when someone picked the most exciting idea rather than the most tractable one.

Selection deserves more rigour than it usually gets, and the criteria are not mysterious.

The four traits that predict success

1. Volume. Hundreds of instances a month at least. Below that, a person doing it carefully is cheaper and better.
2. Tolerable error cost. A wrong answer should be recoverable — an embarrassment, not a lawsuit.
3. A clear definition of correct. If two of your own experts would disagree about the right answer, you cannot evaluate the system, and if you cannot evaluate it you cannot improve it.
4. Existing material. Documents, examples, past decisions. AI works from what exists; it does not conjure institutional knowledge.

The four shapes that reliably work

• Extraction — pulling structured data out of documents, emails and forms. Boring, high volume, verifiable.
• Classification and routing — deciding what something is and where it goes.
• Answering from your own corpus — support, internal knowledge, policy questions.
• First-draft generation — where a human reviews and owns the output.

All four share a property: a person can check the output in seconds. That is what makes them safe to deploy and easy to improve.

The three that usually disappoint

Anything requiring genuine domain judgement with legal or clinical consequence. Not because the technology cannot produce plausible output, but because the liability sits with you and review costs as much as doing the work.

Prediction without enough history. Forecasting needs years of clean data. Most businesses have eighteen months of messy data and a system change halfway through.

“An AI that understands our business.” This is not a use case, it is a mood. Every project that starts here converges on one of the four shapes above after a month of expensive discovery.

Score your candidates in half an hour

Anything scoring 20 or more is a strong candidate. Under 12 and you should expect a research project rather than a delivery project — which is fine, as long as everyone knows that is what they are funding.

Prove it on fifty real cases first

Before commissioning a build, take fifty real historical examples with known correct answers and test the approach against them. This costs days, not weeks, and it settles arguments that would otherwise run for months.

If accuracy on fifty real cases is poor, no amount of engineering rescues the use case — the problem is the problem, not the implementation.

Frequently asked questions