Intelligence where it earns its place: a practical test for AI features

Almost every product roadmap now has an “AI” line on it. Some of those lines will make products dramatically better. Many will add cost, latency and a new class of bugs for no measurable gain.

Our position is simple: intelligence where it earns its place. Before a model goes into a product, it has to pass four questions.

1. What outcome improves, and how will we measure it?

“Make search smarter” is not an outcome. “Reduce the share of searches with no useful result from 18% to under 8%” is. If we cannot name the metric, we cannot tell whether the model helped — and we cannot tell when a new version made it worse.

2. Do we have an evaluation set before we have a prompt?

We write the test before the feature: a few hundred real, representative examples with the answers we expect, including the hard and the ugly ones. Every change — a new prompt, a new model, a new retrieval strategy — runs against it. Demos persuade; evaluation sets decide.

3. What does a wrong answer cost, and who catches it?

A model will be wrong some of the time. The design question is what happens then. A wrong product recommendation is cheap. A wrong figure in a financial summary is not. The higher the cost of an error, the more the design needs citations, confidence thresholds, human review or a deterministic fallback.

4. Do the economics hold at scale?

A feature that costs a few cents per request is invisible in a prototype and decisive in production. We estimate cost and latency per request at expected volume, and we look for the simplest approach that meets the quality bar — sometimes a smaller model, sometimes caching, sometimes no model at all.

Often, the answer is simpler software

Running this test regularly produces an unglamorous result: a good share of “AI features” are better solved by clearer data, a better search index or a well-designed form. That is not a failure of the test. It is the point of it.

When a model does pass, it ships with its evaluation set, its monitoring and a clear owner — so that “intelligent” stays true after launch day.