A practical test for useful AI in software products

Start with the task, not the model

A product idea can begin with a capable model and still lack a clear user need. A more useful starting point is the task: what is someone trying to do, where does it become difficult, and what constraints shape a good outcome?

This framing keeps the proposed capability open to comparison. A smaller interface improvement, a clearer process or no AI at all may address the task better than adding generation or automation.

Name the evidence before building

Before selecting a model or interaction, write down what a helpful result would look like and what failure would cost. The measures depend on the task. They might concern completion, correction effort, time, comprehension, consistency or another observable outcome.

A prototype can help reveal missing assumptions, but a prototype is not evidence of sustained value. Evaluation should match the intended context and include cases where the system should decline, ask for clarification or hand control back.

• Who is the intended user, and what decision or task is in scope?
• What baseline or alternative will the capability be compared with?
• Which errors matter, and how will they be detected?
• What result would make the team stop or change direction?

Include the cost of interaction

Generated output can move effort rather than remove it. People may need to check, correct or reconcile a result. The product question is therefore not only whether a system can produce an answer, but whether the complete interaction supports the person’s task with acceptable effort and risk.

What to examine next

A practical next step is to describe one user task, prototype the smallest capability that could support it, and agree on evaluation criteria before reviewing results. This is a proposed method for exploration, not a finding from a completed experiment.