Frame an AI Product Opportunity

Describe the current work

Map the trigger, people, information, decision, action, delay, and failure cost in the existing workflow. Interview the person who performs and the person affected by the work. Do not begin with “we need a chatbot.” Begin with a friction or decision that can be observed.

Write at least one non-AI alternative: better search, clearer rules, form redesign, automation, training, or an additional human review step. AI must earn its additional uncertainty and operating cost.

Choose the role of AI

Decide whether the system drafts, classifies, retrieves, recommends, predicts, transforms, or acts. Specify the human responsibility around that output. Drafting a support reply and sending one are different products with different evidence and risk requirements.

List required data, permissions, freshness, integration points, latency, cost, and expected failure modes. If you cannot obtain representative data or evaluate quality, the concept is not ready for a model comparison.

Define value and boundaries

Use task metrics such as correction rate, completion time, resolution quality, review burden, adoption, and harmful-error rate. Include an unacceptable outcome and a stop condition. “Users like it” is not enough.

Practice: one-page opportunity brief

Write: user and job, current workflow, pain evidence, non-AI baseline, proposed AI role, human handoff, required data, quality metric, business metric, risk metric, pilot group, and stop condition. Ask a skeptical domain expert to identify assumptions.

Decision

Proceed only when a small pilot can generate trustworthy evidence without exposing users to irreversible or poorly understood harm.