Define AI products with full rigor

Define AI products with structured canvas covering problem, approach, data, evaluation, and responsible AI

Pre-build rigor check for any AI feature; prevents building technically impressive but commercially useless models; required for user-facing AI with safety implications.

• · User problem statement
• · Proposed AI solution approach
• · Available training and evaluation data

• · AI Product Canvas (PDF or MD) with eight sections: problem, approach, data, evaluation, UX design, responsible AI checklist
• · Go/no-go recommendation for build

• · Optional: Model API (GPT-4, Claude, Gemini) or fine-tuning framework
• · Optional: Evaluation framework (test sets, benchmarks, human review panel)

• · Specific user problem (not 'add AI because we can')
• · Minimum acceptable accuracy threshold defined
• · Training and evaluation data identified

• · Accuracy target undefined before build starts
• · No plan for when model is wrong (user-facing without fallback)
• · Training data not audited for bias
• · Metric chosen is not user-facing (optimizing for accuracy when user cares about latency)

• · Eight anti-patterns explicitly called out upfront (no 'add AI' decisions, undefined accuracy, no fallback plan)
• · Task-type taxonomy (classification, generation, summarization, recommendation, search, prediction, conversation, agent)
• · Data audit checklist: bias risk, privacy considerations, data ownership
• · Evaluation framework with offline (test set), online (A/B test), and adversarial components
• · Responsible AI checklist (bias audit, fairness evaluation, hallucination risk, opt-out mechanism)