Begin with the decision, not the model
The most useful AI projects start by naming a specific decision or workflow that is slow, inconsistent or expensive today. When the target is clear, it becomes possible to judge whether a model is needed at all, what accuracy is good enough, and who remains accountable for the outcome.
Make data quality visible early
Most AI delays are data delays. Before building, list the sources the system will rely on, who owns them, how fresh they are and what personal or sensitive information they contain. Gaps found in week one are cheap; gaps found after launch are not.
Evaluate usefulness, reliability, safety and cost together
A prototype that impresses in a demo can still fail in production. Agree an evaluation set that reflects real cases, measure it on every change, and track operating cost alongside quality so the system stays sustainable.
Keep people where judgement matters
Automation should remove repetitive work, not accountability. Where consequences are significant or inputs are ambiguous, design review steps, clear escalation and an audit trail from the beginning.