| What | Question | What we need |
|---|---|---|
| Problem | What problem are we solving, and for whom? | A specific problem articulated in terms a user would recognize — not a technology direction |
| Input | What's the input, exactly? | Precise description of what enters the system — format, variability, messiness, volume — with real examples you can point to |
| Output | What's the output, exactly? | Precise description of what the system produces — structure, fields, format — with a concrete example of a correct output for a given input |
| Scope and boundaries | What's in scope? What's explicitly not? | Clear boundaries — what this system does, where it stops, what adjacent problems you're deliberately not solving |
| Evaluation bar | What does "good enough" look like, and why? | A specific threshold with a rationale grounded in the cost of errors, the current baseline, and the minimum quality that delivers value |
| Errors to catch | Which errors are tolerable? Which are fatal? | The error hierarchy — from minor annoyances to catastrophic failures — and the asymmetry between error types |
| Data | What data do you have to work with? | An honest inventory of what you actually have today — labeled examples, codified knowledge, historical outcomes, and known gaps |
| Approach | What approach fits this problem shape? | A technical choice matched to the problem shape, not to hype — with honest consideration of whether a simpler approach would work |
| Human intervention | Where does the human stay in the loop? | The explicit design of the human role — where, why, and what happens when the human disagrees with the system |
| Proof | How will you evaluate? | How you'll measure during development and monitor in production — ground truth, metrics, failure review, and feedback loops |
Not "we're using AI for procurement." What specific problem, for what specific person.
If we can't articulate the problem in terms a user would recognize, we are solving something abstract.
Not "customer data" or "documents." Exactly what enters the system — format, variability, messiness, volume.
If we can't point to real examples right now, we don't understand the problem yet.
Not "recommendations" or "insights." The precise structure, fields, and format the system produces.
If we can't write out what a correct output looks like for a given input, we can't evaluate.
Not everything. Define where this system stops and what adjacent problems you're deliberately not solving.
AI invites scope creep. If we haven't drawn boundaries, we’ll build something that tries to do everything and does nothing well.