01
Start with the shape of the input
AI earns its cost when the input is unstructured or ambiguous: free-text emails, scanned documents, recorded calls, messy product descriptions. If the input is already structured and the rules are stable, deterministic automation is cheaper, faster and easier to debug.
A useful test: could a competent new hire follow a written checklist to do this task correctly every time? If yes, you probably want automation, not a model.
02
Then look at the cost of being wrong
Every probabilistic system is occasionally wrong. What matters is what happens next. If a wrong answer is recoverable — a customer asks a follow-up, a reviewer rejects a record — AI is usually viable. If a wrong answer is silently written into a financial or clinical system, you need validation, thresholds and human review before you need a better model.
- Recoverable errors: suggestions, drafts, triage, routing, search
- Costly errors: payments, medical decisions, legal commitments, irreversible writes
- Anything costly needs confidence scoring and a human review queue
03
Finally, check whether the process should exist
The most valuable output of a discovery workshop is often the discovery that a step is no longer necessary. Automating a redundant approval makes the redundancy permanent and harder to remove later.
Map the process as it really runs, including the workarounds, before deciding what to build.