Use rules when the answer is defined
If a valid identifier maps to one known record, a deterministic lookup is easier to explain and retest. In the catalog-enrichment project, the runtime moves through exact match, structured fields and normalized descriptions. It does not ask a language model to invent a fiscal classification for each request.
Use AI when the input needs interpretation
Evidence photos, free-form descriptions and speech can contain useful information without a fixed structure. The audit platform uses image interpretation within an explicit scoring workflow. Model output still needs a defined format, checks and a place for exceptions.
Keep consequential decisions reviewable
Ask what happens when an interpretation is wrong. A useful design separates extracting information, proposing a decision and committing an action. Put a human checkpoint where the cost of an incorrect action outweighs the benefit of immediate automation.
Test the exception path
Before release, try missing evidence, repeated submissions, ambiguous input and a model timeout. Record the expected handling for each. An automation that succeeds only on clean examples leaves the team to discover its operating rules after launch.
A brief to bring to discovery
Describe one recurring decision. Include a sanitized typical input, a difficult example, the expected outcome and who currently approves it. That is more actionable than starting with a request to add an AI agent.