Prove the job and the evidence
Describe the workflow in observable terms: who does what, what consumes time, which decisions require judgement and what a better result is worth. Then gather representative examples, including the ambiguous and adversarial cases that a demonstration usually avoids.
- Is the task frequent or valuable enough to justify a system?
- Can quality be evaluated against representative examples?
- Do you have permission and a retention policy for the required data?
Design for incorrect output
Model output is probabilistic. Decide which mistakes are tolerable, which require human review and which must be prevented with deterministic controls. The interface should communicate uncertainty where the user needs it rather than presenting every answer with equal confidence.
Measure the whole product
A model score is only one part of product performance. Track completion quality, latency, cost, user correction, escalation and the effect on the underlying business outcome. Establish a versioned evaluation set so changes can be compared before they reach every user.