Start with the work, not the model
Choose a specific workflow and identify where time, errors, uncertainty, or coordination create a measurable problem. Avoid beginning with a tool and searching for a use case afterward. Interview the people who perform the work and map the current process, constraints, exceptions, and decision rights.
Decide whether AI is necessary. A clearer form, better data integration, a simple rule, or process change may address the problem more safely and cheaply. If AI is appropriate, define its narrow role and the human decisions it will not make.
Set evaluation and stop conditions
Use permissioned data and define a baseline before comparing results. Track more than average accuracy: examine false alarms, missed cases, performance across contexts, review time, and failure recovery. Decide what evidence would support continuing, changing, or stopping the pilot.
Document system ownership, cyber controls, training, vendor responsibilities, and an offline or manual fallback. A successful demonstration is not proof of readiness for deployment.