Readiness for AI begins with one task and a person who can judge the result. A narrow example, such as suggesting a category for a support request, lets the team examine data permissions, quality, and fallbacks before committing to broader use.
Write the task and its boundaries
Describe the input, expected output, and action the system may take. Decide whether the output is advisory, queued for review, or allowed to trigger a change. Identify prohibited inputs and actions explicitly. A proposed answer and a sent message have different risks and require different checks.
Check data and permissions
Identify the owner of every source and confirm the pilot audience may see the information. Sample for missing fields, contradictory instructions, and material that belongs to a different group. Decide who updates the source and how corrected information reaches the pilot. If an owner cannot resolve disagreements in the material, keep a manual fallback.
Decide what a pilot must demonstrate
Create a small set of representative examples, including an ambiguous input and a request the tool should decline or escalate. Decide who reviews outputs, how issues are recorded, and when the trial stops. Compare the review effort and failures with the current process before expanding the scope.
Example
Evaluate a policy assistant with approved documents, unavailable answers, and access boundaries.
What to do next
Proceed when the owner can approve the sources, review expected and unexpected results, and pause the pilot.