A small first deployment to find out whether the idea works. Most demonstrate that the technology functions, which was never in doubt, and fail to establish whether it delivers value in the actual workflow with real users and real edge cases. A good pilot is designed to be capable of failing.
Technical view
Should state in advance the hypothesis, the success measure with a threshold, the population and duration, and what will happen on each outcome. Include a baseline comparison against the current process. Common failure patterns are selecting enthusiastic users unrepresentative of the population, curated inputs that avoid the hard cases, and measuring output volume rather than outcome quality. Plan for the transition to production from the start, since pilot infrastructure rarely survives contact with scale.
Related
Part of AI Terminology Index, section: Product and Adoption.