A team asks for an AI pilot, and the first question is usually which tool to try. I’d start somewhere less exciting: trace one task through the people, documents, approvals, and exceptions that already shape it. The best pilot often begins with a workflow that is tedious, bounded, and important enough for someone to notice if it improves.
Follow the work first
Take a real process, such as preparing a meeting brief or sorting incoming service requests, and map what happens from arrival to handoff. Note where information is missing, who checks it, and which decisions depend on context that never appears in the form. Those details help distinguish a language task from a coordination problem.
Then choose a narrow step where assistance could save attention without taking away judgment. Drafting a first summary might be a reasonable test; deciding which request deserves urgent treatment may not be. A clear boundary gives people a way to challenge the output instead of treating fluency as proof.
Design the human check
A useful pilot names who reviews the result, what they should verify, and what happens when the system is uncertain. If staff must quietly redo the whole task, the experiment is measuring a demo rather than a change in work. Give reviewers a simple way to record errors and recurring edge cases.
The right measure depends on the task: fewer minutes spent assembling a draft, fewer incomplete submissions, or clearer handoffs between teams. Pick a baseline before introducing the tool, and include the review effort in the comparison. A faster first step is not a gain if it creates a slower last step.
Keep the decision reversible
Use information the team is allowed to share with the chosen system, and make the trial small enough to pause if the output is unreliable or the workflow changes. Record what the tool can and cannot do, who owns the process, and what evidence would justify another round. That note becomes more valuable than a polished pilot presentation.
AI in practice is less about finding the most impressive model than making a careful change to how work happens. Start with a task people recognize, keep a person responsible for the consequential call, and measure the whole path. If the pilot teaches the team something useful even when it fails, it was still worth building.
