Find the work before buying seats

A company can give everyone an assistant and still have the same late reports, repeated customer questions, and overfull review queues. The software is present. The work has not been redesigned. An employee may try a few prompts, discover that the assistant cannot reach the right source or that nobody knows whether its answer is safe to use, and return to the old method.

Start with a job people can point to. “Prepare the weekly supplier exception summary” is better than “increase AI productivity.” The first has source records, a due date, a reviewer, and a finished output. The second can produce a training attendance chart without improving anything. Ask the team to show you the last five completed examples. The awkward parts usually become visible within minutes.

Do not make the first pilot a task where a wrong answer moves money, changes a clinical record, or sends a binding promise. Start in draft-only mode when you can: the assistant suggests; a qualified person checks and decides. That is enough to learn whether the tool helps, and it keeps the first experiment reversible.

A good starting task has several examples, a clear definition of “good enough,” and at least one person willing to compare the old and new methods. It need not be glamorous. A service manager might choose first-draft answers for routine warranty questions. A finance lead might choose a variance explanation that still goes to an accountant. A plant supervisor might choose a maintenance handoff summary, not an automated maintenance decision.

Write down the baseline before the team sees an AI demo. Take a small, ordinary sample—perhaps ten recent cases—and record elapsed time, hands-on time, corrections, and whether the recipient had to come back. The sample is not a company-wide statistic. It is a starting point you can compare honestly with the next ten cases.

The sponsor should be the manager who owns the queue and can change it. Information technology can provide access and security; it cannot decide whether a revised supplier summary is operationally useful. A senior executive can clear time and budget, but the working owner must stay close enough to see errors and change the process.

A high-volume task may still be a poor first choice if its source documents are contradictory, its owners disagree on the correct answer, or a mistake is hard to reverse. Fixing the source or approval path may be the first project. That is progress, not an admission that AI failed.

Before any purchase, make a one-page candidate list. Rank each job by recurring volume, pain, clarity of a correct answer, consequence of error, access to approved data, and willingness of the owner to test. You do not need a mathematically precise weighted score. A short explanation for each choice is more useful than decimal points.