Do not start with the loudest annoyance

Teams naturally nominate the task they dislike most. Frustration matters, but it does not establish value or feasibility. Map a real workflow from trigger to outcome, including exceptions, approvals, workarounds and the systems people use outside the official process.

Score candidates from one to five across frequency, time per occurrence, error rate, customer consequence, data availability, rule stability and reversibility. Record confidence beside every score. A confident medium-value candidate is often a better pilot than an exciting use case built on guesses.

Account for failure cost

Automation changes the speed and scale of mistakes. A workflow that drafts an internal summary has a different risk profile from one that sends regulated advice or alters a customer account. Identify who notices a failure, how it is reversed and what evidence must be retained.

Use human review deliberately. It is not a temporary embarrassment: approval can be the correct permanent design where judgement, accountability or rare exceptions matter. The goal is to remove repeat work without pretending uncertainty has disappeared.

Measure the whole system

Time saved is useful but incomplete. Track correction effort, exception volume, completion time, quality and user trust. Include model, platform, monitoring and maintenance costs. An AI step that is cheap per request may still be expensive to evaluate and supervise.

Define a pilot threshold before building: for example, reduce median handling time while keeping correction rates below an agreed level across a representative set of cases. If the threshold is not met, the result is evidence—not a reason to move the goalposts.

Pick a bounded first move

A good pilot has a clear start, end, owner and fallback. It runs beside the existing process long enough to compare results. It produces logs that explain what happened without collecting unnecessary personal data.

The final decision should be build, buy, integrate, redesign the process without AI, or stop. Automation discovery earns its cost when it makes any of those conclusions defensible.