Pick a repeated task that someone can describe from start to finish. Record its trigger, inputs, decisions and completion condition. If those are unclear, automating it will usually make the ambiguity harder to see.
Look for a useful boundary
A first automation might collect form submissions, validate required fields and create a review item. It does not need to replace the entire business process. Choose an output a person can inspect.
Use AI selectively
A fixed rule is often better for structured decisions. A model can help with language-heavy work, but it should have a clear input contract and an appropriate review step. Consider the cost of a plausible but incorrect answer.
Plan retries and ownership
External systems fail or throttle requests. Use stable identifiers, bounded retries and a visible failure queue. Someone must own resolving items that cannot complete automatically.
Measure the pilot
Compare time spent, rework and exception rates before and after. Keep a manual fallback until the workflow is understood. Expand only when the result is useful to the people doing the work.