Test the Workflow Before You Build the Agent

Before building an AI agent, test whether the workflow problem is worth solving.

Pick one repetitive task and ask a few people who actually do it to walk you through the last time it caused delay or rework. With their permission, try a human-run prototype: manually prepare the useful output for one real task, without copying sensitive records. Set a one-week limit and agree who owns the trial.

Watch for a practical signal: does the person choose to use the result again, and does it reduce the specific friction they named? Stop if the task is rare, the existing workaround is good enough, or the trial creates new risk.

That first test may not need an agent at all. It should tell you whether there’s a customer problem worth building around.