Prototype the Agent’s Recovery Path First

An AI agent’s first useful prototype might be the moment it doesn’t know what to do.

Take an agent that drafts meeting follow-ups from notes. If an owner or deadline is unclear, it should flag the gap and ask for approval—not quietly invent a commitment. Prototype that pause-and-clarify path before automating the whole workflow.

For one week, test it with a few people who already handle this task. Keep the scope to drafting and flagging uncertainties. Look for whether they can review the draft quickly and spot what needs correcting. Stop if the agent obscures uncertainty or creates more review work than it saves.

The first question isn’t just “Can it complete the task?” It’s “Can people recover control when it can’t?”