What is a useful first scope for an agent correction pilot?

Start with one correction class that your agent is expected to handle, such as moving an expense between cost centres. Define the trigger, required evidence, permitted actions and point at which the task is complete. A pilot should cover a whole decision even if it touches only a small part of the wider business.

Specify the agent's role explicitly. An assistant that prepares proposals has a different success condition from an agent allowed to post approved entries. The environment should expose enough state to assess that role without implying broader authority.

A useful initial request to MirrorFirm can describe the current agent interface, the records it can inspect and the failure you most need to understand. Ask for a generated sample with an acceptance explanation that ties each task to that concern. If your uncertainty is evidence matching, adding complex payment execution may make results harder to interpret.

The pilot brief should also name excluded activities, such as changing supplier details or issuing refunds. These boundaries keep adjacent capabilities from becoming accidental requirements. Expand the scope when results identify a specific missing dependency or behaviour, rather than assuming that more domains automatically produce a more informative evaluation.