How should I scope a pilot for this workflow?

Scoping starts with the process you want your agent to handle, not with a data volume. Name the two systems that disagree, the roles involved, the actions the agent may take, and the outcomes you consider correct, including escalation.

From that, a pilot can fix a small set of decisions: - **Domain and record types**: for example supplier invoices against receipts, or claim payments against reserves. - **Cause mix**: which root causes to cover, and how many overlap. - **Policy shape**: tolerances, approval roles and correction methods to include. - **Form**: dataset, task pack, simulated environment, or a combination. - **Format**: JSONL, Parquet, or tool definitions for your framework. - **Use**: training, evaluation, or both, since that affects how labels are held out.

Limits apply. The output is generated, so it will not reproduce your internal systems, and it should not be treated as a source of real business records. Scope, delivery terms and licence terms are agreed with MirrorFirm for each engagement; this page does not set them.

A sensible first step is a small sample covering a few causes, reviewed against your own judgement of what a realistic case looks like. Use what you learn to tighten the scope before committing to anything larger.