How should I scope an intake dataset or environment before starting?

Scoping starts with a narrow question, not a large catalogue. Decide what you want the material to do: exposure to business process text, supervised examples of next actions, preference data on notes and requests, or an interactive environment for agent evaluation.

Then choose the slice of intake that matters. Options include the domain (for example lending, claims, utilities connections or patient registration), the channels, the number of policy rules, the balance of duplicates and exceptions you want represented, and the tools an agent will see. Narrower slices are easier to inspect.

Next, decide the format: JSONL for case histories, Parquet for tables, JSON Schema for tools, and a seed-based environment if you need resets.

Licensing is a conversation about how you will use the output: training, evaluation only, internal research or product features, and whether you need to retain models trained on it. Terms, prices and timelines are agreed with MirrorFirm directly; this page does not state them.

Limits apply. The output is generated, so it does not contain real business records, and it should not be treated as proof of any organisation's actual practice. Inspect grounding and disclosure risk in a sample. A sensible first step is a small pilot with a clear inspection plan.