How should a team scope a first intake dataset or pilot?

Scope a first step narrowly. A useful pilot covers one intake path, one policy set and a limited range of exceptions, so that quality can be inspected before anything larger is considered.

Start by naming the case type, for instance claim lodgement for a general insurer or new engagement requests for a professional services firm. Then name the systems your agent will touch, the tool actions it will call and the outcomes you want graded. These decisions shape the schema more than any volume figure.

Next, choose the form. A dataset in JSONL or Parquet suits training and offline checks. A task pack adds expected outcomes for grading. A simulated environment adds live tool calls, state changes and reset. Each takes more effort to specify, so pick the lightest form that answers your question.

Decide the exception mix deliberately. Ask for clean cases to establish a baseline, then add the exceptions that worry you most: duplicates, conflicting fields, missing documents, ambiguous routing.

Licensing terms, volumes and timing are agreed with MirrorFirm directly for your use; none are fixed in this article. Say early whether you plan to train, evaluate or both, and whether outputs will be shared outside your organisation, because that affects the licence you need.

The material is generated, so include a review of grounding and disclosure risk in the pilot plan.