What should we specify when asking for a generated intake dataset or task pack?

Settle these points before you ask for a sample, because each one changes what is generated and how it can be scored.

- **Domain and case type:** for example insurance claim notices, utility connection requests or professional-services engagement requests. - **Channels and formats:** web form, email, portal, transcribed phone note, scanned attachment text, and how messy each should be. - **Completeness policy:** which mandatory items apply, whether versions change over time, and whether alternatives are allowed. - **Routing table:** number of queues, the criteria for each and any tie-break rules. - **Defect mix:** missing items, contradictory documents, duplicates, amendments after lodgement and requests that fall outside policy. - **Relationships:** links between cases, parties and prior cases that the agent must discover rather than be handed. - **Answer-key depth:** final outcome only, or also required checks, tool actions and handover note content. - **Handover and audit fields:** what a downstream role needs to see, and what the event log must record.

If you cannot decide on all of them, pick the two or three that matter most to your evaluation and let the first sample show you the rest.