How should a buyer scope a first sample or pilot?

A useful first step is a small, specific sample, not a broad request. Three decisions define it.

**Domain and obligation type.** Choose one or two settings, such as utility bills, insurance premiums or freight invoices. Each brings its own terms, reference habits and exceptions. Starting narrow makes review quicker and feedback sharper.

**Task shape.** Decide whether you need training examples, preference pairs, a held-out evaluation set or an environment with tools. These need different record structures and different review effort, so say which matters most.

**Difficulty profile.** State which exceptions you care about: short and over payments, third-party payers, disputes, duplicates or reversals. A pilot can weight the sample toward them.

Limits to expect and plan for:

- Output is generated, synthetic material. It is not real business records, and it should not be treated as a stand-in for any specific organisation's data. - Grounding and disclosure risk should be inspected on the sample, by your team as well as ours. - Coverage of any one domain is scoped in the pilot, not assumed.

Ask for a sample that includes the policy files, the schema and a generation manifest, so you can run your own checks. Then agree what a good result looks like before reviewing, for example which checks must pass and how many cases you will read by hand. Terms, volumes and timing are agreed per engagement.