How is intake different from a generic form-filling or document-extraction benchmark?

**Q: How is this different from a generic form-filling or document-extraction benchmark?**

A: Extraction benchmarks ask whether a field was read correctly. Intake asks whether the right decision was made with the fields once they were read. The workflow combines four judgements that are normally tested separately: capture, completeness, matching and routing.

A task pack for this workflow can pose them together, so an error in one shows up as a consequence in the next. A mis-captured postcode, for instance, may send a utility connection request to the wrong regional queue even though every other field was right.

It also adds things a pure extraction set leaves out:

- a policy document the agent must apply, with versions and exceptions - existing customer or account records to match against, including near-duplicates - state that changes as actions are taken, so a case can be re-entered after a request for information - handover notes that a later role depends on

You can then score outcomes (final queue, case status, outstanding items) and process (which checks were run, in what order, and which tool actions were used).

Because the material is generated, ask in a sample how the policies and case patterns were grounded, and inspect them for realism. Generated records do not reproduce any real organisation's files, and they should not be treated as if they did.