Suggested preparation effort: 5–15 internal person-days for a modest pilot; assumption only. Agree the task and measurement before transforming the sample. A buyer should explain why each requested field, label and delivery feature is necessary.
For a retrieval pilot, a useful original acceptance plan might include a fixed query set, expected source references, a no-data baseline, the licensed-data condition, an agreed review method and error categories. Keep evaluation material separate from any allowed training material. Do not let the buyer redefine success after seeing results without recording the change.
Technical QA should cover item counts, stable IDs, parsing, excluded content, rights-cohort linkage, duplicate handling and version integrity. Preserve a release manifest and transformation log. For file delivery, understand that the recipient gains a copy; for APIs or controlled queries, test credentials, entitlements, rate limits and failure behaviour.
AWS documents assets, revisions and datasets as separate building blocks, illustrating why a delivery version needs an identity rather than simply a folder name. AWS — Data in AWS Data Exchange NIST's generative AI profile also supports documented governance of third-party data. NIST AI 600-1 — Generative Artificial Intelligence Profile, July 2024
Completion condition: both sides accept a written pilot result, including failures and limitations. If quality is insufficient, price one bounded remediation step before considering expansion. If task performance is adequate but no budget exists, do not quietly convert the test into free ongoing production access.