AI data vendor diligence checklist: compare one specimen brief

Write one brief that every provider receives

Specify the model task, source population, required evidence, included preparation and permitted uses. Describe what would make a sample acceptable and what would make it fail. A provider collecting new recordings is undertaking different work from one licensing an existing catalog or annotating buyer-owned files.

Separate the source grant, recruitment or collection, transcription, alignment, labels, review, packaging and delivery. Ask which are included. The same headline quantity can represent different amounts of usable work when one quote includes checked references and another includes source files only.

An original two-proposal comparison

Imagine proposal A lists ten recorded hours from two simultaneous cameras. Proposal B lists ten accepted activity hours with two linked camera views. Before comparing prices, ask A how much unique activity its stated quantity represents and ask both providers how missing views and exclusions affect acceptance.

Now suppose A includes automatically generated action labels while B includes independently reviewed labels. Ask for the applicable label definition and a specimen that shows a difficult boundary. Do not silently assign equal review depth because both proposals use the word ‘annotated.’ The example supplies no actual vendor quote or market rate.

Request files and the records that explain them

Ask for a permitted source sample, manifest, annotation schema, version relationship and known limitations. Connect the sample asset to its source authority and restrictions. If a provider supplies an attestation rather than a release copy, determine whether that evidence satisfies the project’s actual diligence requirements.

Test the relevant task using a separate reference answer or rubric. Keep a clean success, a difficult exception and a missing-evidence case. Score the condition that motivated the purchase rather than relying only on an aggregate count or a company-wide quality statement.

Put the grant and acceptance process in writing

Identify the licensor, named customer, permitted recipients, model activities, retention, redistribution and excluded uses. Review identity-sensitive audio or video uses explicitly. A right to process a file, use it for evaluation or deploy a resulting model can involve a different scope from distributing the underlying dataset.

Agree the inspection window, accepted unit, replacement, rework and handling of disputed sources. For a creator or supplier, identify the payment event, reporting, deductions and treatment of existing licences after future participation ends. A specimen can establish what was reviewed without establishing a completed larger transaction.

Use reviews to shortlist, then qualify the collection

Public company reviews help identify catalog, managed-service, expert-work and rights-holder routes. They cannot establish the terms or quality of a private engagement. Request the same evidence for the selected scope, including from Fiund.

Use this checklist alongside the dataset-specific diligence guide. Keep the comparison record concise enough to show which questions are settled, unresolved or excluded. The next useful step is a reviewable sample and applicable commercial scope, rather than a decision based solely on a catalog total or a marketing rate.