Know what AI buyers will ask before they make an offer.

Buyers evaluate evidence before they discuss price.

A company that wants to sell data to AI companies will face a predictable sequence of requests. Buyers do not start with enthusiasm for the asset. They start with questions about what it is, where it came from, and whether the seller has the authority to license it.

Sellers who prepare answers early move more clearly. Sellers who improvise often discover gaps while a buyer is waiting.

This article lists the requests we see most often and how to prepare for them. It is a guide to preparation, not a promise of any outcome. Every dataset, and every buyer, is different.

The data card or manifest.

The data card is usually the first document requested. It describes the asset in plain terms. It lets a buyer decide whether a deeper review is worthwhile.

A useful data card includes:

• Content type, such as text, images, audio, video, code, or structured records.
• Source systems and how the material was collected.
• Date range, languages, and geographic coverage.
• Record counts and total volume, with usable content separated from gross file counts.
• Fields, schema, and a short data dictionary.
• Known gaps, limitations, and excluded categories.
• A version identifier.

The manifest is the file-level counterpart. It lists items or groups of items with identifiers, formats, sizes, and dates. Buyers use it to check that what they receive matches what was described.

Keep the card and the manifest consistent. If a category is removed after review, update both before anything goes out.

The sample.

Buyers ask for a sample so they can test quality for themselves. They may run it through their own evaluation pipeline. They will compare it to the data card.

A sample should represent the proposed scope. It should not be the best records alone. Select across sources, dates, and formats. Describe known weaknesses openly.

Before sharing a sample, decide:

• Which records are cleared for disclosure.
• Whether metadata alone can answer the buyer’s first questions.
• Who may access the sample, and for how long.
• What the buyer may and may not do with it during evaluation.
• Whether an evaluation agreement should be signed first.

Raw records are not needed to explain every aspect of an asset. Start with the least disclosure that supports a real review.

Rights and provenance evidence.

This is often the longest part of the process. Buyers want to see that the seller owns the data or holds the right to license it. They also want to see how it was assembled.

Expect requests for:

• The chain of title: who created or collected the material, and how rights reached the current owner.
• Employee and contractor assignments, where the material was created by people.
• Terms of service, contributor terms, or user agreements in force at the time of collection.
• Third-party licenses that cover part of the content.
• Any earlier license, exclusive or otherwise, that may limit the new one.
• Acquisition or asset-transfer documents, where the data was inherited.
• Records of how the data was collected, including any scraping or purchased sources.

Missing documents should be labeled as missing. Do not reconstruct certainty from memory. If part of the dataset cannot be supported, consider excluding it.

The consent basis.

Where data relates to people, buyers ask what permission supported its collection and whether that permission covers the proposed use.

Prepare a short statement covering:

• What notice or consent users received, and when.
• Whether the language addresses disclosure to third parties or AI training.
• How withdrawal or deletion requests were handled.
• Whether any records came from sources with narrower purposes.

Older terms often did not anticipate AI use. That does not always rule out a license, but it needs review. Do not assume that broad wording covers a new purpose. Your counsel should assess it.

PII handling.

Buyers will ask whether the data contains personal information and what was done about it. A vague answer slows everything down.

Document the following:

• Which fields or content types may contain personal data.
• Whether identifiers were removed, masked, or pseudonymized, and by what method.
• How you tested the result, and what residual risk remains.
• Whether sensitive categories are present, such as health, financial, or children’s data.
• Where confidential third-party content might sit inside free text, attachments, or recordings.
• Who handled the data during processing.

Be careful with the word “anonymized.” If re-identification is still possible, say so. Precision here protects the seller as much as the buyer.

Dedup and quality statistics.

Buyers pay attention to what is usable. A large file count means little if much of it is duplicated or empty.

Prepare figures you can support, and state how you measured them:

• Exact and near-duplicate rates, and the method used.
• Share of empty, corrupted, or unreadable records.
• Language identification results, with the confidence of the method.
• Length or size distributions.
• Labeling or annotation quality, if labels exist.
• Any filtering already applied, such as spam or boilerplate removal.

Do not publish estimates as if they were measurements. Distinguish reviewed samples from extrapolations.

Format and delivery.

Buyers need to know how the data will actually arrive and whether it can be used on receipt.

Expect questions about:

• File formats, encodings, and compression.
• Schema documentation and metadata files.
• Delivery method, such as transfer to a secure location or controlled access.
• Integrity checks, such as hashes for each file or batch.
• Delivery in stages, with acceptance testing at each stage.
• Whether the buyer can retain a copy or only access it in a controlled environment.

Settle how each delivered version is identified. A buyer should be able to say exactly which version they received.

Exclusivity questions.

Exclusivity is a common early question. Buyers may ask for it. Sellers should decide what they can offer before the conversation starts.

Define the terms:

• Which content the exclusivity would cover.
• Which field of use, such as training a specific type of model.
• Which territory and what duration.
• Whether you keep the right to use the data yourself.
• Whether any earlier license already limits what you can grant.

Exclusivity should describe exactly what is covered. A broad promise made without checking prior commitments can create exposure.

Refresh and updates.

If the data keeps growing, buyers ask whether new material will be available and on what terms.

Consider in advance:

• Whether the dataset is a static export or an ongoing feed.
• How often new material could be delivered.
• Whether refreshes would carry the same rights and documentation.
• How deletion requests would be applied to delivered data.
• Whether future material is priced and licensed separately.

A static historical export and a continuing refresh are different commercial offers. Treat them separately.

Warranties and indemnities.

Near the end of evaluation, buyers usually ask for contractual assurances. These are negotiated terms, and they carry real risk for the seller.

Commonly requested items include:

• Ownership or authority to license the data.
• Compliance with applicable law in collection and disclosure.
• No known infringement of third-party rights.
• Accuracy of the data card and the sample.
• An indemnity for certain third-party claims.
• Limits on liability, which the parties negotiate.

Only give a warranty you can support from your evidence. If the record has gaps, narrow the statement or exclude the material. Your counsel should review each clause.

A preparation checklist.

Before approaching any buyer, confirm that you have:

• A versioned data card and a matching manifest.
• An approved sample and a clear list of what may be disclosed.
• A source register with chain-of-title documents.
• Assignments, contributor terms, and third-party license records.
• A consent and notice summary for personal data.
• A PII handling description, with testing notes.
• Dedup and quality figures, with methods stated.
• Format, delivery, and integrity-check plans.
• A position on exclusivity and refresh.
• A list of warranties you can support, reviewed by counsel.
• A named person authorized to approve disclosure and sign.

Not every item will be complete at the start. The checklist shows where work remains. Some material may be dropped from scope, and that can be the right decision.

Where to go next.

Preparation differs by asset type. For inherited material, see acquired archives. For other topics, browse more insights. To see how we structure the work, read how it works.

If you hold data that AI buyers may value, start with a confidential seller brief. It lets us understand the asset and its constraints before any buyer is contacted.