Selling to Labs
Chain of title, in practice
The first call is about the data. The second call is with counsel, and it is about paper. Most sellers lose the deal here, not because their rights are bad but because they cannot show what their rights are. This is the list of documents that gets asked for, what each one has to say, and the four places the chain usually breaks.
What "chain of title" means for a dataset
A buyer needs a continuous, documented line from the person or process that created each record to your right to license it for model training. Continuous means no gaps. Documented means an artefact, not a recollection. Model training means that specific use — a right to "use, reproduce and display" a contribution does not obviously include training a model on it, and counsel will not assume it does.
The test a buyer applies is not "is this lawful?" It is: can we describe the provenance of this corpus in a public summary, and would this documentation hold up if the description were challenged? That is a higher bar than internal comfort, and it is the bar that has been rising since training-data summary obligations arrived.
The four documents
1 · Source agreements
Whatever gave you the data in the first place: a data-sharing agreement with a partner institution, a purchase from an earlier holder, a research collaboration, a corpus acquired with a business. Counsel reads these for scope of use, sublicensing rights, field-of-use restrictions, and survival after termination. The recurring failure is an agreement that grants you broad internal use and is silent on onward licensing — which reads, to a buyer, as no right at all.
Have ready: the executed agreement, every amendment, and a short note on which portion of the corpus each one covers. If different tranches came in under different agreements, say so up front. Discovering a second agreement mid-diligence costs more credibility than the second agreement costs in restrictions.
2 · Employment and contractor IP assignment
If people you paid produced the records — annotators, clinicians writing notes, operators recording trajectories, engineers building the harness — the buyer needs assurance the output belongs to the company. For employees this is usually handled by the employment contract, and usually adequate. For contractors it is the most common break in the entire chain: contractor work frequently defaults to the contractor owning the output unless assigned in writing.
Have ready: the standard clause, evidence that everyone who contributed signed a version of it, and the date the clause changed if it did. A corpus collected over five years often has two or three generations of contract behind it, and the earliest generation is the one that says the least.
3 · The terms-of-service clause, if the data came from users
This is the document buyers scrutinise hardest, because it is the one most often stretched. Questions asked, in order:
- Does the clause grant a licence broad enough to cover training, or only to operate the service?
- Was it in force when the records in question were created? Terms updated in 2025 do not retroactively cover data collected in 2021 in most readings.
- How was acceptance obtained, and can you evidence it per user?
- Is there a right to sublicense, without which you can grant the buyer nothing?
- Were users given a way to opt out, and were those opt-outs actually applied to this export?
A ToS that covers "improving our services" is the classic ambiguous case. Some buyers accept it with indemnity; increasingly, buyers who must publish a provenance summary do not. If your position rests on this clause alone, know that before you are asked, and price accordingly rather than discovering it in the room.
4 · Supplier flow-downs
Any part of the corpus obtained through a vendor — a scraping supplier, an annotation shop, a data broker, a transcription service — carries that vendor's terms into your position. You cannot grant more than you received. Counsel will ask for the vendor contract, the vendor's own representation about where they got it, and whether their warranties flow down to your buyer.
A vendor who will not represent their sources is a defect in your asset, not theirs. It is worth renegotiating that representation before you go to market; afterwards you are negotiating with a buyer watching.
The four places the chain breaks
| Break | How it shows up | Fix before market |
|---|---|---|
| Contractor gap | Records produced by people with no signed assignment, usually the earliest cohort. | Retroactive assignment where reachable; otherwise carve the affected tranche out and sell the rest. |
| Silent sublicence | Source agreement grants use but not onward licensing. | Go back to the source for a written extension. Slow, and the only real fix. |
| Temporal mismatch | Current terms cover training; the records predate them. | Segment the corpus by the terms in force at creation and disclose the split. |
| Third-party content inside your records | User uploads, quoted articles, embedded images, pasted code with its own licence. | Detect and either remove or document separately. This is the one sellers most often have not looked for. |
The manifest that ends the argument
The strongest position is not a folder of contracts, it is a per-record annotation. Every record carries its source, the instrument that grants the rights, the consent basis, and any reservation that applies. Then the corpus can be filtered rather than argued about, and a buyer's question — "can you give us only the records covered by the 2023 agreement?" — becomes a query instead of a project.
{
"record_id": "trg-00184213",
"source": "partner:northside-clinic",
"instrument": "DSA-2023-11 §4.2",
"consent_basis": "explicit-research-consent",
"created": "2024-03-11",
"terms_in_force": "tos-v4",
"reservations": [],
"third_party_content": false
}
Sellers who arrive with this are through diligence in days. Sellers who arrive
with a claim and a promise to look into it spend weeks, and roughly half of
those deals do not restart. We are shipping these annotations as operators for
the curation pipelines people already run, so the manifest is a pipeline stage
rather than a spreadsheet — that work is queued as provenance-ops.
The second call, in the order it happens
- Where did the data come from, and under what instrument? (Source agreements.)
- Who created it, and does that output belong to you? (IP assignment.)
- If it came from users, what did they agree to and when? (Terms, plus acceptance evidence.)
- What did your suppliers promise you? (Flow-downs.)
- What is in the records that you did not create? (Third-party content.)
- Can you produce a filtered subset by any of the above? (The manifest.)
Six questions. A seller who can answer all six with an artefact has just scored well on the heaviest dimension in the rubric — 24 of 100, per what your data is worth — and has removed the most common reason a deal dies.
This post describes what buyers ask for and how sellers commonly fail it. It is not legal advice, and the treatment of any specific clause depends on jurisdiction and on facts we do not have. Get your own counsel to look at the four documents above before you rely on them.
Read next: De-identification that survives counsel — detection rates, free-text sweeps, held-out audit, and what evidence to retain.