Sell Data to AI Labs

How lab demand differs

A product company buys data to fix a known model gap on a schedule. A research lab buys data to test whether an approach works at all, which makes the demand narrower, more experimental, and harder to forecast. One quarter's requirement may be a specific language pair; the next may be something nobody has collected before.

Labs also tend to buy in smaller volumes than their public reputation suggests. A frontier lab may want a few hundred hours of a rare language with expert annotation, not tens of thousands. What they pay for is specificity, and that is good news for small suppliers who can hit a narrow target.

What labs actually look for

Across the requests we see, lab-adjacent demand clusters in four areas.

• Low-resource languages and dialects, where any clean, well-documented collection is scarce.
• Domain expertise: data produced or annotated by people with professional knowledge, where the value is the judgment, not the file.
• Evaluation and preference data, used to test models rather than train them, which often means smaller volumes with very careful annotation.
• Unusual conditions (noisy environments, specific recording setups, code-switching, children's speech with proper guardian consent), where generic suppliers do not go.

Why cold outreach mostly fails

Sending an email to a lab's general contact address with an attached dataset almost never produces a purchase. Research and data teams are small, inbound volume is high, and an unsolicited dataset arrives without the one thing that would make it actionable. Data is bought against a specification, and a specification only exists when a buyer has a live requirement. A dataset that matches none cannot be sold at any volume, and there is no public list of what a lab is buying this quarter.

What works instead is being present when a requirement exists. That is what a broker relationship does: we hold the requirement, we know which suppliers can meet it, and we bring the specification to you. It is less direct than an email to a famous address, and it is how most actual supply moves.

What makes a lab-ready supplier

Three things get a supplier taken seriously: a documented rights chain, a written production process, and the ability to hold a specification exactly. Labs are precise customers, and a supplier who silently improves or simplifies a specification will fail acceptance even when the data is good.

On rights, be blunt with yourself before you are blunt with a buyer. An undocumented consent chain is the single most common reason suppliers are turned away in this market, ahead of quality and ahead of price. If you cannot show consent for every person recorded, and ownership of what you hold, the conversation ends at the first review.

Money follows acceptance, not delivery. A batch is paid when the buyer's QA passes it, and the QA criteria are written before production. Suppliers who understand that, and price their work to survive one round of re-work, stay in the market; suppliers who treat delivery as the finish line do not.

The exclusions are firm on this side of the market too: recorded telephone calls, medical or clinical data, and scraped personal data are declined, by us and by the buyers we work with. A lab's legal review will not accept them, so no intermediary can place them.

How to approach a lab without wasting the attempt

A lab procurement conversation is short and specific. These are the things that make the difference between being asked a follow-up question and being closed out.

None of this requires a relationship with the lab already. It requires being concrete about what you have and honest about what you do not.

• Lead with the specification, not the story. Language, volume, recording conditions, annotation depth, delivery format. In that order.
• Say what the rights position is in one sentence. Who consented, to what, and what evidence exists. If the answer is complicated, lead with that instead.
• Name the capability that is hard to replicate. Access to a specific population or a specific recording environment is the part a lab cannot easily build.
• Do not send a sample without metadata. A file with no provenance is not evidence, and it makes the rest of the offer look unserious.
• Ask what would have to change for it to fit. A no with a reason is more useful than silence, and it tells you what to bring next time.

How a lab pilot actually runs

A pilot with a lab is not a sample request. It is a test of whether your production process can be relied on, and it is scored on that basis rather than on how good the first batch looks.

Understanding what the pilot is measuring changes how you prepare for it. Most suppliers prepare the data and neglect the process evidence, which is the part being assessed.

• What they ask for. A defined subset produced to their specification, with the same documentation that would accompany a full delivery.
• What they check. Whether the specification was met exactly, whether the metadata matches the audio, and whether the consent records are complete and transferable.
• What they are really assessing. Whether a second batch would arrive to the same standard without them having to supervise it.
• Typical shape. A scoped pilot, a review point, and a decision on whether to scale. The decision is usually made on process reliability rather than on the data itself.
• What kills a pilot. Late delivery, a specification that was quietly reinterpreted, incomplete consent records, or metadata that does not match the files.
• What happens after. A passed pilot becomes a supply relationship with volume commitments. A failed one rarely gets a second attempt, which is why the process evidence matters more than the sample.