Sell AI Training Data

What we look for in a supplier

The suppliers we work with repeatedly tend to share a set of traits that matter more than size.

• Native-speaker access. You can recruit speakers of the target language who are genuinely native, not fluent learners.
• Documented process. You have a written annotation guideline and a quality control step, even if it is simple.
• Honest capacity. You can say no. Suppliers who accept every project and then miss deadlines cost more than they earn.
• Consent handling. You collect and retain speaker consent in a form that survives a legal review.

What you need to tell us

A useful first message is short and specific. Vague introductions tend to stall because there is nothing concrete to match against a buyer requirement.

• Which languages you can produce, and at what dialect granularity.
• Your recording capability: studio, home-recording network, mobile collection, or field teams.
• Whether you can annotate, and to what depth: transcription, phoneme, intent, diarization.
• Realistic capacity per month, in hours.
• Where you are located, and which jurisdictions you can operate in.

How the relationship works

We bring a specification, you quote against it, and we agree the quality bar with the buyer before production starts. A pilot batch goes first. It protects you as much as the buyer, because it establishes that your output meets the bar before you commit a full run.

Payment terms are set per project and confirmed in writing before production begins.

What we will not ask you to do

We do not take on medical or clinical data, and we do not source recorded telephone calls. If that is your main capability, we are not the right partner.

We also do not ask suppliers to produce data without a defined use case and a documented consent chain. If a buyer wants data with unclear provenance, we turn it down rather than pass the problem downstream.

What a supplier agreement actually covers

The commercial terms are less mysterious than they look. Every supplier arrangement we sign comes down to five things, and it is worth knowing them before the first call.

None of these are unusual. They are the same terms any buyer of commissioned data needs in order to sell the result to their own customer.

• Scope. Exactly what is being produced: volume, language and dialect, recording conditions, annotation depth, and the delivery format.
• Quality bar. The measurable standard the output is checked against, agreed before production rather than argued about afterwards.
• Rate and schedule. A price per unit and a delivery date, both fixed in writing. Changes to either are agreed, not assumed.
• Rights and consent. What you are granting, what you are keeping, and what evidence of contributor consent has to be handed over with the data.
• Remedies. What happens if a batch fails the quality check. The answer should be that it is fixed at no cost to the buyer, and it should be written down.

Where suppliers lose money

The failures we see are rarely about capability. They are about the commercial details that were never settled.

The most expensive one is producing before the specification is frozen. If the buyer changes the language mix or the annotation depth halfway through a run, the supplier carries the rework. That is why we do not start production until the specification is signed off on both sides.

The second is agreeing to a quality bar in conversation instead of in writing. A buyer who says the data must be clean has not said anything checkable. If the standard is not measurable, every delivery becomes a negotiation.

• Working without a written specification. Rework is almost always the supplier's cost.
• Quoting a date you cannot hold. Missing one deadline costs more future work than the project was worth.
• Accepting a scope you cannot staff, then subcontracting it quietly. The quality drops and the buyer sees it.
• Handing over data without consent records. The buyer cannot use it, so nobody gets paid.
• Pricing a one-off as if it were repeat work. First runs carry setup cost that later runs do not.

How to price what you have

The market quotes per hour, but the rate is driven by things that have nothing to do with duration. Understanding which of them apply to you tells you what to expect before you approach anyone.

We quote per project rather than per hour, because two recordings of the same length can be worth very different amounts. The list below is roughly the order in which the factors move the number.

• Rarity of the language or domain. A well-documented low-resource language commands more than another set of English read speech, for the simple reason that there is less of it.
• Consent documentation. Clean, specific, transferable consent is worth more than the audio it covers. It is also the thing most suppliers underestimate.
• Existing metadata. If the recordings already carry accurate transcription, speaker demographics and channel information, the buyer saves a production step.
• Volume and consistency. A large set recorded under one protocol is worth more per hour than the same total hours assembled from mismatched sessions.
• Exclusivity, if you are willing to grant it. Exclusive rights are rare in this market and priced accordingly.
• What lowers a rate: unclear provenance, missing or generic consent, unusual file formats, small volumes, and inconsistent recording conditions.