AI Data Brokerage

What a data broker actually does

The training data market has a structural problem. On one side are AI teams that know exactly what they need (400 hours of Levantine Arabic call-center audio, recorded on headset, with intent labels), but have no idea who can produce it. On the other side are studios, annotation firms, and field teams with the capability but no way to find that buyer.

A broker closes that gap. We take the specification, find the producer who can actually meet it, define the quality bar with both sides, and stay accountable through delivery.

That is different from a stock marketplace, where listings sit on a shelf waiting to be found, and from a catalogue house, where you buy whatever happens to be in stock. Brokerage starts from your requirement, not from existing inventory.

Brokerage against the alternatives

A marketplace is the right tool when something on the shelf already fits your requirement. This table is about what happens when nothing does.

What we do not do

Being clear about the boundaries saves everyone time.

• We do not pass on a dataset whose licensing chain we cannot evidence. Licensing existing data and trading it on is part of the business, but only where the rights actually reach you.
• We do not take on medical or clinical data, and we do not touch recorded telephone calls. Both carry consent and privacy exposure that we are not set up to carry.
• We do not publish prices on this site. Every quote depends on language, hours, recording conditions, and annotation depth. A number without those inputs would be fiction.

How a project runs

1. Written specification

   Volume, distinct contributor count, language and dialect, recording environment, annotation type and delivery format all get written down before anything is collected. We will not quote against a vague request. It always ends in a dispute about what was promised.

2. Matching

   We go to producers vetted against that specification, confirm they can hit it, and get a real cost and a real timeline. If nobody can, we say so rather than take the order and improvise.

3. Quality gate

   A pilot batch is produced and annotated first, and you review it before the full run begins. This is the single most effective way to avoid a rejected delivery: problems that would cost a full re-collection get caught after a few hours instead.

4. Delivery

   The dataset arrives with the documentation that makes it usable: contributor consent records, collection methodology, and the annotation guideline it was produced under.

Why buyers use a broker instead of going direct

The obvious argument against a broker is cost. You are adding a layer, and layers cost money. The less obvious argument for one is that the layer absorbs work you would otherwise do badly.

Finding a producer for an uncommon language takes weeks of cold outreach, and most of the people who respond cannot actually deliver. Vetting means checking whether they have done this language before, whether their annotators are native speakers, and whether their consent paperwork would survive a legal review.

Then there is the coordination cost. A single multilingual project can involve a dozen production teams across as many countries, each with different standards and different turnaround. Making those outputs land in one consistent format is real work, and it is work that has nothing to do with your model.

How brokers are paid, and what it means for you

The commercial model behind a broker is worth understanding, because it explains their behaviour when a project goes wrong.

There are four common models in this market. They create different incentives, and the differences show up at exactly the moments when a project is under pressure.

• Margin on production. The broker quotes you a price and pays the producer less. Transparent, common, and it means the broker's incentive is to keep production efficient.
• Commission from the producer. The broker is paid by the supplier. This is the model to be careful with, because the broker's interest is in the size of the order rather than in whether it fits you.
• Licensing share. On existing datasets, the broker takes a share of the licence fee. Common in this market, and usually the cleanest arrangement for the buyer.
• Retainer or project fee. The broker is paid for the sourcing work regardless of the outcome. Rare outside very large programmes, and it shifts the risk of a failed search onto you.

What to ask a broker before committing

These questions are not hostile. A broker with a straightforward model will answer all of them without hesitation, and the answers tell you which of the models above you are actually dealing with.

If any of them cannot be answered, that is itself information about how the project will go.

• Who is the producer, and can you tell me about them? A broker who will not name the producer is usually protecting a margin rather than a relationship.
• How are you paid? The answer should be a sentence, not a paragraph.
• What happens if the data fails the quality check? The answer should be that it is fixed at no cost to you.
• Do I receive the consent records directly? You should. Holding them behind the broker creates a dependency you cannot audit.
• What is your role after delivery? A broker who disappears at handover is leaving you with the consequences of their sourcing.
• Is there anything in this project you would turn down? A broker who never says no is not filtering, and you are paying them to filter.