Guides · For brokers and referrers

How to broker AI training data: what actually works

Finding data is the easy part. Here's how intake, quality checks, packaging and routing to AI labs actually work, from a team that does it every day.

Erik Newsham · · First published on X

Short version

Brokering AI training data means finding companies whose records a lab will pay for, checking that the data is real and that the owner can license it, packaging it so a buyer can judge it quickly, and getting the right package to the right lab. Finding data is the easy part. Earning a lab's trust is the part that takes months, and it's the part most new brokers skip.

This guide is the long version, based on what we do every day at Grepped, including the parts that go wrong.

Why everyone suddenly became an AI data broker

Sometime this autumn a lot of accounts on my timeline started calling themselves AI data brokers. The money in this market is real, so I understand why. But I've spent an unhealthy number of hours actually doing it, and a few of the things being repeated online are wrong. So here's how it works, start to finish, grubby bits included.

What AI labs actually want to buy

The labs have all trained on the same internet: the same crawl, the same GitHub, the same Reddit. What they're looking for now is the material that sits behind a locked door. A chip team's real design and verification work. Wet lab and clinical data. Video from a factory floor where people wear head cameras while they do the job. The strange internal workflows of a roofing contractor, a tax firm or a freight company.

That last group sounds odd until you realise that things which are intuitive to people are not intuitive to models. Nobody has written down how a freight company actually reschedules a missed pickup, so no model has learned it.

Demand moves quickly. On most of our calls this autumn, labs have asked about finance, legal, semiconductors, biodata, manufacturing, aerospace, very specific workflow data, and RL environments with proper graders. Very few have asked specifically for generic marketing agency text or a large dump of ordinary code. So if your pitch is "we've got ten years of emails, what's that worth", the honest answer is usually less than you'd hope.

That said, buyers still pay well for a business's internal Slack, Salesforce, QuickBooks and similar systems when they show real work across several tools. We cover what moves the price in what your company's data is worth.

Step 1: Get the data in the door with a proper intake form

Every company that comes to us fills in an inventory before we touch anything. The questions are dull on purpose:

  • What is the data, in plain words?
  • Roughly how much is there?
  • What format is it in?
  • Where did it come from, and is any of it scraped or someone else's?
  • Who owns it?

That one form removes about half the pipeline on its own, because some of the answers are, let's say, creative. In a single week we've had a 40GB zip labelled "HIGH QUALITY CONVERSATIONS" that turned out to be one man's entire WhatsApp history, a dataset of "real human emotion" that was 13 hours of a smart doorbell pointed at a neighbour's front path, and an earnest offer of voter rolls with full names and addresses.

If you're starting out, build the form first. It saves you from spending a week on data nobody can legally buy.

Step 2: Check what's actually in the folder

My favourite story in this industry: a broker took in a dataset and priced it more or less by the terabyte. Nine folders deep, under sensible-looking directory names, were roughly four thousand copies of Frozen 2, put there to make the folder look bigger. The broker never opened it. The lab did. I'd be amazed if that broker ever hears from that lab again.

So before anything moves, we check that the data matches what the owner told us on the form, that it isn't padded, that it really belongs to them, and that a buyer actually wants it this quarter. We also keep any number the owner gives us marked as stated by the owner until we've checked it ourselves. Labs notice that kind of care straight away, and it's a big part of why they keep coming back.

Step 3: Handle personal data carefully, and know when to walk away

Some data you simply decline. If the whole value of a dataset is who the people in it are, there's nothing left to sell once that's handled properly. And if an owner tells us something can't be shared, it doesn't get "mostly redacted". It stays with them.

Every package states plainly where it stands on personal data, so the buyer's legal team isn't guessing. Our security page covers how we access and store data once a deal is signed.

Step 4: Package it so a tired researcher can say yes

The person at a lab reading about your dataset has a hundred others in their inbox. Make it easy for them. Every package we send covers the same ground:

  • What it is, in detail.
  • What it's good for: post-training, evals, RL environments.
  • How big it is, and whether that number has been checked or only claimed.
  • The format and how it gets delivered.
  • Where it came from and who holds the rights.
  • Where it stands on personal data.
  • Its known limitations.

Listing the flaws feels wrong the first time. It's also the fastest way I've found to be believed about the good parts.

To give you an idea of what's in demand right now: RL environments for physical chip design, graded by real signoff tools. Labs read a description like that once and reply the same night. You can see how we present datasets on our marketplace, and builders can read real company data for RL environments.

Step 5: Route it to the right lab

This is the part you can't copy from a thread. We work with the large labs, the newer labs, and many smaller specialist labs you've probably never heard of, including a few that work in sensitive sectors and alongside government. Each wants something slightly different. One only cares about chip design. Another wants factory video and nothing else.

Routing means knowing which is which before you pitch, plus who actually signs, what their legal team will push back on, and roughly what they paid for the last thing like it. Labs tell us which industries they're hungry for, and when something lands in one of those sectors we bring it to them first.

For a lab, a good broker is one counterparty and one agreement instead of forty small sellers who each need their own NDA and their own review.

What a data deal looks like

Our average deal is now around $500,000 to $600,000 and rising, though plenty are smaller and the number depends heavily on the data. The licence can be one-off or ongoing. The lab sets the scope of use and it's written into the agreement. Every licence we arrange is perpetual, and whether a deal is exclusive or non-exclusive depends on the owner and the listing. Both are possible.

The owner keeps ownership throughout. They license a copy and carry on running their business on the same systems. Our guide to selling your company's data to AI companies walks through the process from the owner's side.

How referrers get paid

If you introduce a company to us and a deal closes, you earn 10% of it, capped at $100,000 per deal. Renewals pay you again. You can cold email, run ads and use our name, and US and non-US companies of any size count. The terms are written down, and every partner gets them on paper before they start.

If you know companies with data, or you work with dozens of them through private equity, venture or an industry network, apply to our partner program.

Should you broker AI data yourself?

Maybe, at some point. Just know what's involved: building the intake form and learning which answers aren't true, digging through terabytes looking for Frozen 2, setting up data handling that a lab's legal team will accept, keeping up with what each lab wants this quarter rather than last, and then waiting months for a lab to trust you enough to send money.

Trust is the product here. It takes a long time to build and one bad dataset can end it. The other route is to work with someone who has already done the slow part and learn how it fits together from the inside.

Know a company with valuable data?

Introduce them to us and earn 10% of the deal, capped at $100,000.

Become a partner