For owners who want lab money
Sell Data to AI Labs
Most companies that sell data to AI labs never sign with the lab itself. They license a copy of their old work records to one of the specialist suppliers that gather training material, and the supplier delivers it to the labs that build the models. Your company keeps ownership the whole time.
This page explains what an AI lab is, how labs really get private company records, what they look for in yours, and why an agent who brings you several offers will usually get a better result than an email to a lab's inbox. When you are ready, the estimate on this page takes two minutes and asks for no files.
- Several supplier offers
- Nothing paid by you
- Your exclusions in writing
The basics
What an AI lab is, and what makes one a frontier lab
An AI lab is a company that builds its own AI models from the ground up. That sets it apart from the thousands of firms that build apps on top of a model someone else made. Training a model means feeding it a huge amount of writing, conversation, code and other material so it can learn patterns from them.
A frontier lab is one of the few companies working on the largest and most capable general models. Names people often give are OpenAI, Anthropic, Google DeepMind and Meta. These are the companies most people mean when they ask how to sell data to frontier labs. DillisData is not affiliated with any of them.
Frontier labs spend on a scale that is hard to picture. For 2026, Amazon, Microsoft, Alphabet and Meta have set combined AI budgets of roughly $725B, according to the Financial Times. Only a small share of that goes to data, but even a small share of that sum is a lot of money looking for good material to learn from.
The supply chain
How AI labs actually get company records
Labs use several routes at once. Only some of them are open to a working company with a few dozen to a few hundred staff.
| Route | How it works | Open to a mid-size company? |
|---|---|---|
| Crawling the public web | Software copies pages anyone can read, such as articles, forums and public sites. | No. Your private email, chat and files are not on the open web, which is exactly why they are worth paying for. |
| Large direct licenses | A lab signs with an owner of a giant body of text, such as a big online forum or a news group. | Rarely. These deals are built for owners of millions of public posts or articles. |
| Lab partnership programs | A lab invites outside groups to offer data. OpenAI announced one in 2023 aimed at large data sets that are hard to find online. | Possible on paper, but these programs ask for very large volumes and offer no help with pricing or terms. |
| Specialist data suppliers | A supplier licenses records from many companies, removes personal details, packages them by type and delivers them to labs. | Yes. This is the main path for established companies, and the one we work through. |
| Court auctions of closed companies | When a company is wound down in bankruptcy, its records can be offered to bidders under court supervision. | Only if your company is closing. A healthy company can license on its own terms instead. |
Seen from the lab's side
What AI labs want from a business
When OpenAI opened its Data Partnerships program in November 2023, it said it was most interested in data that expresses human intention, such as long-form writing and conversations. Company records are full of exactly that. Here is what makes one archive stand out from another.
Many years in one place
A lab learns more from ten years of one company's tickets than from a few months. Long history shows how work changes, how people get better at it, and how problems repeat.
People working something out
The best material is back and forth: a customer asks, a coworker checks, a manager decides. A finished report is useful. The thread that led to it is often worth more.
Depth in one kind of work
Labs want their models to be good at real jobs, such as quoting a project, handling a claim or fixing an outage. A firm that has done one kind of work for years holds rare, specific examples.
A clean right to license
Before paying, a supplier needs to know the records are yours and that nothing in your contracts blocks the deal. Clear ownership makes your archive easier to buy.
Records that export cleanly
Material that comes out of a standard system in a standard format, such as a mailbox export or a help desk download, is cheaper for a buyer to prepare. That can show up in the price.
What qualifies
Records labs pay for, and the systems they sit in
One or two of these, covering a few years, is often enough to draw offers. You choose which ones go into a deal.
| Record type | Common systems | What a lab's model learns from it |
|---|---|---|
| Email threads | Microsoft 365 and Exchange, Google Workspace, saved PST or mbox files | How a request turns into a plan, a price and a signed job over many messages |
| Team chat | Slack, Microsoft Teams, Google Chat | Quick problem solving between coworkers, including mistakes and corrections |
| Help desk tickets | Zendesk, Freshdesk, Jira Service Management, ServiceNow | A problem, each attempt to fix it, and the answer that finally worked |
| Sales and account notes | Salesforce, HubSpot, Pipedrive | How a lead is qualified, followed up and won or lost |
| Documents and proposals | SharePoint, OneDrive, Google Drive, Box, Dropbox | Careful writing for a real reader, with real numbers and terms |
| How-to guides and wikis | Confluence, Notion, Guru | Step by step instructions a new hire would follow |
| Project and job records | Jira, Asana, field service and ERP systems | The order in which real work gets planned and finished |
| Source code | GitHub, GitLab, Bitbucket | How a team writes, reviews and fixes software over time |
Going direct vs. using an agent
How to sell data to AI labs without cold-emailing them
It is natural to think the best price comes from going straight to the top. In practice, a frontier lab's data team is set up to buy very large collections, often from a small number of suppliers who have already done the gathering, cleaning and legal checks. A single company of 60 people is a small file to them, and most will point you to a supplier or not reply at all.
Suppliers are the real buyers for most companies. They compete with each other to fill the gaps the labs tell them about, and they publish payouts that run from about $10K to more than $4M per company. They also differ a lot. One may want only support tickets. Another may pay well for email but ask for an exclusive license. A third may pay quickly but cap the price low.
That is the gap we fill. We take a short profile of your company, never the records themselves, to several suppliers at once. We put their offers next to each other, push on price, and point out any clause that would stop you from licensing again. The buyer covers our fee as a separate amount, so your share stays whole. The bigger picture of this market is in our main guide to licensing company records to AI firms, and the wider list of sources labs use is in where AI companies get their training data.
See what labs' suppliers might pay you
Four questions about your headcount, age, industry and records. You get a cautious range in about two minutes, and nothing about your files leaves your hands.
What stays out
What to keep out of a deal with an AI lab supplier
A careful license names what is excluded, so it cannot drift in later. Lawyers at Frankfurt Kurnit Klein and Selz laid out the main legal risks in an October 2026 article. None of this is legal advice. It is general background, and your own lawyer should read any license first.
Other companies' secrets
Email and chat often hold a partner's or vendor's confidential details, shared under an NDA. Folders and threads tied to those agreements can be left out by name.
Customer personal data
Payment details, account files and order histories about your customers stay out. Labs' suppliers do not need them, and they carry the most risk.
Promises in old privacy policies
If your privacy policy or handbook never mentioned AI training, check it first. The FTC has warned that quietly applying looser rules to data you already collected can be unfair or deceptive.
California staff data
Under the CCPA, handing over California employees' personal data for money can count as a sale unless it meets the law's de-identification standard. That shapes how names and details are stripped.
Talk about pay and working conditions
Employees discussing pay, hours or conditions together can be protected under the National Labor Relations Act. Channels used for that kind of talk are good candidates to exclude.
Anything you would need back
Once a model has trained on a record, it cannot practically be removed. If a file still gives you an edge today, keep it out.
The process
How a lab ends up with your records, step by step
Expect to spend roughly one hour of your own time. The supplier does the copying and cleanup, and we handle the search and the paperwork.
Answer four questions
Tell us your staff size, years in business, industry and the record types you hold. No logins, no uploads.
2 minutesWe approach the suppliers
We share your profile with the suppliers most likely to want it and gather what each one would pay and on what terms.
1 to 2 weeksYou pick an offer and sign
We go over every offer with you in a single call, including its limits and exclusions. If none feels right, you owe nothing and walk away.
One callThe supplier copies, cleans and pays
The buyer takes its copy straight from your systems, removes names and personal details, and pays you within the window in its terms. Buyers' terms set that window at 7 to 60 days.
Per contractProof labs are buying
Public signs that AI labs pay for company records
These are reported facts, not promises about your deal. They show how much the largest players care about this kind of material.
Google won the bankruptcy auction for Spirit Airlines' internal records, including about 100 million employee emails and 500 million Teams messages. Customer bookings and payments were excluded. The sale is pending court approval as of October 2026.
Axios, Aug 2026A court-appointed privacy ombudsman recommended approval of the Spirit sale. Spirit hired an outside firm to strip personal details from the email before any of it moves to the buyer.
The Next Web, Oct 2026The price Meta paid in 2025 for a 49% stake in a training data supplier. Labs lean on suppliers that heavily.
TechCrunch, Jun 2025Speaking at a research conference in late 2024, OpenAI co-founder Ilya Sutskever told the room the field had hit "peak data", with no new supply of public text coming.
The Verge, Dec 2024Timing
Why selling data to labs is a now question
Nobody knows how long labs will keep paying for old business records. Here is why we suggest owners decide soon.
Gaps get filled one at a time
Suppliers pay the most for a kind of work they cannot yet offer the labs. Once they have a few firms like yours, the next one is easier to pass on.
Labs are trying other methods
Labs keep testing new ways to train, including data they generate themselves. If those methods improve, the need for real company records could shrink, and no one can say when.
Your archive is idle
Old mailboxes and closed tickets cost something to store each month and return nothing. A non-exclusive license changes that without deleting anything.
Our monthly intake is capped
Each month we accept only as many new companies as we can shop with care. Once those places are taken, the next owner waits for the following month.
Questions
Questions owners ask
How do I sell data to frontier labs directly?
You can write to a lab's data or partnerships team, and some run public programs. Expect them to ask for very large volumes and to handle pricing on their terms. For most companies under a few thousand staff, the practical route is a supplier that already sells to those labs.
Do AI labs buy from small or mid-size companies?
Mostly through suppliers. Most buyers start at around 20 employees and three or more years in business. Below that, offers are fewer and smaller, and the estimate will tell you quickly where you stand.
Can we choose which labs get our records?
Your contract can spell out which companies are allowed to use the copy and whether the supplier may hand it on. We point out resale and sublicense terms in every draft so you can limit them before you sign.
Is selling data to labs the same as selling customer data?
No. A deal like this covers your company's own work records. Customer personal data stays out, and whatever goes in is stripped of names and personal details by the buyer.
Will a lab's model repeat our private emails word for word?
The buyer removes personal details before training, and it will usually let you check samples first. No cleanup method is perfect, which is why sensitive folders should be excluded outright.
Does DillisData have a deal with any AI lab?
No. DillisData is not affiliated with OpenAI, Anthropic, Google, Meta or any other lab. We work for you and take your profile to the suppliers that serve the labs.
If we sign with one supplier, can a second one license our records too?
Yes, if the first license is non-exclusive. That is the default we push for. An exclusive deal should pay noticeably more, and we flag it clearly when one is offered.
Start here
Find out what AI lab suppliers might offer you
Two minutes, four questions and a number rounded down on purpose. If it looks worth pursuing, we take it to market. If not, you have shared nothing.
No cost to your company. No access to your systems. Limited spots each month.