Selling data to AI labs: buyer audits

By Eric Do Couto · Updated October 15, 2026

Disclaimer: The author is not a lawyer, accountant, or financial advisor. This article is for general information only and does not constitute legal, tax, or financial advice. Laws and market terms change, so consult a qualified professional before making any decision.

Selling data to AI labs comes down to one test on the buyer's side: does the dataset show real work, from request to outcome, that the lab cannot scrape and is free to use? Before any money moves, expect a lab to check seven things: workflow completeness, history, structure, the rights chain, anonymization, uniqueness and language. History and uniqueness depend on what you already hold. Rights and anonymization are the two you can still fix before you apply.

Why AI labs pay for workflow data

One of the people who built today's models says the public web is tapped out. At the NeurIPS conference in Vancouver in December 2024, OpenAI co-founder Ilya Sutskever told researchers, "We've achieved peak data and there'll be no more," adding, "There's only one internet."

The US Copyright Office's May 2025 report on generative AI training says model performance "depends heavily on the quality of the data used to train them." It cites research from major developers suggesting quality "may even be a more important consideration than quantity," and notes that scraped internet text often holds content "with limited or negative training value."

What labs are short of is records of work getting done. A 2026 Epoch AI explainer, built on interviews with 18 people at AI labs and training-environment startups, named enterprise workflows a major growth area, with tasks like "filing reports, or manipulating spreadsheets." Your SOPs, ticket queues, wikis and project histories record that work inside a real company. One founder told Epoch there has been "substantially more in-housing" as labs build their own data teams, so a dataset earns its fee by showing work a lab's staff cannot recreate.

How an AI lab audits a dataset

OpenAI's data partnerships page, posted in November 2023, asks for "large-scale datasets that reflect human society and that are not already easily accessible online to the public today." It singles out "data that expresses human intention (e.g., long-form writing or conversations rather than disconnected snippets)." After its February 2024 licensing deal with Reddit, Google described the access as "real-time, structured, unique content." Read together, those statements point to seven checks.

Language demand varies by buyer; OpenAI asks for such data "across any language, topic, and format." Volume matters, but given the Copyright Office's point about quality, a large archive of fragments can lose to a smaller one where every record is complete.

The rights chain

Provenance now has a visible price. In July 2026 a US federal judge approved a $1.5 billion settlement under which Anthropic will pay authors about $3,000 per book after training its chatbot on pirated copies. The earlier ruling in that case found the training itself lawful and the acquisition of millions of books from pirate websites wrongful. How the data was obtained was the costly part.

Expect a buyer to ask who created each record, under which contract, and whether a client or other third party has a claim on it. Contact centres and agencies often hold data on behalf of clients, and that data may be licensable only with the client's permission. The AI training data licensing guide has a rights checklist.

Anonymization

The same OpenAI page says the company is "not seeking datasets with sensitive or personal information, or information that belongs to a third party." Free text is the hard part: a customer's phone number typed into a ticket body survives a scrub that only clears the phone field. Canadian companies should test the result against the serious possibility standard explained in PIPEDA and anonymization and, for Quebec residents' data, the stricter Law 25 anonymization rules. Ticket-heavy archives have their own support ticket prep checklist.

What gets a dataset rejected, and how to avoid it

When selling data to AI labs stalls, it stalls on one of the checks above. The usual forms, with fixes:

• Client or third-party data with no written permission. Get the permission, or carve that data out before you apply.
• Personal information left in free text, attachments or call transcripts. Scrub with automated detection, then sample by hand.
• Fragments, such as closed tickets with the resolution stripped or final reports with no drafts or reviews. Export the full thread or project record, status changes included.
• Too little volume. Our partner looks for companies with 30 or more employees, partly so there is enough volume and variety. Data brokers also commonly report that archives under roughly 150 GB in total are hard to place, so check whether more systems could be included.
• Too little history. Brokers commonly look for about 3 or more years of retained records. Check how far back each system really goes, because a company can be old and still have deleted its history.
• No internal communications. Brokers commonly report that datasets without internal communications are harder to place. Whether email or chat threads can be included depends on rights, privacy review and anonymization.
• Numbers that do not verify. Record counts, date ranges and headcount are checked during review, and figures that do not hold up commonly end the conversation. Pull them from your systems first.
• A language mix you cannot separate. Tag records by language so the English material can be offered on its own.

The volume, history, communications and verification points above are commonly reported by data brokers, not formal requirements set by any one buyer. They tend to lower how strong a fit looks rather than rule a company out. The eligibility checker asks about all four.

What an AI lab's sample review looks like

Whether you reach a buyer directly or through an intermediary (see how AI training data marketplaces work), reviews tend to move from description to evidence to terms.

1. You describe the data: types, source systems, date range, rough volume, language mix and ownership. With our partner, that is the application, followed by a fit review on data types and volume.
2. The buyer looks at structure and quality, often through a field list or a small anonymized extract. Send only scrubbed records, and state volume as record counts.
3. Rights and privacy questions follow: ownership, client contracts, your privacy policy and the anonymization method.
4. Your team and counsel review the terms: scope, permitted use, exclusivity and the anonymization standard. Data is anonymized and delivered only under a signed licence.

Most of these deals are licences: you keep ownership and grant defined rights to an anonymized copy. Our partner publicly states that companies have been paid $100K to $2M+ per approved data package, though that is not a guarantee. For a rough sense of where your archive might land, try the data value calculator or read what drives the price of company data.

If your records pass the seven checks on paper, check if your company qualifies. It takes about two minutes and nothing is stored.

A checklist before selling data to AI labs

1. List each source system (help desk, wiki, CRM, project tracker, QA tool, code repository) with its date range.
2. Mark which records belong to clients or other third parties, and collect written permission where you need it.
3. Read your privacy policy and customer contracts for limits on secondary use.
4. Pull a small test extract, scrub it, and have someone who knows the data try to re-identify people in it.
5. Separate Quebec-resident data and non-English records so each can be assessed on its own.

Application steps are in the Canadian guide to selling company data to AI.

FAQ

What kind of data do AI labs buy from companies?

Records of real work that are not on the public web: SOPs, knowledge bases and wikis, support tickets and chat logs, CRM metadata and project histories, QA processes and engineering records. Full workflows, from request to outcome, are what buyers look for first.

Is selling data to an AI lab a sale or a licence?

Usually a licence. You keep ownership and grant the lab defined rights to use an anonymized copy for an agreed scope and term.

How much do AI labs pay for company data?

There is no public price list. Our partner publicly states that companies have been paid $100K to $2M+ per approved data package. That is not a guarantee, and price depends on volume, history, uniqueness and how clean the rights are.

Will an AI lab accept data that contains personal information?

Generally not. OpenAI, for example, says it is not seeking datasets with sensitive or personal information. In Canada, remove or irreversibly anonymize personal information before delivery, and apply Quebec's Law 25 standard to Quebec residents' data.

Does my company qualify to sell data to AI labs?

The qualifying signals are 30 or more employees, mostly English data, a US, Canadian or other Western market, and the rights to license the data. The eligibility checker takes about two minutes and stores nothing.

Sources

• The Verge, "OpenAI cofounder Ilya Sutskever says the way AI is built is about to change" (December 2024)
• U.S. Copyright Office, "Copyright and Artificial Intelligence, Part 3: Generative AI Training" (pre-publication, May 2025)
• Epoch AI, "An FAQ on reinforcement learning environments" (2026)
• OpenAI, "OpenAI Data Partnerships" (November 2023)
• Reuters, "Exclusive: Reddit in AI content licensing deal with Google" (February 2024)
• Google, "An expanded partnership with Reddit" (February 2024)
• Associated Press via WIFR, "Judge approves a $1.5B Anthropic settlement over pirated books used to train the Claude chatbot" (July 2026)

General information only, not legal advice. Have Canadian counsel review data rights, anonymization and any licence before data leaves your company.

Could your data qualify? Answer six questions in about two minutes. Nothing is stored.
Check if your company qualifies

Ready now? Apply to license your data (partner link).