Sell side · Data licensing

Sell Your Data to AI Labs

AI labs have a problem money can solve: the public web is mined out, and the data that teaches models how real work gets done sits inside companies — in email, documents, tickets, codebases and media libraries. A real B2B market now exists for that data. This page is the reference for the sell side: who buys, what commands value, how deals are structured, and the liability work to do before you sign anything.

The market, in cite-ready facts

Who buys — the four buyer types

• AI labs directly (Google, OpenAI, Anthropic, Meta) — mostly for platform-scale content: news archives, community corpora, stock media. Anthropic told WIRED it does not buy workflow data from live companies; labs’ appetites differ by data type, so do not generalize from one headline deal.
• Data vendors with buying programs — Micro1 (operational records), Nyne AI (US company records), Appen’s Enterprise Data program (it states it covers extraction, anonymization and transfer costs), plus expert networks such as Mercor and Handshake that press coverage names licensing professional documents and company records. Vendors package and resell to labs.
• Brokers and marketplaces — FileYield (private brokerage), DataVendor (code and RL environments), Troveo and Versos AI (video), Defined.ai (dataset marketplace with a partner program reporting $600K+ average partner revenue — company figure). Full profiles are in the directory’s Data brokers & marketplaces group.
• Wind-down and specialist buyers — companies closing or selling divisions transact data as an asset (the Spirit Airlines auction is the public example); specialist buyers exist for codebases and professional archives.

What data commands value

• Operational workflow records — email, chat, tickets, CRM history, documents: prized because they show long-horizon, real work. This is the category WIRED and Fast Company describe buyers cold-pitching for.
• Expert demonstrations and work product — professional documents, graded work, reasoning traces: Handshake confirmed to WIRED that it licenses professional documents so labs can test how models handle real knowledge work.
• Production code with history — codebases with commits, pull requests and issue context; systems languages and human-written code price higher (DataVendor).
• Domain corpora — legal, medical, financial, scientific archives with clear rights.
• Media libraries — rights-clean video above all: unpublished creator video has been reported at roughly $1–4 per minute; licensed footage typically $0.75–$3 per minute (2025 reporting).
• Preference and evaluation data — judgments, rankings and rubrics from people whose expertise is verifiable.

How deals are structured

Almost everything legitimate is a license, not a sale: you keep ownership; the buyer gets training and evaluation rights for a defined scope and term. Standard moving parts: exclusivity (worth more, costs you future deals — price it separately), field of use (training only vs. any product use), sublicensing to named labs, delivery format and de-identification duties, payment on delivery-and-acceptance vs. royalties, and what happens to derived models (you cannot untrain a model — negotiate deletion and non-use commitments for breaches instead). The mechanics, term by term, are in our data licensing guide; the buyer-by-buyer comparison is here.

Before you sell: rights, consent, liability

This is the section that separates a licensing business from a future lawsuit — and the reason law-firm guidance (Frankfurt Kurnit Klein & Selz publishes on AI data deals) and compliance vendors (Tonic.ai, on de-identification) rank for this query alongside the brokers. This is general information, not legal advice.

• Rights & consent audit. Do you actually own — or hold commercialization rights to — every record class you plan to license? Customer contracts, employee agreements, privacy notices and the original consent terms under which data was collected rarely contemplated AI-training resale. If the consent does not cover it, the deal does not fix that.
• Privacy law exposure stays with you. CCPA/CPRA, GDPR and sector law (HIPAA for health data) attach to the data’s origin. The Spirit Airlines sale stalled precisely here: a union argued the corpus contained employees’ private information. One concrete example: US tax-return information is governed by IRC §7216 — preparers generally cannot sell compilations of return data, anonymized or not, outside narrow exceptions.
• De-identification is a discipline, not a checkbox. Names are the easy part; quasi-identifiers, free-text details and rare-attribute combinations re-identify people. Insist on a documented de-identification standard, expert review for sensitive corpora, and contractual limits on re-identification attempts by the buyer.
• PII redaction scope. Define whose job redaction is (seller, broker or buyer), what standard applies, who verifies it, and who pays — some buyers (Appen’s program states it covers anonymization and transfer costs) absorb this; many push it to the seller.
• Contract terms to negotiate. Training/evaluation-only use; no public release of the raw data; no identification of your company; security standards and breach duties; audit rights; indemnities that match the risk you are actually taking; termination and post-term data handling; and payment triggers you can verify (delivery + acceptance beats open-ended royalties for most sellers).
• Employee and stakeholder governance. Fast Company’s reporting is blunt: businesses are paid for employee-generated data while employees receive nothing — and at Meta, employee revolt suspended an internal data-collection program. Decide your internal position before a buyer’s deadline decides it for you.

Red flags

• Consumer “sell your data” apps. Browser-history and receipt-sharing apps paying a few dollars a month are a different business — personal-data brokerage, not B2B licensing. Nothing on this page applies to them, and conflating the two is how people get burned.
• Upfront fees to “list” or “appraise.” Legitimate brokers and marketplaces in this category charge on transaction (a fee or split), not for the privilege of an introduction.
• Deadline pressure. Business Insider’s reporting (via Fast Company) describes six-figure offers with 48-hour decision windows. Real diligence takes longer; a buyer who cannot wait for your rights audit is telling you something.
• Unverifiable price talk. Broker marketing circulates $150K–$1M per-dataset figures (LinkedIn posts, pitch decks). Some real deals land there; most published evidence supports five-to-six-figure outcomes for ordinary operating companies. Price your data from a buyer’s written offer for your actual records — not from a pitch.
• Ownership transfer at license prices. If the contract says assignment but the money says license, stop.

For data labs and buyers

This platform covers all three sides of the AI data market — buyers, the experts who produce data, and the owners who license it. If you run a data-acquisition program and want it represented accurately here, see partner with us.