How to Sell or License Company Data to AI Companies: The Complete Seller Guide

Selling company data to AI companies usually means licensing records the company already holds, such as email, chat, tickets, documents or code, to a developer or intermediary that uses them to train or evaluate models. The process runs in 14 steps: readiness, inventory, rights audit, consent and employee notice, PII removal, valuation, choosing a buyer, the data manifest, due diligence, the contract, delivery, payment, tax and accounting, and life after the sale.

DataPacts covers the seller's side of AI training data licensing as an independent reference: it is not a broker, takes no cut of any sale, and treats not selling as a valid outcome.

DataPacts Editorial Team checked the figures on this page against their primary sources on November 9, 2026; the page on how we source, date and check every figure sets out the rules.

What Selling Company Data to an AI Company Actually Transfers

In most deals the company keeps its data and grants a license to train on it, and the license terms decide more than the price. micro1's Data License Master Terms (last updated 2026-11-08) take a worldwide, royalty-free, nonexclusive, irrevocable and perpetual grant: the seller can license the records again but cannot end the buyer's right. Labelbox states that Prism licenses are exclusive by default and that the seller keeps ownership (checked 2026-11-09).

The license goes to an AI developer, a data vendor or a marketplace; brokers, sell-side advisers and aggregators place data with these buyers without acquiring it.

Disclosed values are rare, and each has a stating party. Reddit's S-1 (2024-02-22) discloses January 2024 data licenses worth USD 203.0 million in aggregate over two to three years, with the counterparties not named. Tempus AI's 10-K for fiscal 2025 (filed 2026-02-24) states USD 200 million in data license fees from Pathos over three years, including USD 50 million upfront. In the Spirit Airlines chapter 11 case (Bankr. S.D.N.Y. 25-11897), Google LLC placed the successful bid of USD 10,000,000 for de-identified business data on 2026-08-14 (Dkt. 1463), a sale not approved as of November 9, 2026, with a hearing set for November 17, 2026 (Dkt. 1779). Each is logged among AI data licensing deals with its source.

The Seller Process in Order: 14 Steps

The 14 steps run from checking readiness to living with the license after the sale, and each row links the guide for that step once it is published. The table sets out, as of 2026-11-09, what each step decides and one public reference for it.

Rows marked "upcoming" name guides that are not yet published; their titles become links when each guide appears. Public reference cells are dated claims by the named party or cited legal provisions, not advice.

What Each Phase of the Process Decides

The 14 steps fall into four phases, and a seller can stop after any of them.

Steps 1 to 3: Knowing What You Hold and Whether You May License It

The first three steps decide whether there is anything to sell: buyers set size thresholds, systems limit what can be exported, and third-party terms limit what can be licensed. Among the published thresholds, Endless, an aggregator that places company data with buyers, states about 30+ employees and several years of history; Handshake AI states 20+ full-time employees and 3+ years in operation; and micro1 states 30+ full-time employees (all checked 2026-11-09). These are program entry rules, not legal limits.

Export limits come next. Slack lets workspaces on the Free and Pro plans export public channels only; private channels and direct messages are exportable on Business+ after the owner applies and is approved, and on Enterprise (checked 2026-11-09). Slack's API Terms (2025-10-10) bar third-party apps from using API data to train a large language model and prohibit bulk export without an additional agreement. Microsoft 365 for business deletes data no later than 180 days after cancellation (Microsoft, 2026-06-08). Rights close the phase: micro1's s.5.1 has the licensor warrant that it granted no third party rights to the same data on substantially similar terms.

Steps 4 and 5: Clearing Customer, Staff and Personal Data

Customer contracts, privacy notices and employee rules decide which records may leave the company, and de-identification decides in what form. Under the CCPA, licensing personal information for money is a sale, and the exemption for employee data has been inoperative since 2023-01-01 (Cal. Civ. Code 1798.145(m)(4)). Staff rules differ by country: in Germany the works council co-determines technical monitoring devices (BetrVG § 87(1) Nr. 6), in the Netherlands it must consent under WOR art. 27(1)(k) and (l), and in France it is informed and consulted on new technologies (Code du travail L2312-8 II 4°). FTC staff wrote on 2024-02-13 that surreptitious, retroactive privacy-term changes for AI training may be unfair or deceptive.

The HIPAA Safe Harbor method removes 18 identifier types (45 CFR 164.514). The Spirit Airlines Bill of Sale (Dkt. 1463, 2026-08-14) requires de-identification under Cal. Civ. Code 1798.140(m), plus 45 CFR 164.514 for health data, and names Tonic.ai as Deidentification Agent; that sale is not approved as of November 9, 2026. The EDPB adopted Guidelines 02/2026 on Anonymisation on 2026-07-07 for consultation until 2026-10-30, with three criteria: no record isolation, no linkage and no inference. For EU records, the rules on GDPR and AI training data apply first.

Steps 6 to 9: Pricing, Matching and Letting a Buyer Inspect

Value, buyer and inspection are decided together, because each buyer prices from its own sample review and its own terms. micro1 states three payout tiers of USD 100K to 500K, 500K to 3M and 3M+ (checked 2026-11-09). Handshake AI's page carries the hero line "Earn $100K to $4M+" and, on the same page, an estimator headed "Your data could be worth up to $100,000" (both checked 2026-11-09). No seller-confirmed figure exists for amounts received under either program.

Buyers are labs, data vendors and marketplaces, and an intermediary can sit between seller and buyer. Regulr AI, Inc., a sell-side adviser, states that its service is free to the data owner and that the buyer pays Regulr a fee when a license closes (checked 2026-11-09). Turing states a 90-day exclusive selling window for Project Lazarus, with non-exclusive licenses to labs (checked 2026-11-09).

Datasheets for Datasets (first arXiv version 2018-03-23) is the research model for describing a dataset without handing it over. Turing states 5 business days for sample QA review, and Patronus AI states 5 business days for a sample review and up to 30 days for a full review (both checked 2026-11-09). The directory of who buys AI training data lists each buyer's check date.

Steps 10 to 14: Signing, Handing Over, Getting Paid and Living With the License

The last five steps turn an offer into money, and the terms signed in step 10 decide acceptance, payment, deletion and what the buyer keeps. micro1's master terms give 30 business days to reject a delivery, with deemed acceptance on silence (s.1.3), while its program page says "30 days" (checked 2026-11-09); liability is capped at three times the fees (s.7.2). GDPR Art. 33(1) requires breach notice within 72 hours where feasible, and 28 CFR Part 202 (effective 2025-04-08) requires a contractual bar on onward transfer to countries of concern, such as China and Russia, in data brokerage with a foreign person.

Among the published payment claims, Mercor states payment on delivery, typically 2 to 4 weeks after the signed agreement; micro1's FAQ says within 30 days of invoicing while its homepage says in full within 30 days of acceptance; Mode states about three months from first conversation; and Polyshares states Net 30 to 60 after approval (all checked 2026-11-09). US-source royalties to a foreign corporation carry 30 % withholding under 26 U.S.C. 1442(a), reduced only by treaty. In the EU a B2B data license is supplied where the customer is established (Directive 2006/112/EC Art. 44), with reverse charge under Art. 196. IAS 38.63 keeps internally generated data off the balance sheet, IAS 38.113 bars classifying a gain on sale as revenue, and ASC 606-10-55-58A and 55-58B (per ASU 2016-10) decide over-time or point-in-time recognition.

After the sale, micro1's s.4.4 requires deletion within 120 days after delivery but no retraining of models, and Labelbox states that "Models already trained on the data cannot be unlearned" (checked 2026-11-09). Protege's Provider Terms keep access for existing licensees for 12 months after termination and for academic licensees for 3 years. These clauses appear among payouts and program terms compared with check dates.

22 More Seller Guides, by the Step They Belong To

Twenty-two further guides answer questions that sit inside or before one of the 14 steps, such as whether to sell at all, NDAs, data rooms and M&A. The table lists them as of 2026-11-09 by step, then alphabetically.

Before Step 1: Gains, Exposure, Costs and the Option Not to Sell

Licensing company data can turn idle records into revenue, but the grant usually cannot be recalled once a model is trained, so the decision belongs before step 1. The four lists below set out, as of 2026-11-09, what a seller gains, what it is exposed to, what it gives up and what else it can do.

Benefits of licensing company data:

• Revenue from records that usually carry no book value under IAS 38.63.
• Disclosed nine-figure contracts, in Reddit's S-1 (counterparties not named) and Tempus AI's 10-K.
• Continued ownership and own use under a nonexclusive license.

Risks of licensing company data:

• Trained models keep what they learned, whatever the contract later says.
• EDPB Opinion 28/2024 (adopted 2024-12-17, para. 114) names erasure of the dataset or the model itself as a possible corrective measure.
• Buyer-side breaches: the California Attorney General lists Mercor breach dates of 2026-03-24 and 2026-03-30; In re Mercor.io Corporation Data Breach Litigation (N.D. Cal. Lead Case No. 3:26-cv-02821-WHO) is stayed pending mediation as of 2026-11-09.
• GDPR fines of up to EUR 20,000,000 or 4 % of worldwide annual turnover, whichever is higher (Art. 83(5), applicable since 2018-05-25).

Downsides of licensing company data:

• An irrevocable, perpetual grant outlives the commercial relationship.
• Exclusivity shuts out other buyers, as with Labelbox's default or Turing's 90-day window.
• Staff time before any money arrives, about three months by Mode's own account.

Alternatives, including not selling at all:

• Keep the data and use it internally.
• License non-exclusively, or only a de-identified subset.
• Take a revenue share: Turing states Project Lazarus pays per sale with nothing upfront (checked 2026-11-09).
• If insolvent, sell through an administrator or trustee under 11 U.S.C. 363(b)(1) and 332.
• Not selling or licensing at all, especially where income can fall. Reddit's 10-Q for the third quarter of 2026 states that "substantially all" licensing contract value comes from two partners, not named, "approaching the end of their respective current terms". Shutterstock's Data, Distribution, and Services revenue fell 76 % to USD 15.8 million in that quarter of 2026 (10-Q), and Wiley's AI license revenue was USD 13.7 million for the quarter to 2026-07-31, against USD 28.9 million a year earlier (10-Q).

The decision can wait: steps 1 to 3 only establish what the company holds and may license.

Where the Seller Guide Meets Sellers, Buyers, Law and the Market

The seller guide covers the order of steps; the pages below cover who sells, who buys, which law applies and what deals and prices are on record. They answer questions that sit before or beside the steps.

• Who sells data to AI companies, by seller type
• Selling data from a shut-down or defunct company
• How administrators sell data assets in insolvency
• Company data partnership programs and their terms
• AI training data law by jurisdiction
• Dated AI training data prices per unit
• Tools for AI data sellers
• How DataPacts is funded and what it never does

This page is information, not legal, tax or accounting advice.

Frequently Asked Questions

Do you sell company data to an AI company or license it?

Most deals are licenses: the company keeps the data and grants a right to train on it, often nonexclusive and perpetual. micro1's master terms take a nonexclusive, irrevocable and perpetual grant, while Labelbox states that Prism licenses are exclusive by default (checked 2026-11-09).

Can a closed or insolvent company sell its data to AI companies?

Yes, through its officers, an administrator or a bankruptcy trustee, and in US bankruptcy a sale of personal data against a restrictive privacy policy needs a consumer privacy ombudsman and court findings. In the Spirit Airlines case, Google LLC placed the successful bid for de-identified data, a sale not approved as of November 9, 2026.

How much do AI companies pay for company data?

No public rate card exists; buyer programs publish ranges that are their own claims, and few deal values are disclosed by a party. micro1's tiers start at USD 100K (checked 2026-11-09), and Reddit's 2024 S-1 discloses an aggregate contract value with unnamed counterparties rather than a price per record. No seller-confirmed figure exists for amounts received under buyer programs.

Can individuals sell their own data to AI companies?

This guide covers companies; individuals selling their own data are covered in a separate guide. The steps on this page assume a company with records, staff and customer contracts. The page on individuals selling their own data for AI training covers the case of a person licensing their own material.