Whether a company should license its data to AI companies depends less on the price than on three things: whether it holds the rights, whether personal data can be removed or lawfully used, and whether it can accept a grant it cannot take back. Not selling is a valid outcome, and DataPacts takes no cut of any sale. Buyers' own terms show the weight of a yes: micro1's Data License Master Terms (updated 2026-11-08) take an irrevocable, perpetual grant, and Labelbox states "Models already trained on the data cannot be unlearned" (checked 2026-11-09).
The decision is the first question of AI training data licensing from the seller's side, and income is uneven: Shutterstock's Data, Distribution, and Services revenue was USD 15.8 million in the third quarter of 2026 against USD 65.7 million a year earlier (10-Q).
DataPacts Editorial Team dates each figure here with its stating party; the page on how we source, date and check every figure sets out the rules.
What a Yes Commits the Company To
Saying yes usually means granting a license to train, not handing over ownership, and the license often outlasts the commercial relationship. micro1's Data License Master Terms take a worldwide, royalty-free, nonexclusive, irrevocable and perpetual grant. The company keeps its data, but warrants in s.5.1 that it has granted no third party the same rights on similar terms.
Deleting raw data does not undo training: under s.4.4 micro1 deletes within 120 days after delivery but has no duty to retrain its models.
Terms can differ within one buyer: Defined.ai's Data License Agreement s.8(a) runs "until terminated", while its FAQ says "The license agreement is perpetual" (both checked 2026-11-09). Each of these terms is a clause of the AI data licensing agreement.
Six Questions That Decide the Answer
Six questions decide whether licensing makes sense, and a clear no to any of the first three ends the process before price is discussed. The table compares what a yes and a no point toward, as of 2026-11-09.
| # | Question | A yes points toward | A no points toward | Public reference | Source |
|---|---|---|---|---|---|
| 1 | Does the company hold the rights to every item it would license? | rights audit, then inventory | stop, or license only cleared records | micro1 s.5.1 no-prior-grant warranty; Slack API Terms bar LLM training by third-party apps | micro1.ai (2026-11-08); slack.com (2025-10-10) |
| 2 | Can personal data be removed, or lawfully used? | de-identification under a stated standard | stop, or exclude those records | 45 CFR 164.514: 18 identifier types; CCPA employee exemption inoperative from 2023-01-01 | ecfr.gov; cppa.ca.gov |
| 3 | Can the company accept a grant it cannot recall? | contract review | stop | micro1 grant irrevocable, perpetual; Labelbox: trained models "cannot be unlearned" | micro1.ai (2026-11-08); prism.labelbox.com, checked 2026-11-09 |
| 4 | Does the company meet program thresholds? | approach programs | smaller or no offers | micro1 30+ full-time employees; Handshake AI 20+ full-time employees and 3+ years | buyers' pages, checked 2026-11-09 |
| 5 | Would the income matter, and would it last? | valuation and offers | keep the data | Reddit: "substantially all" from two counterparties near the end of their terms; Shutterstock down 76 % | SEC 10-Q filings |
| 6 | Have staff and customers been told, where rules require it? | proceed with notice | wait | BetrVG § 87(1) Nr. 6; WOR art. 27(1)(k), (l); Code du travail L2312-8 II 4° | national statutes |
Rights and Personal Data Come Before Price
A company can license only what it has the right to license, and personal data in the records decides how much of the rest can leave the building. A rights audit settles which records the company may license; Slack's API Terms bar third-party apps from using API Data "to train a large language model" (2025-10-10).
Personal data narrows the set further. The CCPA employee-data exemption has been inoperative since 2023-01-01, and HIPAA Safe Harbor (45 CFR 164.514) removes 18 types of identifiers, the standard named for health data in the Spirit Airlines sale, which is not approved as of November 9, 2026.
Program Thresholds and Company Size
Buyer programs publish entry thresholds, so a small company may get no offer or a small one. micro1 states 30+ full-time employees, Handshake AI 20+ full-time employees and 3+ years in operation, and Endless, an aggregator, about 30+ employees (all checked 2026-11-09).
Replay, a data vendor that licenses de-identified data on to labs, states USD 10K to 100K for 20 to 50 employees (checked 2026-11-09). The full data sale readiness check starts from bands like these.
Whether Licensing Income Lasts
Licensing income from AI buyers has been large for a few sellers and uneven for most, and filings show how fast it can fall. Reddit's S-1 of 2024-02-22 discloses USD 203.0 million in aggregate value of data licenses, counterparties not named, and Tempus AI states USD 200 million in license fees from Pathos over three years (10-K, 2026-02-24).
Reddit's 10-Q of 2026-10-29 states that "substantially all" licensing contract value comes from two counterparties, not named, "approaching the end of their respective current terms". Shutterstock's revenue fell 76 % (see above), and Wiley's AI license revenue was USD 13.7 million in the first quarter of fiscal 2027 against USD 28.9 million a year earlier (10-Q).
Benefits: What a License Can Bring the Company
The benefits of licensing are income from records that usually carry no book value, continued ownership under a nonexclusive grant, and in some programs costs the buyer carries.
Benefits
- Income from records off the balance sheet: IAS 38.63 bars recognition of internally generated customer lists and similar items as intangible assets.
- Large contracts with a stating party exist (Reddit, Tempus AI).
- Ownership and own use continue, and the company may license again where no warranty such as micro1's s.5.1 bars it.
- The buyer can pay extraction, anonymization and transfer: Mercor states that it does (checked 2026-11-09).
- Some programs combine an upfront fee with revenue share: Labelbox states it (checked 2026-11-09).
Program ranges are each buyer's claim, checked 2026-11-09, and some contradict themselves: Handshake's page says "Earn $100K to $4M+" and, on the same page, "Your data could be worth up to $100,000". No seller-confirmed amount received under any program is public. See how to estimate what company data is worth to AI companies.
Risks: Legal, Security and Model Exposure
The main risks are legal exposure for data the company had no right to license, models that keep what they learned, and a counterparty that is breached or disputes the deal.
Risks
- GDPR (applies from 2018-05-25): Art. 83(5) sets fines up to EUR 20,000,000 or 4 % of total worldwide annual turnover, whichever is higher.
- EDPB Opinion 28/2024 (adopted 2024-12-17), para. 114: corrective measures may include erasure of the whole dataset used to develop an AI model and/or the model itself.
- FTC: the final Everalbum order of 2021-05-07 requires deletion of models and algorithms developed using users' photos and videos, and FTC staff stated on 2024-02-13 that retroactive, surreptitious changes to privacy terms to allow AI training may be unfair or deceptive.
- Employee data: staff rules apply in Germany (BetrVG § 87(1) Nr. 6), the Netherlands (WOR art. 27(1)(k) and (l)) and France (Code du travail L2312-8 II 4°, information-consultation only).
- Transfers abroad: 28 CFR Part 202 (effective 2025-04-08) requires, for data brokerage with a foreign person, a contractual bar on onward transfer to countries of concern (202.302).
- Buyer-side breach: Mercor's breach dates are 2026-03-24 and 2026-03-30 (California Attorney General), and In re Mercor.io Corporation Data Breach Litigation, N.D. Cal. Lead Case No. 3:26-cv-02821-WHO, has a stipulation to stay pending mediation (Dkt. 44, filed 2026-07-21), as of 2026-11-09.
A breach or lawsuit at a buyer is a fact about that counterparty, not a finding against it. The seller checks a buyer's record, including entries on data buyer incidents, before signing.
Downsides: What the Company Gives Up
The downsides are the costs of a yes that are certain rather than possible: a grant that cannot be recalled, exclusivity that shuts out other buyers, and staff time before any payment.
Downsides
- An irrevocable, perpetual grant (micro1) and deletion of raw data without a duty to retrain models (s.4.4, backups excluded).
- Exclusivity: Labelbox states exclusive by default, and Turing states it is the sole seller for the first 90 days.
- Acceptance by silence and a capped remedy: micro1 s.1.3 deems acceptance after 30 business days, and s.7.2 caps liability at 3x fees.
- Time to money as each buyer states it: Labelbox under 21 days from first call, Mercor 2 to 4 weeks after signing, Mode about three months (all checked 2026-11-09).
- Revenue share pays only when the data sells: Turing states "Nothing upfront".
When and how buyers pay decides how much of the downside falls early, as set out in when and how AI data buyers pay.
Alternatives: Other Paths for the Data, and Keeping It
A company that decides against a full license still has narrower options, and keeping the data is one of them.
Alternatives, including not selling or licensing at all
- Keep the data and use it internally.
- License nonexclusively or only a de-identified subset: micro1's terms are nonexclusive, while Labelbox states exclusive by default.
- Take revenue share instead of a fixed fee: Turing states "Nothing upfront", and Replay states upfront cash plus perpetual revenue share.
- For a closing or insolvent company, sell through a court-supervised process under 11 U.S.C. 363(b)(1), with a consumer privacy ombudsman (11 U.S.C. 332) where consumers' personally identifiable information is sold.
- Do not sell or license at all, a valid outcome when the income may not last, as the filings above show.
The four monetization models are license, sale, revenue share and data-as-a-service, and each is set against the others in alternatives to selling data to AI labs.
How the Answer Differs by Company Situation
The same six questions give different answers for an operating company, a company that is closing, an insolvent estate, a regulated business and an EU employer. The table compares what changes the answer, as of 2026-11-09.
| Situation | What changes the answer | Who signs | Public reference | Source |
|---|---|---|---|---|
| Operating company | program thresholds; customer and staff rules | the company (board) | micro1 30+ and Handshake AI 20+ full-time employees | buyers' pages, checked 2026-11-09 |
| Company closing down | accounts and exports may disappear before a deal closes | officers before dissolution | micro1 states USD 100K to 3M+ company compensation in its defunct-company program (checked 2026-11-09) | micro1.ai/company-referral |
| Insolvent estate (US) | court approval; ombudsman for consumer PII | trustee or debtor in possession | 11 U.S.C. 363(b)(1), 332; Spirit successful bid USD 10,000,000, not approved as of 2026-11-09 | uscode.house.gov; Bankr. S.D.N.Y. 25-11897, Dkt. 1463, 1779 |
| Health records holder | HIPAA de-identification before any license | covered entity | 45 CFR 164.514 Safe Harbor | ecfr.gov |
| EU employer | GDPR; works-council and notice rules for staff communications | the company after consultation | GDPR Art. 83(5); BetrVG, WOR, Code du travail | EUR-Lex; national statutes |
In the Spirit Airlines estate, Google LLC is the successful bidder (Dkt. 1463, 2026-08-14); the sale is not approved as of November 9, 2026, and a hearing is set for 2026-11-17. A company that is closing faces a different clock, covered in data from a shut-down or defunct company.
When the Answer Is No, for Now
The answer is no, at least for now, when the company cannot show it holds the rights, cannot remove personal data, or is being pushed to decide fast.
- Customer contracts or confidentiality clauses bar the use.
- Third-party platform terms bar it, as Slack's API Terms do.
- Personal data cannot be de-identified under a stated standard.
- Staff are not yet informed where notice or co-determination applies.
- An exclusive, perpetual grant is offered for a low fixed fee.
- An unsolicited approach sets a short deadline.
A no now can become a yes after the rights and consent work, so the decision can wait. Separate checks show how to tell whether a data licensing email is legitimate.
Where the Decision Sits in the Process to Sell or License Company Data
The decision comes before step 1 of the seller process, and every later step can still end in no. The guide on how to sell or license company data to AI companies covers what follows a yes:
- Data inventory for an AI data sale
- Company data that cannot be licensed for AI training
- Customer consent and confidentiality clauses
- Employee notice and consent
- PII removal and de-identification
- Choosing a buyer for your data
- Evaluating an offer for your data
Who Decides Inside the Company, and Who Is Affected
A data license is usually a board-level decision because it grants rights in a company asset for years, and staff and customers whose records are in it are affected without being parties. A guide covers board and shareholder approval for a data licensing deal. Another covers the employee perspective on company data sales.
Program Pages Cited on This Page
The buyer and program terms above come from these programs' own pages, each checked November 9, 2026.
- Handshake AI data buyer profile
- Mercor data buyer profile
- micro1 data buyer profile
- Company data partnership programs
- AI data partnership payouts and program terms compared
Also cited: Endless (an aggregator), Labelbox Prism, Replay (a data vendor), Turing Project Lazarus, Defined.ai and Mode.
Frequently Asked Questions
Is it worth selling company data to AI?
It is worth considering when the company holds the rights, can remove personal data and can accept a license it cannot recall; the price comes after those answers. micro1 states payout tiers from USD 100K to 3M+ (checked 2026-11-09), but no seller-confirmed amount received under any program is public.
Is licensing company data to AI companies legal?
Licensing is lawful when the company holds the rights and meets the privacy and employment rules of each jurisdiction its records touch. The GDPR, applicable since 2018-05-25, sets a maximum fine of EUR 20,000,000 or 4 % of worldwide annual turnover (Art. 83(5)). The rules sit in AI training data law by jurisdiction. This page is information, not legal advice.
Can a buyer be made to stop using the data later?
Usually not for models already trained: buyers' own terms say trained models cannot be unlearned and set no duty to retrain. Labelbox states this (checked 2026-11-09), and micro1 s.4.4 sets no duty to retrain.
Is selling data to AI ethical?
The ethical question turns on whether the people in the records were told and whether the use matches what they were promised. FTC staff stated on 2024-02-13 that retroactive, surreptitious changes to privacy terms to allow AI training may be unfair or deceptive. The page on ethics and risk of selling data for AI training covers the question in full.