Definition. Licensing company data to AI labs means granting an AI developer, or a licensing house that supplies them, the right to train models on a de-identified copy of your company's internal records: email, chat, project files, tickets, drawings and similar working documents. You keep the originals and ownership. The buyer pays a one-time license fee, typically after a sample review and a contract that sets the permitted use, the term and what was removed.
This guide covers what labs are buying, who qualifies, the step-by-step process, what gets removed, the contract terms that matter and how long it takes. It is written for owners and executives of US companies with 20 or more office staff and three or more years of records. It is not legal or tax advice.
What are AI labs actually buying?
Labs learned to write from the public internet. The internet does not show how a quote gets priced, how a change order gets approved, or how a support ticket moves from complaint to fix. Your inbox does. One buyer's intake page calls it "the working record of the business".
MIT Technology Review reported on October 13, 2026 that AI developers are buying the archives of shut-down startups through wind-down services, and that demand rose suddenly as labs "hit the limits of public and synthetic datasets." The wind-down firm SimpleClosure told the magazine it had handled more than 150 such deals and that 47 percent of its deals now include code or data sales. Fast Company reported that the former CEO of the defunct software company cielo24 sold every Slack message, internal email and Jira ticket as training data for "hundreds of thousands of dollars," and that SimpleClosure's payouts on about 100 earlier deals ran from $10,000 to $100,000 (Gizmodo's account of the original Forbes report gives the same range).
The largest public deal is Spirit Airlines. Axios reported on August 17, 2026 that Google won the bankruptcy auction for Spirit's internal business data with a $10 million bid. The set includes about 100 million emails and 500 million Microsoft Teams messages, and Mercor was the backup bidder at $7.5 million. TIME reported that Google said personal information would be scrubbed by a third party and that no customer data was included. Court filings summarized by Chapter11Cases show the de-identification must be done by an independent party selected by the debtors, and that the sale still needed court approval as of September.
Why pay for this? TIME quotes a Surge AI executive on reinforcement-learning environments that "are only as good as the data that populates them," and notes that teaching models office work "may require data that is largely private, buried in corporate emails and chats."
What buyers screen for:
- Contextual continuity. Full decision loops that can be traced across systems over years: the idea in a chat thread, the discussion in email, the approval in a ticket, the result in a project folder. Polyshares states that "the length and texture of the record matter more than headcount or revenue" and that visible reasoning in the record affects price more than raw volume. A migration that dropped history lowers value.
- Number of systems. micro1 says companies with "several years of records across multiple systems" are a stronger fit. Handshake AI lists Gmail, Outlook, Slack, Google Drive, Dropbox, Notion, NetSuite, QuickBooks, Figma and AutoCAD among the sources it takes. No buyer publishes a tool count. In Vaquero's intake, companies with five or more systems carrying history draw the most interest.
- Years of history. Handshake AI asks for "3+ years in operation". Polyshares sets no floor and prices on the record.
- Desk headcount. micro1's enterprise program names "30+ full-time employees". Handshake AI names "20+ full time employees". Polyshares names none. Office staff produce the records. Field headcount counts for less unless the field work generates written records.
- English and US-based. micro1 states it is "currently prioritizing US companies, followed by other Western markets" and that data "should be primarily in English."
Judgment-heavy work is worth more than routine work. Buyers' pages say "all industries," and the asks change with lab projects. MIT Technology Review notes that buyers' quotes vary widely because valuations rest on "research theses" about how many useful tasks a dataset can yield. Records with no narrative (a ticket opened and closed with no notes) add little.
Who qualifies, and who does not?
The ideal profile:
- 20 or more office employees, now or at your peak
- 3 or more years of email, chat and project records still on hand
- US-based, working mainly in English
- Operating, winding down or closed
Vaquero's rule of thumb on fit: data about your own work has a low barrier to license. Data about other people has a high barrier. This is Vaquero's framework, not a buyer's published rule, but the buyer pages back it. Polyshares' license "does not cover what you hold on behalf of your own clients". Frankfurt Kurnit's Daniel Goldberg puts it directly: "A company that holds data does not necessarily have the right to sell it."
| Business type | Typical fit | Why |
|---|---|---|
| Product manufacturers, wholesale distributors, e-commerce | Clean | The records are about your own products, prices and operations |
| Commercial construction GCs and subs, engineering firms | Clean to moderate | Own estimates, change orders and drawings; engineering firms retain instruments of service |
| Software and IT product companies | Clean | Own code, tickets and design discussions |
| Freight brokers and 3PLs, field and energy services | Moderate | Own operations; call recordings need a consent review |
| Contract manufacturers and job shops | Moderate | Customer prints and specs may belong to the customer |
| Marketing agencies, staffing, recruiting | Usually out | Deliverables and communications are client property |
| Law firms, accounting firms | Usually out | Privilege, confidentiality rules, client data |
| Healthcare with patient data, financial advisors, insurance agencies | Usually out | Regulated personal data (HIPAA, GLBA) |
The Journal of Accountancy advised CPA firms in November 2026 that permission to use client data for an engagement does not extend to training an AI model, that removing identifiers does not make information nonconfidential, and that unauthorized disclosure of tax return information under IRC Sec. 7216 can carry criminal consequences.
Agencies and professional firms are not a hard no. Their internal SOPs, playbooks and operations records can be licensed. Client work cannot. Ask anyway and a broker will tell you on the first reply.
Fewer than 20 employees? Polyshares states it buys from "any company with real operating history, in any sector" with no headcount floor. If you have years of records, it is worth a profile.
How does the licensing process work?
Vaquero runs a four-phase process. Nothing leaves your systems until the final license is signed, and you can stop at any phase and owe nothing.
Phase 1: Profile and mandate
You share headcount, years of records, the systems you used and whether the company is operating or closed. No files, no names. A broker confirms your company is the kind buyers are paying for, and you get a valuation range from published deals. You sign a short sell-side mandate. The broker represents you, and the buyer pays the broker's fee at close.
Time: five minutes for the profile, one business day for the reply.
Phase 2: Anonymous buyer quotes
The broker registers an anonymous profile with two to three buyers. They see counts and date ranges only, never content. One 15 to 30 minute call confirms what you have. Buyers return indicative prices you can compare side by side. Polyshares publishes "offer within one to two weeks of your intake". Handshake AI's on-page estimator is labeled "an indicative estimate only and is not an offer to purchase your data".
Time: one to two weeks.
Phase 3: Rights review and negotiation
This is where most of the work happens, and it only runs if the quotes justify it. Licensing counsel reads your client and partner agreements and checks ownership. You walk through your systems on one call. Each system is mapped and the records are sorted into include and withhold. Then the broker negotiates price, permitted use, exclusivity term, liability caps and payment timing.
You get a written report on what to include and hold back, then final offers.
Time: two to three weeks.
Phase 4: Delivery and payment
You pick one offer, or none. If you sign, your IT contact exports the agreed scope to the buyer's de-identification partner. The broker checks the de-identified sample and the written record of what was removed before anything is released. You are paid on delivery or from escrow, per the license. Your originals never leave your hands.
Through Phase 3, buyers see an anonymous profile, counts and date ranges, never content.
What gets removed before a buyer sees anything?
Three categories are out of scope from the start:
- Personal information about customers and employees
- Anything a client or partner agreement restricts
- Investor, fund and ownership materials
Before delivery, every name and identifier is replaced with a placeholder. The substance of the work stays. An example of what a buyer receives:
Original (stays with you) From: Karen Doyle kdoyle@ridgelinemech.com Re: CO #14, rooftop unit relocation Structural won't carry RTU-3 at grid line C. Moving to D adds 40 ft of duct and a crane pick. $18,400 plus 4 days.
Licensed copy From: [PM_07] <[EMAIL]> Re: [CO 0014], rooftop unit relocation Structural won't carry RTU-3 at grid line C. Moving to D adds 40 ft of duct and a crane pick. $18,400 plus 4 days.
De-identification is a transformation, not a scrub. The buyer's partner replaces identifiers consistently so the same person is [PM_07] across every message, which preserves the decision loop. Labelbox Prism describes the same method: "names, emails, and internal IDs are tokenized before egress". You see a sample before you sign, and the written record of what was removed.
Two points to settle in writing with every buyer: who performs de-identification and whether it happens before or after the transfer, and who bears liability if something identifying gets through. Buyer terms differ on both. Polyshares says identifying fields are stripped "before anything moves" and that the handling is its responsibility. micro1's FAQ says the company does not need to anonymize first and that identifying information is "removed or transformed before your data is shared with third parties", which puts the step after transfer.
Hard exclusions that no de-identification cures:
- Protected health information. HIPAA treats health data as de-identified only under expert determination or the safe-harbor removal of 18 categories of identifiers.
- Payment card data. The PCI standards define cardholder data as the full card number with or without name, expiration and service code, and bar storing sensitive authentication data at all.
- Privileged communications, HR and personnel files, payroll and tax records, health and benefits information. Frankfurt Kurnit lists each, along with data covered by children's privacy and biometric laws.
- Export-controlled technical data. ITAR defines technical data as information "required for the design, development, production" of defense articles, and the Commerce Department treats the release of controlled technology to a foreign person inside the US as a deemed export.
Recorded calls are a separate problem. Thirteen states require the consent of all parties to record a call, and when the parties sit in different states the strictest standard generally applies. Vaquero withholds recordings unless the consent record is documented.
Which contract terms matter most?
A data license looks like a software license and behaves like a sale. Files cannot be clawed back once a model has trained on them. That shapes every term below.
| Term | Market default | What to negotiate |
|---|---|---|
| Permitted use | Perpetual. micro1's master terms grant a "worldwide, royalty-free, fully paid-up, nonexclusive, irrevocable, perpetual license" | Accept perpetual use. It cannot be undone in practice. Narrow it to model training and evaluation. |
| Exclusivity | Varies. Labelbox Prism states its partnerships are "structured as exclusive by default". micro1's master terms are nonexclusive on their face but refer to "the exclusivity and audit provisions of the Order Form", so the real term sits in the order form | A fixed term, then non-exclusive. Perpetual exclusive means you can never license again and is a problem in any future sale of the company. |
| End customer | Often undisclosed. Polyshares states it does "not disclose which labs take which data" | Named end customer under NDA. No multi-tier sublicensing. |
| Representations | Broad, and they survive. micro1's terms list the warranty section among those that survive termination | Knowledge-qualified, limited to the approved manifest of systems and dates. |
| Liability | Seller indemnifies the buyer for IP and privacy claims. micro1 caps each party at three times the fees paid, with carve-outs for fraud, gross negligence and willful misconduct | A single mutual cap at fees received, with a fraud carve-out only. |
| Deletion | Raw data only. micro1 deletes the data 120 days after delivery but owns the derivatives, models and synthetic data | Extend deletion to derivatives, rejected and voided data. |
| Payment trigger | After acceptance. Polyshares pays "net 30 to 60, once the data has been approved, shared, and anonymized". micro1 takes 30 days to review, then pays within 30 days of acceptance | Payment at delivery, or escrow. No payment on signature with acceptance rights that run for months. |
| Audit | Where it exists, it sits in the order form (micro1) | Audit right on exclusivity and deletion. |
| Confidentiality | Mutual | Exception for acquirers and lenders, so the license does not block a sale or refinancing. |
| Forum | Buyer's home state. micro1: California law, JAMS arbitration seated in California | Seller-friendly forum, or at least neutral. |
Three rules from how deals actually close:
One buyer at a time. Clauses like "offer it to us first" are exclusivity by another name. Run quotes in parallel in Phase 2, then negotiate with one.
Pending exit changes the math. If you are selling the company in the next 12 to 24 months, a data license becomes a diligence item for your acquirer. Tell your broker early. In some cases the right answer is to wait, or to license non-exclusively with an acquirer exception.
Expect ownership diligence. Buyers verify that the person signing owns the company and the records. Handshake AI ties its payout to "the outcome of our due diligence process", and the micro1 master terms make the seller warrant its rights.
Should you use a broker or go direct?
You can go direct. Most buyers have an intake form (Handshake AI, Telegraph Lab). Polyshares states that it is the buyer and that "you are selling to us, not through us". The question is whether one quote and one set of contract terms is the right outcome.
| Route | What you get | What you give up |
|---|---|---|
| Direct to one buyer | Their price, their paper, their timeline | Price discovery. No one reads your client contracts before the buyer prices the scope. |
| Affiliate referrer | A link to a buyer's intake form | Everything above. The referrer is not at the table and has no duty to you. |
| Seller-side broker | Two to three quotes side by side, counsel review before anything is offered, negotiated terms, delivery and payment tracked | Nothing. The buyer pays the broker's fee at close, and the arrangement is disclosed to you before you sign. |
Buyers pay differently by industry and by what their lab needs this quarter. Published ranges run from $100K to $2M at Polyshares, $100K to $3M+ at micro1 and $100K to $4M+ at Handshake AI, while Labelbox Prism says smaller datasets start around $50K. Those are the buyers' published figures, not offers. The same record set can draw different prices from different houses in the same month. That spread is the broker's job.
Two things to check before signing with any broker: that they disclose who pays them, and that they will tell you when not to sell. Some records cannot be licensed today because of client agreements or a pending exit. A broker who sends everything to a buyer is a referrer with a mandate.
How long does it take, and when do you get paid?
| Step | Time |
|---|---|
| Profile to first reply | One business day |
| Buyer quotes | One to two weeks |
| Rights review and negotiation | Two to three weeks, only if quotes justify it |
| First reply to signature | Usually four to six weeks |
| Delivery to payment | Within days of delivery when paid on delivery or from escrow. Buyer-standard terms run net 30 to 60 after acceptance (Polyshares, micro1). |
The long pole is Phase 3. Companies with clean ownership and few client agreements move faster. Companies with many customer contracts spend the time on the include-and-withhold sort, and the report from that sort is useful even if you decide not to license.
Common mistakes
- Sending files to get a number. No buyer needs content to give an indicative price. Headcount, years and systems are enough. Anyone asking for an export before a signed license is a red flag.
- Signing perpetual exclusive. Covered above. It forecloses every future deal and complicates a sale of the business.
- Skipping the contract read. Client NDAs, reseller agreements and platform terms of service can restrict what you export. Frankfurt Kurnit notes that Slack's API terms prohibit third-party applications from using API data to train a large language model. Finding that out after a buyer has priced the scope wastes the buyer's goodwill and your time.
- Ignoring your privacy policy. A policy that promised "we will not sell your data" cannot be cured retroactively. The FTC's February 13, 2024 post states that a business "cannot then unilaterally renege on those commitments after collecting users' data." Frankfurt Kurnit adds that under the CCPA, a transfer of workplace data for money is a sale of personal information unless the data meets the statute's de-identification standard. Counsel checks this in Phase 3.
- Accepting uncapped indemnity. A seller should not carry unlimited liability for a buyer's de-identification process.
Find out what your records are worth
Enter your office headcount and years of records on the Vaquero estimator for a range built from published deals. A reply from a person follows within one business day. No files, no names, nothing to sign.
Rivers Colyer is a co-founder of Vaquero Data Co. in Austin, Texas. Vaquero represents companies licensing their operating records to AI labs through licensing houses. This guide is general information, not legal or tax advice.
Sources
All read November 12, 2026.
- MIT Technology Review, "AI labs are buying data from dead startups to train their models," Oct 13, 2026. Demand drivers, SimpleClosure deal counts, pricing on research theses.
- Axios, "Google buys Spirit Airlines emails, chats, documents out of bankruptcy," Aug 17, 2026. $10 million winning bid, data counts, Mercor backup bid.
- TIME, "Why Google Bid $10 Million for a Failed Airline's Data," Aug 25, 2026. Third-party scrubbing, no customer data, why labs want workplace records.
- Chapter11Cases, Spirit Airlines deidentified data sale summary, Sep 13, 2026. Auction record, de-identification condition, approval status.
- Fast Company, "Shuttered startups are selling old Slack chats and emails to AI companies," Apr 2026. cielo24 sale, SimpleClosure payouts.
- Gizmodo, "Failed Companies Are Selling Old Slack Chats and Email Archives to Train AI," Apr 17, 2026. SimpleClosure deal count and $10,000 to $100,000 range, citing Forbes.
- Frankfurt Kurnit Klein & Selz, "Selling Company Data for AI Training: What to Know Before You Say Yes," Oct 1, 2026. Rights to sell, NDAs, Slack API terms, CCPA sale definition, categories to exclude.
- Journal of Accountancy, "Asked to sell firm data for AI training? Be very cautious," Nov 2, 2026. CPA confidentiality rules, IRC Sec. 7216.
- Polyshares home page. Working record, length and texture over headcount, $100K to $2M, client-held data excluded, net 30 to 60.
- Polyshares, "Sell your company's operational data". Offer in one to two weeks, buyer not broker, de-identification before transfer, no disclosure of end labs.
- micro1, "Turn existing workflows into revenue". 30+ employees, US priority, English, tiers, 30-day review and payment, de-identification after transfer.
- micro1, Data License Master Terms, updated Nov 8, 2026. Perpetual nonexclusive grant, order-form exclusivity and audit, 3x cap, 120-day deletion, derivatives ownership, California law and JAMS.
- Handshake AI, data license intake page. 20+ employees, 3+ years, source list, $100K to $4M+, indicative estimate disclaimer, due diligence.
- Handshake AI, Data Partnerships overview. Five-step process, scope and exclusion of sensitive information.
- Telegraph Lab. Rights review, licensing, de-identification and structuring as the stated process.
- Labelbox Prism FAQ. Exclusive by default, tokenization before egress, $50K starting point.
- FTC, "AI (and other) Companies: Quietly Changing Your Terms of Service Could Be Unfair or Deceptive," Feb 13, 2024. Retroactive privacy-policy changes.
- FTC, Gramm-Leach-Bliley Act business guidance. Coverage of financial advisors and insurance.
- Kilpatrick Townsend, "Wiretap Laws in the United States," Jul 29, 2024. All-party-consent states and the strictest-standard rule.
- 45 CFR 164.514, HIPAA de-identification standard. Expert determination and safe harbor.
- PCI Security Standards Council glossary. Cardholder data and sensitive authentication data.
- 22 CFR 120.33, ITAR definition of technical data.
- Bureau of Industry and Security, deemed exports. Release of controlled technology to a foreign person in the US.
