How to sell your company data to AI companies: the 2026 guide

Why AI companies pay for company data

AI labs have trained their models on most of the public internet. What they lack is a record of how work actually happens inside a business: the email thread that led to a decision, the ticket that tracked a problem to its fix, the invoice that closed it out. Forbes reported in April 2026 that labs had started buying old Slack messages, email archives and Jira tickets from companies, because training AI agents to do real jobs takes examples of real jobs (source: Forbes, Apr 16, 2026). In that story, the transcription company cielo24 sold its records through the wind-down service SimpleClosure for "hundreds of thousands of dollars" (source: Forbes, Apr 16, 2026).

By October the market had grown well past shutdowns. MIT Technology Review reported on October 13, 2026 that SimpleClosure had handled more than 150 data deals since April and expected to return more than $10 million to sellers by year-end (source: MIT Technology Review, Oct 13, 2026). The same report describes Surge AI buying archives to build practice environments where AI agents learn tasks, and notes that buyers value records full of judgment calls and disagreements (source: MIT Technology Review, Oct 13, 2026). In August, Google won the auction for bankrupt Spirit Airlines' business records with a $10 million bid (source: MIT Technology Review, Oct 13, 2026), and micro1 later offered $12.5 million for the same records (source: Fortune, Sep 14, 2026).

Then on October 9, 2026, micro1 committed $1 billion over 12 months to buying operational data from enterprises, financed by Citi and Hercules Capital (source: micro1, $1B announcement (Oct 9, 2026), The Next Web, Oct 9, 2026). Mode, Handshake AI, Nyne and a wave of newer brokers now approach operating businesses directly. If you have received an offer, or expect to, the rest of this guide is the background you need before you answer it.

Licensing vs selling your company data

Most of these deals are licenses. You keep your systems and your originals, and the buyer receives a de-identified copy of agreed records with the right to use it on agreed terms. micro1 says companies keep ownership of their data and license a copy (source: The Next Web, Oct 9, 2026). Mercor says it handles extraction and anonymization and that no AI lab sees raw data (source: Mercor, Monetize your data).

Some wind-down services work differently. Sunset buys a shutting-down company's datasets outright and then licenses them to labs itself (source: Sunset blog, Apr 14, 2026). For a company that is closing, a clean sale can be simpler. For an operating business, a license is usually the better fit, because you can limit what is shared, for how long, and for what use.

Three license terms change what a deal is worth to you:

• Exclusive or non-exclusive. An exclusive license stops you licensing the same records to anyone else. Troveo, for example, says exclusivity is time-limited and must be earned (source: Troveo, business data).
• Term. A fixed period, or perpetual. A perpetual license never ends, so price it that way.
• Permitted use and resale. Whether the buyer can only train its own models, or can sublicense your records to other labs.

What buyers want from your records

Buyers describe the same core set of records in slightly different words. micro1 lists operating procedures, knowledge bases, internal documents, CRM data, project histories, QA processes and team decision records (source: micro1, data partnerships). Mercor lists data from Slack, Teams, Gmail, Outlook, Confluence, Salesforce, GitHub, Drive and Notion, and says the more tools you share, the higher the payout (source: Mercor, Monetize your data). Nyne names accounting systems, customer records, support data, scheduling and dispatch, and jobs and estimates (source: Nyne, sell company data).

Three qualities matter more than any single system:

1. Records that show work being done. Procedures, tickets, quotes, work orders and the messages around them. A drive full of finished PDFs shows outcomes. A ticket history shows the steps.
2. Linkage across systems. Buyers typically want records kept connected to one another, so a support ticket still ties to the email thread and the invoice (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026). That is also what makes de-identification harder, which is why scoping matters.
3. Years of history. Handshake AI asks for 3+ years in operation and its estimator asks how long you have been in business (source: Handshake AI, license your data, Handshake AI, data partnerships). Sunset's CEO told MIT Technology Review he values a company's records by headcount, years in operation, data volume and how the team used its tools (source: MIT Technology Review, Oct 13, 2026).

Here is how the systems buyers name map to what a buyer actually learns from them. The middle column lists systems that at least one program names on its own page.

You do not need every row. A company with clean tickets, a CRM people actually used and a decade of email often has more to offer than one that has every system and three months of history in each.

Who buys company data for AI

We track 21 programs that buy or broker company records. They fall into four groups, and the group matters as much as the name, because it decides who pays whom.

Beyond the programs, large AI companies sometimes buy records directly, as Google did with Spirit Airlines. Our directory of companies that buy company data for AI lists every program with its eligibility, advertised pay, seller fees, referral terms and the date its site was registered. For the three largest buyer programs side by side, see micro1 vs Mercor vs Handshake.

How much AI companies pay for company data

Every figure below is the company's own advertised term, linked to where it says so. None is an offer, and none describes what any particular business received.

Why the ranges are so wide, and what moves a company up them:

• Volume and breadth. Mercor prices on data volume and the number of tools connected (source: Mercor, Monetize your data). Troveo says value scales with systems, team size and years of history (source: Troveo, business data).
• Headcount and years. Handshake AI's estimator example for a 100-person US company with 10 years in business shows $383,083 to $896,969, and Handshake labels it not an offer (source: Handshake AI, data partnerships).
• Uniqueness. micro1 reserves its $1M+ tier for highly unique data (source: micro1, data partnerships).
• Exclusions. Every system or team you carve out lowers the price. That is often the right trade, and it should be a deliberate one.

Wind-down services advertise smaller figures. SimpleClosure describes "tens to hundreds of thousands of dollars for unique assets" (source: SimpleClosure AssetHub). Our company data value checker shows which programs you are likely to qualify for and what each one advertises, based on your headcount, years and systems.

Who qualifies to sell business data

Most programs set a floor of about 20 to 30 full-time employees and 3 or more years of records. A few go lower for professional firms, and one publishes no minimum at all.

Size is a proxy for volume. A 25-person company with ten years of email, tickets and CRM history can have more to offer than a 60-person company that moved systems last year. Language and geography also matter: micro1 and Endless want English-language records and start with the US (source: micro1, data partnerships, Endless).

If you are closing or selling your business

Most of the early deals came from companies that were shutting down (source: Forbes, Apr 16, 2026, MIT Technology Review, Oct 13, 2026), and that route is still open. Several programs are built for it:

• micro1 runs a separate track for companies that have shut down, asking for 50+ employees at peak and substantial Slack or Teams history (source: micro1 defunct-company form).
• Turing's Project Lazarus buys a company's full operating record from operating or winding-down companies. Up to $1M in one payment; standalone codebases $10K to $100K+. (source: aitrainer.work summary of Project Lazarus)
• SimpleClosure sells shutting-down companies' code and workspace data through its AssetHub, and Sunset buys shutting-down companies' datasets outright (source: SimpleClosure AssetHub, Sunset blog, Apr 14, 2026).

If you are selling the business itself, settle the records first. A buyer of your company will want to know whether its records were licensed, to whom, for how long and on what terms, and an exclusive license you signed can limit what the new owner can do with the same records.

In a bankruptcy, a records sale goes through the court. In the Spirit Airlines case, a privacy ombudsman reviewed the deal, booking and passenger records were carved out, and the de-identification firm must certify each transfer to the court before it happens (source: The Next Web, Oct 5, 2026). Those safeguards are a good template for any company selling records on the way out.

How to sell company data to AI, step by step

Every program runs some version of the same sequence. The order below puts the work that protects you before the work that gets you paid.

1. Inventory what you have. List each system, how many records it holds and the date range. Frankfurt Kurnit's checklist starts here (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026). This is the step that tells you which buyers to approach.
2. Review your rights. Read your customer, partner, vendor and data license agreements, your privacy notices, and the terms of the platforms you would extract from (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026). If you cannot license something, it does not belong in the inventory.
3. Scope and set exclusions. Decide which systems, teams and years are in, and write down what never leaves: client files, personnel records, regulated data, NDA material. High-risk data should be excluded outright (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
4. Agree how de-identification works. Ask who does it, to what standard, and how you can check it. In the Spirit Airlines sale, a specialist firm de-identifies the records and must file a certification with the court before every transfer, and the buyer must publicly commit not to re-identify the data (source: The Next Web, Oct 5, 2026). That is a useful model even for a private deal.
5. Collect offers. Approach more than one buyer with the same scope. Nyne says it will beat a qualifying written offer from another buyer (source: Nyne, compare data licensing offers), which only helps you if you have one.
6. Negotiate the contract. Price, payment schedule, exclusivity, term, permitted use, deletion, warranties and the cap on your liability. More on each below.
7. Deliver. The buyer or its contractor extracts the agreed records, de-identifies them and transfers them. Mercor says it covers extraction, anonymization and transfer at no cost to the business (source: Mercor, Monetize your data).
8. Get paid. Payment usually follows delivery or the buyer's acceptance of the data.

Before your first call with any buyer, have these ready:

• Headcount today and at peak, and the year you started keeping records in each system.
• A list of every system with a rough record count and date range.
• Your draft exclusion list, and the contracts that mention data use or confidentiality.
• The privacy notices and employee policies that applied when the records were created.
• Who in the company is authorized to sign a license.

In practice the clock that matters most is your own. A rights review and scoping can take weeks before anything is signed, and it is the part no buyer will do for you.

What stays out of a data license

The records worth the most to a buyer often sit next to records you have no right to share. These are the usual exclusions:

• Client files and privileged material at law firms. ABA Model Rule 1.6 protects all information relating to the representation of a client (source: ABA Model Rule 1.6). Client files, matter content, client communications and work product are out. See our guide for law firms.
• Client books and tax return information at accounting firms. 26 U.S.C. 7216 makes a preparer's unauthorized disclosure of tax return information a misdemeanor punishable by a fine and up to a year in prison (source: 26 U.S.C. 7216). See our guide for accounting firms.
• Protected health information. HIPAA sets its own de-identification standard (source: 45 CFR 164.514 (HIPAA de-identification)), and most operating businesses are better off leaving PHI out entirely.
• Personnel and disciplinary files. Frankfurt Kurnit lists HR files among the high-risk data to exclude, and notes that a disciplinary thread stays sensitive even without a name on it (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
• Third-party confidential material. Anything you received under a customer's or partner's NDA, or a contract that limits its use (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
• Data your platforms restrict. According to Frankfurt Kurnit, Slack's API terms prohibit third-party application providers from using API data to train a large language model (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026). How the records are extracted matters, not just what they are.

Some buyers publish their own exclusions. Endless says it excludes personal data, PHI, privileged client-matter content, client books and proprietary pricing logic (source: Endless). Pewter notes that regulated, privileged or third-party records may be excluded (source: Pewter). Treat a buyer's list as a floor and write your own.

Is it legal to sell company data to AI companies?

No single federal statute allows or bans it. The answer depends on your facts, and Frankfurt Kurnit's October 2026 analysis puts it plainly: most companies can sell this kind of data, but not everything a buyer asks for (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026). The rules that decide it come from several places:

• Your privacy notices. You are bound by the promises in effect when you collected the data. The FTC has warned that quietly changing a privacy policy or terms of service to allow AI training may be unfair or deceptive (source: FTC Office of Technology, Feb 13, 2024, Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
• Your contracts. Customer, vendor and partner agreements can restrict what you do with data they gave you (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
• Platform terms. The software you used to create the records may limit extraction for AI training (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).
• Privacy law. State laws such as California's set their own test for when data counts as de-identified (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026), and HIPAA has its own for health data (source: 45 CFR 164.514 (HIPAA de-identification)).
• Employment and labor law. Covered in the next section.
• Professional rules. For lawyers and accountants, the client confidentiality rules above decide most of the answer.

Buyers do not take this risk off your hands. Handshake AI's page says data partners remain responsible for making sure their transfer complies with applicable law (source: Handshake AI, data partnerships). Expect the contract to ask you to warrant that you have the right to license everything you deliver.

This is general information, not legal advice. For a law firm, the state bar's ethics hotline is the right first call. For everyone else, a privacy or technology transactions lawyer can review the scope, and that review is worth having before you sign.

What you owe your employees

Much of the most valuable material, email and chat, was written by your staff. Three points matter here.

First, labor law. Employee messages about pay, scheduling and working conditions can be protected concerted activity under Section 7 of the National Labor Relations Act, whether or not employees are unionized, and a collective bargaining agreement may restrict a sale or require bargaining (source: Frankfurt Kurnit Klein & Selz, Oct 1, 2026).

Second, employees notice. In the Spirit Airlines sale, the flight attendants' union filed an objection in bankruptcy court, arguing that disciplinary records, payroll history and internal communications stay sensitive even when de-identified, and asking for flight attendant information to be excluded (source: Bloomberg Law, Aug 2026).

Third, the practical steps are simple. Tell employees what is being licensed and what is excluded. Update your workplace technology and privacy policies going forward. Exclude HR systems, personnel files and channels used for personal or grievance discussions. None of this replaces advice from employment counsel, but it removes the most common objections before they start.

How to choose a buyer and compare offers

Two offers with the same headline price can be worth very different amounts. Compare them on these terms:

Fees vary more than most owners expect. Mercor, Mode, Pewter and Endless say they charge the business nothing (source: Mercor, Monetize your data, Mode, data program, Pewter, Endless). Troveo keeps a share of each sale (source: Troveo, business data). DataSupply.ai charges the seller a tiered success fee starting at 20% of the first $50K (source: DataSupply, how it works). Nyne quotes its initial payment after mandatory fees it does not itemize publicly (source: Nyne, compare data licensing offers).

Many programs also pay people who introduce companies to them. micro1 advertises up to $50,000 per referred company that signs, and Mode pays the greater of $50,000 or 6% of license fees paid (source: micro1, company referral, BizBuyData, $50K referral). If someone is steering you toward a single buyer, ask how they are paid. Our comparison of AI data referral programs lists each one's terms.

Warning signs in a data offer

Most programs here are run by real companies. Several are also weeks old, and the market moves faster than anyone's paperwork. These are the points worth checking before you reply:

• A short deadline. Business Insider reported in October 2026 that some companies receive six-figure offers they are asked to accept or reject within 48 hours (source: Business Insider, Oct 2026, Reclaim The Net summary of Business Insider, Oct 14, 2026). A real buyer can wait for your lawyer.
• An unnamed buyer. Pewter's site does not name the licensing partner that pays the business (source: Pewter), and DataSupply.ai licenses to buyers it does not name (source: DataSupply, how it works). Ask who will hold your records before you share anything.
• "Up to" figures. Endless advertises "Get up to $4M" (source: Endless), and Mercor's page carried an "up to $100K per referral" line in July 2026 that was gone by September (source: UsagePricing tracker of Mercor's page). Nyne itself says estimates and "up to" quotes do not count as offers (source: Nyne, compare data licensing offers). Only a written offer with a fixed number does.
• Fees paid by the seller. Most programs are paid by the buyer. If yours is not, get the fee schedule in writing and compare the net figure.
• A very new site. Several brokers' domains were registered in September or October 2026. The buyer directory shows each registration date. A new site is not a problem in itself, but it means there is no track record yet.
• Requests for access before a contract. No one needs admin credentials to your systems to make an offer. Scope first, sign second, extract third.

Why use a sell-side advisor

Every buyer program works for the buyer. That is fine, and several publish clear terms, but no buyer will tell you that a competitor would pay more, or that your scope is wider than it needs to be.

A sell-side advisor works for you. In practice that means:

• Building the inventory and exclusion list with you before any buyer sees it.
• Taking the same scope to several buyers so the offers can be compared line by line.
• Negotiating price, exclusivity, payment schedule, deletion and liability.
• Keeping your counsel in the loop on the rights review.

Regulr is one such advisor. We introduce established businesses to companies that license de-identified records, shop the records to several buyers, and negotiate terms. Records never pass through Regulr: the business licenses directly to the buyer. When a deal closes, we may be paid by the buyer or through a program's referral terms. Regulr has not closed a data deal yet, so we will not quote you results. Other seller-side advisors are listed in the directory, and you should compare us with them as carefully as you would compare buyers.

Frequently asked questions

Sources

Program terms checked against each company's own pages on October 15, 2026.

1. Forbes, Apr 16, 2026
2. MIT Technology Review, Oct 13, 2026
3. micro1, $1B announcement (Oct 9, 2026)
4. The Next Web, Oct 9, 2026
5. Fortune, Sep 14, 2026
6. The Next Web, Oct 5, 2026
7. Bloomberg Law, Aug 2026
8. micro1, data partnerships
9. micro1, company referral
10. Mercor, Monetize your data
11. UsagePricing tracker of Mercor's page
12. Handshake AI, data partnerships
13. Handshake AI, license your data
14. Nyne, sell company data
15. Nyne, compare data licensing offers
16. Mode, data program
17. Mode press release, Oct 13, 2026
18. BizBuyData, $50K referral
19. Pewter
20. Endless
21. Troveo, business data
22. Centific, enterprise data
23. DataSupply, how it works
24. DataSupply, refer a company
25. HUMXN, sell company data
26. Yahoo Finance, Oct 13, 2026
27. Sunset blog, Apr 14, 2026
28. SimpleClosure AssetHub
29. micro1 defunct-company form
30. aitrainer.work summary of Project Lazarus
31. Polyshares
32. SHFA.ai
33. micro1 operating-company form
34. Frankfurt Kurnit Klein & Selz, Oct 1, 2026
35. FTC Office of Technology, Feb 13, 2024
36. ABA Model Rule 1.6
37. 26 U.S.C. 7216
38. 45 CFR 164.514 (HIPAA de-identification)
39. Business Insider, Oct 2026
40. Reclaim The Net summary of Business Insider, Oct 14, 2026