How selling company data to AI labs works

Nobody can tell you precisely, and the buyers' own calculators prove it. For the same company (100 desk staff, 10 years of history), Handshake's calculator returned $383K–$897K and Nyne's returned $49K–$115K when we captured them on 16 October 2026. That is a gap of roughly 8x before any real review of your records.

Published ranges are just as wide: from about $20K for a small closed startup up to ceilings of $5M or more at the biggest programs. A ceiling is what the best case could reach, not what a typical company receives.

Four things move the number most: how many people worked at a desk, how many years of linked history you can export, how many separate systems the history spans, and your industry and status. Chat or email history that stops when you switched tools halfway through counts for much less than an unbroken run.

For a starting band built from the same published figures, use the estimator. Treat it as a way to decide whether this is worth a conversation, not as an offer.

Sources: Who's buying data; how to (actually) broker ai data; AI labs are offering companies big money to hand over their employees' private Slack DMs and emails

AI labs are building agents that do ordinary office work, so they want complete workflows rather than isolated files: the chat thread, the email, the ticket, the document and the code change that all belong to one piece of work, with the history that connects them. Communications usually anchor a deal, and the other systems add depth around them.

Mercor's CEO published which tools appear most often across the 300+ enterprises it has partnered with: Slack, GitHub, Google Drive, Gmail, Figma and Jira lead the list. Several programs set a bar of about five connected tools and two or more years of history.

Some material is off limits almost everywhere. Buyers and their partners list legal privilege, health records, payment data, Social Security numbers and anything involving minors as excluded. Also keep out anything you do not own, such as client data held under a confidentiality agreement or open-source code.

The inventory worksheet shows how to count what you have, system by system, without exporting or pasting any records.

Sources: Mercor has now partnered with > 300 enterprises to help them monetize their data, paying companies up to $10M per enterprise.; How To Master AI Data Brokering With Micro1; micro1 Offers $100K+ for CRM Records That May Not Be Yours to Sell

The usual path is: you describe the company, the buyer reviews that profile, you may share a small sample, the buyer makes an offer, you sign an agreement, you run an export, the data is de-identified, and payment follows acceptance. Each buyer fills in those steps differently, so the comparison table lists who does what.

Published timelines run from about two weeks from upload to payment at Grepped to about three months from first call to payment at Mode. Acceptance checks and payment terms account for most of the difference.

Two rules hold throughout. You never hand anyone a password or an admin login; exports are run by someone inside your company. And no raw data moves before you have signed. Minewell never receives raw data at any stage: we collect a description of your company and your systems, we introduce you to buyers, and everything after that happens between you and them.

If a company is closing, speed matters more, because the value depends on the workspaces still being live.

Sources: AI Data Broker Master Guide; AI labs are buying data from dead startups to train their models

Often it can be done lawfully, but having the data is not the same as having the right to sell it. The answer depends on your customer and partner contracts, the privacy notices you gave people when you collected their information, the terms of each platform you export from, and the law where your people and customers live.

The main questions lawyers raise are: do any contracts or confidentiality agreements stop you licensing the records; do platform terms restrict using exported data to train models; do employee communications touch protected activity under US labour law (NLRA section 7); does a transfer for value count as a sale under California's CCPA; and, for EU and UK staff, is there a valid lawful basis under GDPR and UK GDPR. Employee consent is a particular worry: critics call it an ethical gray area, and a court filing over the Spirit Airlines sale argues de-identification does not remove it.

This is general information, not legal advice. Have your own counsel check your contracts and employee notices before you sign anything. Our legal checklist lists what to bring to that conversation.

Sources: Selling Company Data for AI Training: What to Know Before You Say Yes (via Passle); EPIC Files Amicus Brief Challenging Spirit Airlines’ Bankruptcy Sale of Employee Data to Google

Buyers say they remove names and other identifiers, usually replacing each person with a stable pseudonym such as "employee 632" so that conversations still make sense. Who does that work differs by buyer: the buyer itself, a third-party vendor, or a tool running inside your own environment. That choice affects who sees the raw records, so ask for it in writing and compare it on the buyer table.

Be realistic about what scrubbing can do. Chat and email are unstructured, so a name can sit inside a sentence, a nickname or an attachment, and linked records can let a long-serving employee be recognised from context alone. Critics argue that de-identification does not prevent inferences about workers.

For that reason the safer approach is to leave high-risk categories out rather than rely on scrubbing them: HR and legal channels, health information, anything about customers you hold under contract, and any group where a small number of people would be easy to pick out. Tell your staff what you are planning before it happens, whatever the law requires.

Sources: he made $1,000,000 in 30 days as an AI data broker; Selling Company Data for AI Training: What to Know Before You Say Yes (via Passle)

That is set by the license, so read it closely. The first question is whether you are selling the data outright or licensing it, and whether the license is perpetual or runs for a term. Next is exclusivity: an exclusive license can pay more, with exclusive deals in the reinforcement-learning market reported at roughly 4 to 5 times non-exclusive ones, but it blocks every other buyer for the whole term. Then check whether the buyer can sub-license to others, and what it may use the data for: training only, or also inside its products.

A common structure is about two years of exclusivity. Shorter terms lower the price, and a carve-out limited to one field of use is a middle path. A perpetual, exclusive license is the one to avoid unless someone gives you a specific reason to accept it and a price that reflects it.

Ask also what happens to models already trained when the license ends, and who is liable if a claim arises about material you should not have included. Our guide to comparing offers takes each term in turn.

Sources: An FAQ on reinforcement learning environments; Exclusive vs Non-Exclusive Data License, Explained; he made $1,000,000 in 30 days as an AI data broker

Records are only worth something while they exist, so the first rule is to export before you cancel anything. Workspaces on Slack, Google Workspace, GitHub and similar tools can be deleted or lock you out soon after a subscription lapses, and a buyer cannot value a history that is gone. Keep the accounts live until you have decided what to do.

Who can sign depends on the stage. Before a formal process, the board or a founder with authority can. Once an administrator or trustee is appointed, they usually control the sale and the proceeds, and money normally goes to creditors and investors before anyone else.

The extreme case is the Spirit Airlines bankruptcy, where Google won an auction for the carrier's business data with a $10 million bid, with other bidders in the room. Closed startups sell too, though at smaller sums: cielo24's founder sold 13 years of records for hundreds of thousands of dollars, and others report amounts in the tens of thousands.

Some buyers accept closed companies and some do not, and the buyer directory shows which.

Sources: AI labs are buying data from dead startups to train their models; Why Google Bid $10 Million for a Failed Airline’s Data; AI's New Training Data: Your Old Work Slacks And Emails

Almost always by the buyer, through a referral fee for bringing a company to the program. Published examples: Mode's referral page offers $50K or more per qualifying company, and another account of its terms describes the greater of 6% or $50K; Pewter pays advisors 10% up to a $100K cap. Ask any broker or referrer exactly what they are paid and by whom.

Minewell is paid by buyers too. Every link that earns us a referral is labelled "Featured partner · paid link" and tagged as sponsored, paid placement stays inside a labelled strip rather than the neutral directory, and the fee comes from the buyer, not from your offer. You pay Minewell nothing.

Where a listing is a broker rather than a buyer, we say so on its card and add only what that broker's own pages support: Pewter, for example, routes the company to a single licensing partner it does not name.

The full details are on our disclosure page. If you are an advisor, accountant or lawyer, the rules on being paid for introductions are different in your profession, and you should check yours.

Sources: How To Master AI Data Brokering With Micro1; Who's buying data

Parts of it are real and parts deserve suspicion. Real: named buyers have paid published sums, reporters have confirmed sales, and every row we know of is on our deals page with its source and date. Unproven: how long demand lasts, whether it reaches beyond a few labs, and whether the data is as useful as sellers hope. Critics point out that a failed company's chat logs may be low-quality, and that prices quoted by people paid on commission are promotional.

Also real are the warning signs: an offer that expires in 48 hours, a buyer that will not name its legal entity, any fee charged up front, a request for your passwords or admin credentials, and a pitch to sell rights you do not hold. The red flags guide shows what each looks like and what to do.

The sensible approach is to compare more than one buyer, take your time, get a lawyer to read the contract, and treat any single number, including ours, as a starting point rather than a promise.

Sources: AI labs are offering companies big money to hand over their employees' private Slack DMs and emails; if you sell AI training data, see how you compare: we found 159 companies selling data, RL envs and evals, every fact sourced; Why Data Labeling Startups Are Paying HVAC Companies $150,000 For Data