Who's buying company data in 2026, and what they pay

Three kinds of buyer pay for company records in 2026. Data companies license an operating company's records and pass a de-identified copy on to the AI labs. Wind-down services sell the archives of startups that have closed. Now and then a lab buys directly, as Google did at Spirit Airlines' bankruptcy auction⁴. This report covers the buyers that publish their terms or have a deal on the public record, and we last checked every figure on October 29, 2026.

Why AI labs pay for company records

Public text is running short. Epoch AI puts the stock of human-written public text at about 300 trillion tokens and expects models to use it up between 2026 and 2032⁷. Websites are also shutting crawlers out: a 2024 audit by the Data Provenance Initiative found that 28% or more of the most actively maintained sources in C4, a widely used training set, had become fully restricted⁸.

Training on data without permission has become expensive too. Anthropic's $1.5 billion settlement with authors covers more than 482,000 books at about $3,000 each, and a judge gave it final approval in July 2026⁹. Company records solve both problems for a lab. No crawler can reach them, and a licence comes with the owner's consent in writing.

Labs mostly want the process. They're training agents to do office work, so they need to see how a request moves from the inbox to a fix: who picked it up, what they traded off and how it turned out. Before a lab sees the copy, the names, amounts and client details in it are swapped out.

Who's buying, and what each says it pays

These buyers publish their terms or have a deal on the public record. The chip on each one says who produced the number.

Bigger firms sit between the labs and these sellers. Scale AI, Surge AI and Mercor mostly pay experts by the hour to write and grade training examples. Meta paid $14.3 billion for 49% of Scale AI¹³, and Surge had more than $1 billion in 2024 revenue, Reuters reported¹⁴. None of them publishes a price for company records, though Mercor's $7.5 million bid for Spirit's archive shows it buys them too⁴.

Deals with a price you can check

Only a few prices come from somewhere other than the buyer's own pages:

• Google's $10 million winning bid for Spirit Airlines' archive comes from an August 14 court filing⁴. The sale still needs the bankruptcy court's approval⁵.
• SimpleClosure's deals for failed startups pay $10,000 to $100,000 per company¹. One founder netted about $25,000 for his crypto startup's code alone¹¹.
• Reddit's IPO filing disclosed data licensing contracts worth $203 million over two to three years¹⁵.
• HarperCollins' AI licence pays $5,000 per nonfiction title, split 50-50 between the publisher and each author who opted in¹⁶.

Reddit and HarperCollins license published content. A company licence covers private records of how the work got done, so their numbers show the size of the market more than the price of one firm's archive.

Who qualifies, from no minimum to 30 staff

• Our main data partner wants 6 or more staff at a law firm, 10 or more at an accounting or finance firm and 30 or more elsewhere, with 3 or more years of digital records. It scores remote and hybrid teams higher, because more of their work is written down².
• Polyshares sets no minimum headcount and says the length and texture of the record matter more than headcount or revenue¹⁰.
• SimpleClosure works only with startups that are shutting down¹.

What labs buy, and what stays out

A licence covers a de-identified copy of how the work happened. Email and chat show the requests and handoffs. Tickets record the fixes, and SOPs show how the job is meant to go. Client records, HR files and credentials stay with you. So does the CRM, because it's mostly client records.

De-identification takes more than deleting names. In a March 2026 finding on Loblaw's PC Optimum program, Canada's privacy commissioner wrote that "stripping accounts of names, phone numbers and email addresses alone is insufficient" and put the burden of proof on the company⁶. A 2019 study in Nature Communications found that 99.98% of Americans could be re-identified in a dataset with 15 demographic attributes¹⁷.

Five questions to answer before you sign

1. Whose data is in the archive (staff, clients, suppliers), and which privacy law covers each group?
2. Does the purpose you collected the records for cover an AI licence? If it doesn't, have you given notice and asked for consent?
3. Do your client contracts, NDAs and data agreements limit how you can use or share these records?
4. Will a qualified reviewer, ideally an independent one, test the de-identified copy before it leaves?
5. Does the licence spell out exclusivity, a ban on re-identification, audit rights, indemnities and what happens if the buyer resells?

Read the exclusivity terms closely. Polyshares' four published deals were all perpetual exclusive licences, which means those sellers can't license the same records to anyone else³.

A deal takes two to three months

Plan on two to three months from application to payment, which is our main data partner's timeline, plus a few hours of your team's time for the export². Polyshares says it makes an offer within one to two weeks of intake and pays 30 to 60 days after delivery¹⁰.

Get a first estimate

The estimator gives you a first range in about a minute. Send your answers and we'll come back with a firm range within one business day.

In southern Ontario, a licence payment can also cover a company's share of an AI adoption project. FedDev Ontario's AI program offers interest-free, repayable funding of $125,000 to $5 million for up to 50% of a project¹⁸. Applications for projects that start before March 31, 2027 are due November 24, 2026¹⁹.