What data do AI companies buy from businesses, and why?

Why are AI companies buying business records now?

Because AI developers are trying to build agents that do office work, and the public web does not show how a business handles a refund, a delayed shipment or a contract dispute. Those decisions sit inside company systems. Buyers say public and synthetic data cannot stand in for them^3,4.

The clearest statement of intent came on 9 October 2026, when micro1 committed to spend USD 1 billion over the next 12 months acquiring and licensing company operational data². It says it uses de-identified company data "to build reinforcement learning environments that reflect real business operations"². Mercor puts the reason plainly: frontier models are increasingly trained to work inside companies, and "public data and synthetic datasets cannot replicate that"⁴. Scale AI says AI companies want to improve their models for "corporate personas such as marketers or financiers", and need real-world environments rather than synthetic ones³.

Supply of public text is also a factor. Epoch AI projected in 2024 that, if trends continued, language models would fully use the stock of human-written public text at some point between 2026 and 2032¹⁰. MIT Technology Review reports that demand for defunct companies' data "emerged suddenly last year, as AI training labs hit the limits of public and synthetic datasets"⁸. The best-known example so far is Google's USD 10 million winning bid for Spirit Airlines' internal data in its US bankruptcy, a proposed sale that needs the court's approval^8,12.

What is a reinforcement learning environment, in plain English?

It is a practice world for an AI agent. Developers rebuild a realistic version of a workplace, with inboxes, tickets and records, give the agent a task, and score how well it does. Over many attempts the agent learns what good work looks like. Real records are what make the practice world realistic.

An illustrative example: an agent is handed a customer's complaint, access to a rebuilt order system and the firm's refund policy. It has to find the order, check the delivery record, decide whether a refund is due and reply. The real email thread behind that case shows what an experienced member of staff actually did, including the judgement call the policy did not cover.

Buyers describe the same idea in their own words. Surge AI builds "reinforcement-learning gyms" where agents practise tasks such as debugging an IT system or drafting appellate briefs, and looks for archives that capture "disagreements to tricky judgment calls"⁸. micro1 says real operations involve "incomplete information, competing priorities and exceptions that require judgment"². Real archives also avoid the slips of synthetic worlds, such as references to dates that do not exist⁸. Records are used for testing as well as training: micro1 describes its examples as data that helps "train and evaluate" AI models¹.

Which records do buyers ask for?

Buyers ask for the systems where staff do and discuss their work: email, chat, documents, tickets, CRM and project tools, plus finance and operations systems. Scale AI, Mercor and micro1 each list these by name. Records that connect across systems, such as a ticket linked to its emails and outcome, are worth most^3,4.

Mercor says it supports more than 100 tools and that "the more tools you share, the higher your payout will be"⁴. Scale AI also lists HR platforms such as Workday among business systems³; we treat HR records as out of scope, as explained in how personal data is removed. Workplace chat and email raise particular questions for staff, covered in licensing Slack, Teams and email archives.

What makes one company's records worth more than another's?

Depth and connection matter more than size. Buyers say they value years of continuous history in the same tools, records that link across systems, visible reasoning and disagreement, specialist know-how and mainly English content. Surge AI says one company's Slack, if varied and deep enough, can be worth more than another's Slack, email and calendars combined⁸.

• History. micro1 says companies with several years of records across several systems are generally a stronger fit¹; Scale AI looks for "years of records" in everyday tools³.
• Links. Scale AI describes the records it wants as "rich, interconnected"³, and Mercor pays more as more tools are connected⁴.
• Reasoning. Polyshares says depth of history and "how much visible reasoning the record contains" move the price "far more than raw volume does"⁶. Labelbox Prism ranks human decisions, escalation paths, expert corrections and real failures highest⁷.
• Specialism. Scale AI looks for "domain depth" that is hard to replicate³; micro1 weighs domain expertise and uniqueness¹.
• Language and size. micro1 wants data primarily in English and 30 or more full-time staff; Scale AI prefers 40 or more employees^1,3; Mode's floor is 6 staff for legal, 10 for accounting, finance and medical, and 30 for others⁵.
• Clear rights. Mode reviews ownership before making an offer⁵.

Prices are uncertain. Surge AI told MIT Technology Review that buyers "quote wildly different numbers" because valuations rest on "research theses"⁸. See what your company's data could be worth, and how this plays out in logistics, manufacturing and software companies.

What do buyers not want?

Buyers do not want customer lists, raw personal data or records the seller has no right to license. Most also avoid HR and payroll files, material held for clients, demos and synthetic data, and archives that are thin, very short or mainly not in English^1,3,6,7.

• Identities. Scale AI lists "employee and customer identities" among what never leaves the seller³; Polyshares strips customer names, emails and account numbers before anyone sees a record⁶.
• HR records. Labelbox Prism excludes HR data, performance reviews and compensation records by default⁷.
• Client material. Polyshares' licence "does not cover what you hold on behalf of your own clients"⁶. For law firms, accountants and clinics that rules out much of the archive; the sectors guide sets out what remains.
• Staged or synthetic work. Scale AI wants day-to-day processes, "not demos or synthetic data"³.
• Records without rights. Surge AI was once offered a large company's codebase for millions and concluded the sellers did not have the rights to sell it⁹.
• Material already public, which developers can usually find anyway.

Will synthetic data make real records worthless?

Not yet, and possibly not soon, but nobody can promise. Synthetic data keeps improving, and buyers already use real records to seed synthetic versions. Research shows models trained repeatedly on machine-made data degrade. Yet no company has published tests of how much real archives improve results, so prices could fall^1,8,11.

micro1 says some datasets "may undergo synthetic rewrites that preserve workflow structure"¹, and Labelbox Prism says synthetic data "still depends on real-world grounding"⁷. In other words, real records increasingly feed synthetic data rather than compete with it.

A 2024 paper in Nature found that "indiscriminate use of model-generated content in training causes irreversible defects", which the authors call model collapse, and concluded that data about genuine human interactions will become more valuable¹¹. That supports demand for real records.

The other side matters as much. MIT Technology Review notes that no company has released tests comparing training on real data with the alternatives, and expects prices to "stabilize in the coming year", unless an AI bubble bursts first⁸. The sensible response is neither to rush into poor terms nor to assume the window will stay open indefinitely.

Who are the buyers?

Most buyers are data companies that license records and supply AI labs, rather than the labs themselves. They include micro1, Scale AI, Mercor and Labelbox Prism, newer company data buyers such as Mode and Polyshares, and occasionally a large AI developer bidding directly, as Google's winning bid for Spirit Airlines' data shows^8,12.

micro1 says it licenses data only to labs and enterprises in the US and allied countries¹. Polyshares says it supplies AI labs but does not disclose which labs take which data⁶. MIT Technology Review lists buyers of wind-down data as foundation AI labs, vendors that build simulated workspaces for AI training, and mature tech companies building their own models⁸. Our dated comparison of who buys company data for AI shows each buyer's size floor, published pay and terms, and our figures on the UK company data licensing market show how many UK businesses are in the size range buyers look for. Unfamiliar terms are explained in the glossary.

Find out which of your records qualify. The valuation is a free, confidential 30-minute call, and no data changes hands.