Why do AI labs pay for workflow data?

What are the labs actually trying to teach?

The newest models are being trained as agents: software that takes a goal, works through several tools, and finishes a job. Resolve the refund. Reconcile the invoice. Move the bug from triage to release. To learn that, a model needs examples of the job being done, in order, with the result at the end.

The public web is thin on exactly this. It holds finished documents, forum questions and marketing pages. It rarely holds the inside view: the ticket that bounced between three queues, the approval that needed a second signature because the amount crossed a threshold, the project that slipped and why. That record exists only inside operating companies, in their help desks, trackers, CRMs and accounting systems.

One plain sentence on how it is used: buyers turn real records into practice tasks, sometimes called RL environments, where a model attempts the work and is scored on whether it reached the outcome the real staff reached.

Why don’t the labs just generate it?

They do generate a great deal of synthetic material. The problem is that generated processes look like what a model already expects. Real processes contain the exceptions, workarounds and odd rules that make work hard. A rule like “refunds over $500 need a second approver” is the kind of detail that only shows up because a real company once needed it.

Real records also carry outcomes. A ticket that closed with no repeat contact in 30 days tells a buyer the steps worked. Generated data has no ground truth of that kind.

Who is buying, and what do they say they pay?

The labs mostly do not buy from operating companies directly. Training-data companies sit in between: they license records from companies, prepare them, and supply the labs. Four publish terms for company workflow records on their own pages.

Figures condensed from each company’s own wording, opened 11/07/2026. Disclosure: several of these firms run referral schemes, and a fee from one may follow an introduction we make; neither the order nor any figure depends on it.

The full side-by-side, with payment timing and ownership terms, is on who buys workflow data.

Is this a real market or a passing burst?

The licensing of text and media to AI developers has a public paper trail going back several years. When Reddit filed its S-1 in February 2024, it put the total value of its data licensing contracts at $203.0 million (TechCrunch covered the filing). Wiley, the academic publisher, booked $23 million from AI licensing in its 2024 fiscal year and $40 million the year after. Shutterstock’s licensing to AI developers brought in $104 million during 2023, by Bloomberg’s account.

Company workflow records are the newer part. Then came Spirit Airlines. Reuters wrote in August 2026 of a $10 million Google bid for the carrier’s internal records (staff mail, Teams chat, spreadsheets, calendars), against $7.5 million from Mercor, pending a judge’s sign-off. Two months later MIT Technology Review covered labs acquiring records from failed startups, with one founder putting his code base sale at roughly $25,000.

Those two figures show the spread. A whole airline’s internal records attract an eight-figure bid. A small company’s code base attracts five figures. Most operating companies that qualify will sit far nearer the second than the first.

Why do published ranges run so high?

Published ranges are written to attract sellers. They describe the top of what a buyer might pay for the best-fitting company, not the middle. micro1’s own estimator labels its output “not a quote or binding offer”. Scale AI calls its range illustrative and says it is not a guaranteed payment.

Three things push a company up a buyer’s range:

• History. Several years of records show processes changing over time; micro1 tells applicants that a longer record usually fits better.
• Links between systems. A support ticket tied to the bug it raised and the release that fixed it is worth more than either alone.
• Volume across teams. micro1 reserves its $500K–$3M band for big operational sets, or for several teams contributing over time.

Things that pull it down are just as plain: records mostly not in English, short history, missing outcomes, and material you cannot prove you own.

What does a company give up by selling?

Under the printed terms, not ownership. micro1 leaves the underlying data with the seller, and Prism leaves the seller in control of data and rights. What a company grants is a license, and the license terms decide the rest: scope, length, whether it is exclusive, and what happens when it ends.

What a company does take on is work and risk. Someone has to inventory the systems, confirm rights, review de-identified samples and read the agreement with a lawyer. Material owned by clients, and customers’ personal details, are off the table unless both contract and statute let them in.

Where should an operations lead start?

With an inventory, not a call. List each system, its date range and who administers it. Then check the rules each buyer publishes against that list. The overview of how to sell workflow data sets out the full deal sequence, from first call to payment.

If your help desk has years of closed tickets with clean outcomes, you already hold the kind of record this market is short of; the open question is only which buyer's rules it fits.