Micro1 said Friday that it would spend $1 billion over the next year acquiring and licensing the records of ordinary business operations. The AI training company is looking for material that can help machines work through the decisions, interruptions and unfinished problems that employees encounter on the job.
The October 9 announcement puts a large sum behind an unusually specific bet: a company’s working history can become a place where an AI system practices. Micro1 says it will use the data to build reinforcement learning environments, settings in which a model attempts a task and receives feedback on its performance.
Citi and Hercules Capital are providing financing, according to the announcement. It does not disclose how much either institution is supplying, the financing terms or how much of the commitment has already been spent.
The billion dollars is a spending commitment, not a report of payments completed.
For a business owner reading the news, a more immediate uncertainty concerns the files themselves. A company can possess years of correspondence and still have difficulty showing a buyer how any particular decision was made.
The missing explanation may be in the head of an employee who left.
Consider a hypothetical manufacturer asked to deliver an order two weeks early. The customer’s request sits in email. A production schedule shows that the original date was impossible to move. A later drawing allows a substitute component, and a supervisor’s approval makes the revised schedule workable. The final order records a successful delivery. On its own, it says little about how the company managed it. Put those records together and a training task becomes possible: determine whether the order can be accelerated, using only the information available at the time. Leave out the revised drawing, or let the model see the eventual delivery date, and the exercise tests something else.
This is the commercial difficulty behind the pledge. The useful material may consist of relationships among files, people and decisions. A business that kept those relationships intact has something different to offer from a business with an equally large, poorly explained archive.
The work between a request and a result
In a September 22 account of Flow, its platform for preparing and reviewing training material, Micro1 described specialists supplying tasks, evidence and tools, then setting the standards by which an AI attempt should be judged.
“Experts define the work and what a good result looks like,” wrote Rumi Elias Calles, its director of research engineering, and Tyler Houchin, a member of its technical staff.
That description gives human expertise a role beyond supplying the original records. Someone has to know which piece of information changes a decision and which apparent mistake is an acceptable way to solve the problem. A rule written too narrowly can punish a sound answer. A loose rule can reward an answer that would get an employee into trouble.
Appen, another company selling data services to AI developers, makes a related distinction on its enterprise-data page. Alongside operating documents and business-system records, it seeks evidence of decisions and escalations, including rejected approvals and cases that had to be reopened. These are the occasions when the standard procedure did not settle the matter. Appen presents that category as especially sought after by advanced AI developers; the page does not provide an independent measure of demand.
A manual can tell an assistant what normally happens. A record of an exception may show what a person did when the manual was insufficient. Whether that judgment was correct still needs to be established.
An old decision does not become good advice merely because someone preserved it.
The problem extends to deciding whether an AI system actually accomplished anything. In a January engineering guide, Anthropic uses an airline reservation to explain the difference between what an agent says and what it does. “Your flight has been booked” is a message. A reservation in the database is an outcome. An evaluation needs to examine the latter.
That distinction changes what a data buyer may need. A conversation ending with a confident answer can be incomplete evidence. The underlying system must show whether the requested change took effect. In the hypothetical factory, an assistant might promise the earlier delivery without securing the substitute part. Judging the reassuring message alone would miss the failure.
Even carefully designed work tests have limits. OpenAI’s GDPval evaluation, introduced in September 2025, covers tasks from 44 occupations. Experienced professionals developed the work and supporting materials. But OpenAI also describes a limitation of giving a model a defined assignment and reference files in one attempt: in actual employment, “tasks aren’t always clearly defined.” A lawyer may first have to speak with a client to establish what work is needed.
The difficulty begins before the first document is opened. That makes the history of how a request changed potentially useful, while also making it harder to prepare a fair test. Which information should the model have at the beginning? What should it have to discover? A company’s records may help answer those questions, but they will not necessarily answer them unaided.
A price for what can be used
Micro1’s data-partnerships page advertises compensation levels starting at $100,000 for qualified companies and reaching $1 million or more for highly distinctive material. It says pricing depends on factors including quality, volume, complexity and uniqueness.
Those figures describe its offer to prospective suppliers. They do not establish the likely price of a particular company’s files.
A useful way to read the numbers is to ask what a buyer would still have to do after receiving the data. If every job can be connected to its instructions, revisions and outcome, less reconstruction may be necessary. If those links are missing, a buyer could be paying for an investigation before it can begin preparing training tasks. The second archive might still contain valuable material, but storage size alone would not reveal the cost of finding it.
Preparation also has a bill, and the identity of the payer matters. Mercor’s company-data program says it covers extraction, anonymization and transfer, with payment to the owner on delivery of the de-identified dataset to the buyer. The stated arrangement is a reminder that a license price and the work required to deliver usable material are separate parts of a transaction.
For the hypothetical manufacturer, removing names would be only one part of preparing the records. A reviewer would also have to decide which customer details could be shared, preserve the connection between the drawing and the approval, and explain any internal shorthand the model could not reasonably understand. These are questions to settle for that collection, not costs that can be calculated from a headline commitment.
There is another possible limit on the value of an archive: models can help produce additional training examples. Micro1’s Flow account describes generating variations and drafts from an expert’s starting material, followed by review. The company is pursuing acquired business data and model-generated material together. Its own description suggests a role for real records as the basis of exercises that can later be expanded, rather than a need to buy a fresh historical case for every attempt.
That leaves an economic question the announcement cannot resolve. How much does an additional company’s history improve a model, once similar work is already represented? A rare procedure might justify considerable preparation. Another collection of familiar tasks might not. Establishing that difference requires testing, and the answer can change as a model learns.
The billion-dollar commitment will draw attention to what businesses have stored. In the manufacturer’s case, the delivery record supplies the ending. The revised drawing and the supervisor’s decision supply the work that made it possible. Before that history can teach a machine anything useful, someone has to recover the connection.
