micro1 to spend $1 billion on corporate data to train AI agents

micro1, a company that supplies data for AI training, announced on Friday that it will spend $1 billion over the next 12 months to buy and license operational data from other companies. Citi and Hercules Capital are providing the capital. The money will go through the Company Data Partnerships program. micro1 uses the de-identified corporate data to build materials for reinforcement learning, the technique that underpins modern AI agents.

What exactly micro1 is buying

This is not about text scraped from the internet or labeled images. Operational data is the trace of real business work: how employees process requests, make payments, close tickets, fill in a CRM and approve documents. For an agent that has to carry out such chains of actions on its own, this material is more useful than any corpus of articles. It shows not only the outcome but also the order of steps, the exceptions and the mistakes.

Why data has become its own budget line

The public web corpus for large models has largely been exhausted, and quality gains increasingly come from fine-tuning on narrow tasks with verifiable outcomes. That is how reinforcement learning works: an agent acts in an environment, receives a signal of success or failure, and adjusts its behavior. An environment for this kind of training has to be built from something, and real company workflows suit that purpose better than synthetic data.

That is why labs and contractors like micro1 are looking for sources that are not openly available. The stated $1 billion for one year shows the scale: data is turning from a byproduct of business into a commodity with its own price.

Key deal parameters

• Size: $1 billion
• Term: 12 months
• Buyer: micro1, San Francisco
• Funding: Citi and Hercules Capital
• Mechanism: the Company Data Partnerships program
• Data format: de-identified operational data from companies
• Purpose: materials for reinforcement learning, agents

The involvement of Citi and Hercules Capital also says a lot. One is a bank and the other a venture lender, which means the data purchases are effectively financed with debt and structured capital. Such a scheme assumes that the data will bring in revenue faster than the money has to be repaid.

What to check if you are considering selling data

Companies that have accumulated work logs and operation histories gain a new way to monetize them, but also new risks. The first question is the quality of de-identification: removing names and numbers does not guarantee that a customer or counterparty cannot be reconstructed from a combination of indirect attributes. The second is rights. Contracts with customers and partners may prohibit passing data to third parties, even de-identified data. The third is the license terms: duration, the possibility of revocation, and the buyer's right to train models that may later compete with the seller itself.

For teams building agents, the takeaway is simpler. Competition is shifting from model size to access to real processes, and a shortage of data will affect prices more than a shortage of compute. It is worth deciding in advance which of your own data you are ready to license, and which data constitutes your competitive advantage and should stay in-house.