micro1, a San Francisco data lab that helps artificial intelligence research companies train and evaluate their models, announced on Friday, October 9, 2026, a commitment to spend $1 billion over the next 12 months acquiring and licensing companies’ operational data. The initiative runs through its Company Data Partnerships program and is financed with capital provided by Citi and Hercules Capital.
The underlying idea is that information about how companies actually work has become a scarce and valuable resource for the next generation of artificial intelligence agents. micro1 explains that it uses de-identified company data to build reinforcement learning environments that reflect real business operations. In those simulated environments, models and agents learn to navigate workflows, make decisions and complete complex tasks.
“Real business operations involve incomplete information, competing priorities and exceptions that require judgment,” the company says in its announcement. Training artificial intelligence to handle that complexity, it adds, requires environments that preserve the context behind how businesses work. According to micro1, the investment will expand the range of industries and workflows represented in its training environments.
The program page details what the company is looking for: standard operating procedures, process documentation, internal knowledge bases, templates, customer management data, project histories and quality assurance processes. It also wants records of how teams evaluate information and reach operational decisions, as well as human feedback on artificial intelligence outputs in real business settings.
The examples span many industries: code documentation and bug tracking in engineering, reconciliations and approval chains in finance, ticket workflows in customer support, sales playbooks, contract review processes in legal, and inventory and logistics procedures. What they have in common, as the outlet Aivio summarized, is that they are not the output of work but the documentary trail of how it gets done.
The company has also published a payment scale, which depends on the quality, uniqueness and value of the data. A qualified partnership can earn more than $100,000; large datasets or ongoing participation by several teams, more than $500,000; and highly unique proprietary data of significant value for frontier artificial intelligence development, more than $1 million. There is also a referral program paying up to $50,000.
The program targets companies with more than 30 employees and established documentation. micro1 prioritizes US companies, followed by other Western markets, and requires the material to be in English. According to The Next Web, companies keep ownership of their data, micro1 removes personal information “when relevant,” may rewrite some datasets synthetically to loosen their link to the original records and lets companies review samples before they are used.
It is worth being precise about what the figure means. As the outlet DevCuration pointed out, the $1 billion describes a planned purchasing commitment, not an announced funding round. micro1 has not disclosed how much Citi and Hercules Capital are contributing, what financial instrument they are using or on what terms. Hercules Capital generally focuses on lending to venture-backed companies, but that does not clarify the specific terms of this deal.
The announcement comes amid an intense search for quality data. With much of the public text on the internet already used for pre-training, labs need new sources to teach agents how to handle real tasks, with their exceptions and internal rules. In September, micro1 made a competing bid for Spirit Airlines’ customer records, although the privacy official appointed in the airline’s bankruptcy ultimately backed a $10 million deal from Google.
The initiative raises important questions. The first is who the data really belongs to: customer records, support tickets and project histories also describe customers, suppliers and employees, not just the company selling them. De-identification reduces that exposure, but on its own it does not resolve the contractual and privacy rights attached. For companies’ legal teams, that will probably be the first issue to examine.
The second question is who will end up using the data. The public terms do not name the labs or model developers that will train on environments built from a company’s information, nor do they say whether partnerships are exclusive. And the third is what happens afterwards: whether a company can withdraw from the agreement and what obligations apply if the data leaves the agreed framework. These details, common in private contracts, are not included in the published information.
For companies, the program opens a new decision alongside the one about adopting artificial intelligence: whether their operational knowledge, built up over years, should become raw material for training someone else’s systems. micro1 presents it as a new revenue stream for something that already exists. But that knowledge is also a competitive advantage, and licensing it could help build agents able to do the same work for other companies, including competitors.
The scale of the commitment is what is truly new. Devoting $1 billion to a year of corporate documentation is a bet that the bottleneck for agents is no longer open text on the web but the record of how work is done inside organizations, and that companies will be willing to sell it. What to watch is whether any well-known company signs publicly and whether any customer whose records are included raises objections.
