What do AI labs do with company data? Sourced uses
The short answer
AI labs use company data to train and test AI agents. Google told Time the Spirit Airlines data "can be helpful in improving our products and AI models." Time linked the bid to RL environments. No source shows real company data in a training run yet.
Every use below comes from a named report. None comes from a lab's own filing about a specific purchase.
In plain words
A lab trains an agent on real work so it can learn how work gets done.
Think of it like this
Think of a library. The lab borrows real work to read and practice from, under rules.
How do labs train agents on company records?
An agent is software that acts for a user instead of only writing text. In its report on micro1's offer for the Spirit records, Fortune wrote that an agent "needs to learn how to choose actions, use tools, respond to intermediate results and recover from mistakes." That takes workflows documented across several systems. Fortune calls the record of those steps a trajectory.
What does an RL environment do with the data?
Epoch defines an RL environment by the actions a model can take and the context that decides their effects. The model tries a task. A grader, usually automated, scores the attempt. The scores update the model.
Time tied Google's Spirit Airlines bid to reinforcement learning environments, where an agent practices tasks and gets scored, and wrote that such environments "are only as good as the data that populates them." Google itself told Time only that the data "can be helpful in improving our products and AI models." Our RL environment page has the prices Epoch reported, with a Slack-grade clone near $300,000.
What other uses are named?
Epoch AI's review of the RL market lists three uses of an environment. They are reinforcement learning, benchmarking and supervised fine-tuning on successful trajectories. Epoch says reinforcement learning is the main use.
micro1's workflows track, as its CFO described it in an undated recorded forum talk, has employees grade a general-purpose agent on their real daily work. micro1 receives the evaluation plus the agent's trajectory (self-reported).
What is not shown in public?
Epoch's FAQ describes clones and expert-written tasks. It does not say real company data fills them. It notes that some labs use their own user data to build tasks.
Marketplace pages name no lab buyers. Our comparison of who buys company data lists the few names reporting supports.
The plain bottom line: the link from your archive to a training run is a plan buyers have stated.
What should a company ask?
Ask which of these uses the buyer plans. Ask whether derived products can be resold. Ask what deletion means after training. Our deletion page covers the last one. The rights check lists what to settle with counsel.
micro1 may pay us a referral fee if a company we introduce goes ahead with it. How we are paid
Have a lawyer read any contract before you sign. They like that.
Apply to micro1's data partnership (paid link)micro1's own data-partnerships pagePaid link. How we are paid
Vaultfish helps you run the process. We are not lawyers and this is not legal advice. If you need counsel who knows this area, we can point you to one. Ask for a referral
