The reason AI firms now pay for everyday business records

Researchers at Epoch AI put the stock of usable public text at around 300 trillion tokens. On current trends, they expect training to consume it somewhere between 2026 and 2032.

Quantity is only half the issue. Published writing is the cleaned-up version. It seldom shows how a dispatcher recovers a missed pickup or how a project lead handles a client who wants more for the same price. Surge AI, which builds training material for agents, says that is why it hunts for archives full of arguments and judgment calls.

Labs can generate artificial data, and they do. MIT Technology Review reported that this approach has limits: invented records contain giveaways, like calendar dates that do not exist. Genuine archives are free of those flaws.

The target is AI agents, programs that carry a task through to the end, whether that is answering a vendor, scheduling a technician or resolving a customer ticket. Agents practise in training environments: simulated offices where they try the job and get a score.

Labs need far more of these environments than exist today. micro1 describes its work as building environments modelled on real company workflows, and ordinary operational paperwork is what it builds them from.

Buyers quote $100K to $2M as the usual band for a business that is still trading. Polyshares cites a $110,000 license with a marketing agency, and Pewter cites $350,000 for an accounting firm of 60. Neither identifies the client.

Research published in Nature Communications in 2019 showed that 15 demographic attributes were enough to single out 99.98% of Americans. That is why the quality of de-identification deserves close attention in any license.