If an outside firm runs your AI agent, it keeps what the agent learns from your work
Glean founder Arvind Jain says the risk is not just relying on tech giants, but losing years of hard won lessons that collect inside their agents.
20VC with Harry Stebbings · Why OpenAI and Anthropic Won't Win the App Layer | Glean FounderArvind Jain, the founder of the workplace AI company Glean, was asked whether large businesses have good reason to fear the makers of the most advanced AI models, known as frontier systems. His answer, he said, depends on the business, but in general, yes.
His reason was not the usual worry about a rival stealing customers. It was about who keeps what is learned from doing the work. If most of the work people do today is eventually done by an agent powered by one of those providers, he argued, a business has handed much of its operations to that provider. That, he said, is "more than technology dependence."
What he meant, he explained, is the way ordinary jobs actually improve. The first time someone does a task, he said, they might document the 10 steps it takes. Then people tweak and optimize the process. Much of that later improvement never gets written down, and workers simply apply those lessons in real time the next time they do the work.
In Jain's view, once an agent does that work, all of that institutional learning accumulates inside the agent. He argued that a business that does not run the agent itself, and does not retain the learning the agent picks up, must then rely on outside AI firms to keep its own operations running.
So he described a fundamental question facing enterprises today, which is how to use AI while retaining control. The compounding lessons that build up as AI does more work, he said, belong to the enterprise doing the work.
Questions people ask
What is an AI agent in this context?
An AI agent is software that can carry out a multi step job on its own, like resolving a support ticket, reviewing a document, or handling an IT request. It uses a large AI model to understand instructions, look up information in company systems, take actions, and report back.
What does it mean for a company to own or control its agent?
In practice it means the business decides where the agent runs, which AI model it uses, what company data it can see, and where its records of past work are stored. That way, if the company switches models, the procedures, feedback, and improvements stay with the business rather than with the outside provider.
Why can't a company just write down what it learns and keep it?
Some of it can be, like written procedures and training manuals. But much of workplace skill is informal, including tiny fixes, exceptions, and judgment that people learn by repetition and rarely record. Jain's point is that an agent doing the work would absorb that unwritten part automatically.