The real boundary of AI agents in a company

Buying a chat box is not the same as hiring an employee.

Chat models are good at producing text. Companies usually need a change of state: a ticket, a published post, inventory, a reply, a ledger entry you can reopen. Once state changes, an error stops being “a sloppy sentence” and becomes “something we cannot take back.” That is the real line between an agent and a chatbot.

I treat a production agent as four things at once: a goal, tools, a human who can stop it, and a record that opens later. Miss one, and it is still a demo.

Work that fits usually looks like this: the steps repeat, right and wrong can be seen, and a mistake can be redone. Xiaoyuan’s rule — actually open a public link once — is that kind of work. Work that does not fit looks like this: permissions are unclear, logs cannot be kept, and a lawyer has to clean up. I will not put an agent on the latter.

Responsibility is another boundary. A model will not sign for the company and will not appear in court. If a proposal says “fully automatic decisions, no human needed,” I treat that sentence as deleted unless the client can name who is legally on the hook.

Narrow is not the same as useless. A browser helper that only acts in a locally logged-in account, without uploading cookies, already solves a class of “I know I should publish this, I do not want to click it.” Companies do not need a universal assistant. They need a worker whose failures can be explained.

If you bring a scene, write this first: what state it changes, who can stop it, how a mistake is replayed. If that paragraph exists, we can talk hours. If it does not, do not buy a talking skin.