Where AI agents fail in real companies — and where they don't
Agents work when the task has a clear definition of done and a cheap way to check. Most business tasks have neither, until you design them in.
The demonstrations are convincing: an agent books the travel, files the ticket, writes the report. Inside a company, the same agents stall. The difference is rarely the model. It is the environment.
Three conditions
Agents perform reliably when three things are true: the task has an unambiguous end state, the agent can verify its own progress cheaply, and a mistake is recoverable. Drafting a reply satisfies all three. Changing a customer's contract satisfies none of them — unless you build the checks.
Build the checks, not the hype
Most of the engineering in a successful agent deployment is unglamorous: evaluation sets built from real cases, permission boundaries, review queues, logging that a person can read. Skip these and the pilot works and the rollout does not.
The question is not "can the model do this?" but "how will we know when it didn't?"
Where they already earn their keep
Knowledge retrieval with cited sources. Pre-processing of documents ahead of a human decision. Preparing context — briefs, summaries, drafts — for people who remain accountable for the outcome. These are not consolation prizes; they are where most of the value is today.