The Myth That AI Agents Can Run the Office Unattended

If you sat through any tech keynote this spring, you met the AI agent: a tireless digital worker that books your travel, answers your customers, reconciles your invoices and never asks for a weekend off. Microsoft used its Build conference to declare agents the new operating system, and OpenAI recast its Codex tool for every office role. The pitch is seductive. The reality, so far, is messier.

Start with the arithmetic, because it is unforgiving. Suppose an agent is reliable 85 percent of the time at each step of a task. That sounds respectable. But string eight steps together, and the odds of getting all the way through cleanly are 0.85 raised to the eighth power, which works out to roughly 27 percent. A workflow that looks solid in isolation fails somewhere about three times out of four. Errors in a long chain of actions do not simply add up. They compound.

Demos are clean, the real world is not

This is why agents dazzle on stage and stumble in production. A demo runs on tidy inputs, a cooperative user and a scenario the presenter picked in advance. A real deployment runs on misspelled requests, half-finished forms, systems that time out and people who do not behave the way the script expected. The gap between those two settings is not a rounding error. It is where most agent projects quietly live and die.

The analysts have noticed. Gartner predicts that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, pointing to runaway costs, fuzzy business value and weak controls for when something goes wrong. The firm polled more than 3,400 organisations to get there. It also estimates that of the thousands of vendors now selling "agents," only around 130 are doing anything that genuinely earns the name. The rest is what Gartner calls "agent washing": familiar software wearing a fashionable label.

Useful, but not unsupervised

None of this makes agents a hoax. Used inside clear boundaries, with a human checking the output, they can take real drudgery off people's plates. The honest way to picture today's agent is less a robot employee and more an eager junior hire: fast, confident, and wrong often enough that you cannot sign off on its work without reading it first.

The technology is improving, and the next generation will fail less often. But the version being sold from the keynote stage, of agents running the back office through the night while everyone sleeps, is still a fantasy. For now, someone has to stay awake.