A large share of agent products are sold with a staffing metaphor: hire an AI sales development representative, an AI support agent, an AI researcher. It is an effective pitch because it maps onto a budget line that already exists. It is also a poor predictor of where these systems work, and people are making expensive decisions on the strength of it.
**The metaphor gets the failure mode wrong.** When a junior employee is out of their depth, they usually know, and they ask. An agent does not have that signal. It proceeds with the same confidence at the edge of its competence as in the middle of it, which means the useful question is never "is this task at the right level" but "how will I find out if it went wrong". Tasks where the answer is "immediately and cheaply" are viable. Tasks where the answer is "in three weeks, from a customer" are not, regardless of how well the agent handles them in the average case.
**What actually predicts success is verifiability, not difficulty.** Agents do well on hard tasks with tight feedback loops — writing code that has to pass a test suite, for instance, which is genuinely difficult and completely checkable. They do badly on easy tasks with no feedback loop, such as writing a short email to an important customer, which any competent person handles without thinking and where being subtly wrong is expensive. This inverts the intuition the staffing metaphor gives you.
**Scope is the second axis.** An agent given a well-bounded task with a defined output performs far better than the same agent given an open-ended objective, even when the open-ended version is nominally simpler. "Upgrade this dependency and make the tests pass" works. "Improve our test coverage" produces a great deal of activity and an unclear result. The staffing metaphor encourages the second phrasing, because that is how you would brief a person.
Partner
VoicePen AI
Audio & MusicQuickly transcribe or convert audio and voice memos into blog posts; with the most powerful AI speech models.
We earn a commission if you sign up — at no extra cost to you.
Partner
VoicePen AI
Audio & MusicQuickly transcribe or convert audio and voice memos into blog posts; with the most powerful AI speech models.
We earn a commission if you sign up — at no extra cost to you.
**Unattended running compounds errors rather than amortising them.** The appeal of an agent working overnight is real, and so is the fact that an error in step three shapes steps four through forty. A human doing the same work notices the ground shifting. Systems that check in — that open a pull request rather than merging, or draft rather than send — recover most of the value while keeping the failure bounded. Most of the products worth using have converged on this, which is itself informative.
**Cost is not proportional to output.** A task that takes an agent forty tool calls and one that takes four hundred can produce the same result, and the second is not ten times better. Because pricing is usually per-token or per-run, the economics of agents are much more variable than the economics of an employee, and much harder to forecast. Budget for the distribution, not the average.
**Where this leaves the category.** Agents are genuinely good at a specific shape of work: well-scoped, verifiable, repetitive enough to be tedious, expensive enough in human hours to be worth automating, and cheap to check. That is a large amount of work and it does not look like a job description. It looks like a list of tickets.
The teams getting value from this are not the ones who hired an AI employee. They are the ones who looked at their backlog, found the twenty items that were boring and checkable, and pointed something at those. That is a less compelling pitch and a much better decision procedure.
