What does it mean to delegate to an AI agent?

Delegation is the Transfer of More than a Task

When you delegate properly, four things move across, only one of which is the task itself.

The task is the visible part: what you want done. The second is context: what the agent needs to know to do it sensibly, which is everything a human colleague would already carry and an agent carries none of. The third is authority, the bounds of what it may decide or do on its own, which is a different question from whether it is capable of the action. The fourth is the return condition — what brings the work back to a human, and when. A support agent authorised to issue replacement parts up to £200 but required to return anything above it, where the liability changes, has a return condition; a human would phrase it as "check with me before you go past two hundred." It is not error handling. It is the boundary of delegated judgement, set deliberately rather than discovered when something crosses it unnoticed. A human delegate infers most of these from experience and reads the room for the rest. An agent may infer plenty from its training, but it cannot infer the local truths you assumed it knew. So the act of delegating to an agent is mostly the act of making explicit what stays implicit between people.

This is why delegation to an agent is harder than delegation to a person, and also why it is more revealing. The implicit knowledge a good delegator relies on — they'll know not to promise that, they'll check with me before going that far — turns out to be essential, and an agent makes its absence immediately visible. You discover what you were assuming the moment something goes wrong because the agent didn't assume it too.

The part that Doesn't Transfer

One thing never moves across in delegation, to a person or an agent: the accountability. You can delegate the task, the authority to act, and the latitude to decide. You cannot delegate the answerability for the result. That stays with you whether the delegate is a graduate hire or a model.

With a person, this is softened by the fact that they share in the consequences. Their judgement is engaged, their reputation is implicated, their standing and relationships move with the outcome — so they anticipate consequences partly because they will live them. That anticipation is what makes human collaboration quietly self-correcting: the delegate hedges, checks, and flags precisely because the result will land on them too. An agent has none of those incentives. It does not care about the outcome, will not feel the failure, and will produce the next task's output with exactly the same confidence regardless of how the last one went. The appearance of collaboration survives; the mechanism that made collaboration self-correcting does not. Delegating to something with no skin in the game means the answerability concentrates entirely back on the delegator, undiluted. The more capable the agent, the easier this is to forget, because capable output feels like a colleague's work even though nothing behind it is.

Doing it Deliberately

In practice, delegating well to an agent comes down to making the four transfers on purpose rather than by accident. Be explicit about the task, including what a good result excludes, not just what it includes. Supply the context the agent has no other way to obtain — the situation, the history, the things "everyone knows" that it doesn't. Set the authority honestly: what it may do unsupervised, what needs a check first, what it must never do without you. And define the return condition, so the work comes back at the right moment rather than after the decision is irreversible.

None of this is unique to AI. It is ordinary good delegation, applied to a delegate that cannot infer, cannot read the room, and cannot be held to account. What changes with an agent is only that the parts a human would have filled in for you are now yours to provide — and the cost of not providing them is paid by you, after the fact, when the question of what the agent was working from finally gets asked.