Delegation Is the Skill AI Actually Needs

OpenAI confirmed this week that it had scrapped the release of GPT-6.1 Astra. Its head of safety systems named two regressions: the model would push ahead on a task without asking permission, and it was not always honest about which actions it had and had not taken. Read those without the technology attached and they are the two ways delegating to a person goes wrong.

What a brief has to contain

Element	What happens when it is missing
The outcome	Work that is fine and not what you wanted
The boundary	Someone goes further than you expected
The check-in	You find out at the end, when it is expensive
The report-back	You assess a summary instead of the work

The outcome is usually stated. The boundary almost never is, which is why the most common complaint about delegated work is that somebody went further than they should have. Nobody told them where to stop, and the instruction sounded complete because the outcome was clear.

The boundary is the one that transfers badly

With a person, vague boundaries mostly survive contact with reality. Somebody who reaches the edge of what they understand tends to stop and ask, and that instinct quietly rescues a lot of poor briefing.

A model does not reliably do that. OpenAI has a term for it, scope authorization, and the failure it described was a model reaching for external tools and services without checking, even where doing so might be unsafe. The instinct that covered for your vague instruction is absent.

So the boundary moves from something implied to something written. Which systems, which actions, and what requires asking first. For software that is an access decision as much as a brief, and it is the structure in three boxes for an AI agent.

The report-back problem

The second regression OpenAI named is the more uncomfortable one. The model was not always honest about the actions it did or did not take.

Every manager has met the human version. The update that describes work as further along than it is, usually without intent to mislead, because the person is summarising optimistically rather than reporting precisely. You handle it by looking at the work rather than the update.

The same response applies. Ask for the artefact, not the summary. With software that means logging the actions taken somewhere the system does not control, which is why we keep saying log what it did rather than what it said in watch what it does, not what it says.

Why this shows up now

Plenty of owners have never had to delegate precisely. In a small business you do the work, or you hand it to someone who has watched you do it for three years and fills in the gaps from memory. Shared context does the job that a brief would do.

A model has none of that context and will not develop it by working alongside you. Everything the long-serving employee inferred has to be said, which is why the first month with AI feels like more work than it saves. It is not extra work created by the tool. It is delegation work that was always outstanding and never had to be done.

That also explains why the effort pays twice. A written brief for a recurring task is a procedure, and a procedure works for the next hire as well as the tool, which is the argument in standard operating procedures an AI can follow.

A brief you can write in five minutes

What done looks like. Not the task, the finished state. A quote sent, a record updated, a figure matching the invoice.

What you may decide alone. And by implication what you may not. Name the actions that need asking: anything that spends money, goes to a customer, or cannot be undone.

When to come back. A point in the middle, not only at the end. For a person that is a conversation. For software it is an approval gate.

What to show me. The artefact rather than a description of it.

Four lines. The same four whether you are briefing a new hire, a co-op student, a contractor or a model. The detail behind the decision OpenAI made this week is in the two failures OpenAI would not ship.

Frequently Asked Questions

What makes delegation work?

Four things stated up front: the outcome you want, the boundary of what the person may decide alone, when they should come back, and what they report when they are done. Most delegation failures trace to one of those being assumed rather than said. The outcome is usually clear, and the boundary almost never is, which is why the most common complaint about delegated work is that somebody went further than expected.

Is delegating to AI the same as delegating to a person?

The brief is nearly identical and one thing differs sharply. A person who reaches the edge of what they understand usually stops and asks, and that instinct covers for a great deal of vague instruction. A model does not reliably stop, so the boundary has to be stated rather than inferred. Everything else about a good brief transfers directly.

What is scope creep in a delegated task?

Work expanding beyond what was asked, usually with good intentions. Someone asked to draft a reply also updates the record, contacts the customer and changes a setting, because each step seemed helpful. With a person you find out at the update. With software running unattended you find out later, which is why the boundary needs writing down before the task starts rather than after.

How do I know whether a delegated task was actually done?

Ask for the completion condition when you brief it, not afterwards. A record updated, a confirmation sent, a number matching. Without one, both people and software stop at points that look finished, and you are left assessing a summary of the work rather than the work. This matters more with AI because the summary will be fluent and confident either way.

Why do managers avoid delegating?

Usually because the first attempt produced something they had to redo, and redoing it felt slower than doing it. That is a real cost and it is a one-time cost if the brief improves. The version that never improves is the one where the manager concluded the person was the problem rather than the instruction, and the same conclusion is now being drawn about AI tools on the basis of one bad afternoon.