An AI task contains five decisions, and one person usually makes all of them

Watch a piece of AI-assisted work go through a small team and you can pull out five distinct decisions:

1. Is this worth doing at all?
2. Is it prepared well enough to hand to somebody else?
3. Who runs it, and on whose subscription?
4. Did the outcome actually meet the bar?
5. Did it land?

On most teams I have seen, one person makes all five. Usually the person who had the idea. That is the whole problem, and it does not look like a problem while it is happening - it looks like being efficient.

What each collapse costs

1 and 2 collapsing means an idea goes straight to a work item without anyone deciding it is worth preparing. You end up with a board that is a list of thoughts, and nobody can tell which entries are real.

2 and 3 collapsing is the common one. The person who prepared the task is the person who runs it, so preparation never actually has to be legible to anyone else. The brief lives in their head, the task looks tiny in the tracker, and it cannot be reassigned. You discover the brief was never written the day that person is on holiday.

3 and 4 collapsing is the expensive one. The person who ran the task judges whether the result is good. With a human doing the work that is merely optimistic. With an agent it is close to meaningless: the agent reports success, the runner reads the report, the report becomes the verdict. Nobody lied and nothing was checked.

4 and 5 collapsing is subtler. Accepting the work and the work actually landing are different facts, and only one of them is verifiable by a machine. If "accepted" also means "merged", then either you are blocking acceptance on CI, or you are letting a human assertion stand in for git. Both are worse than keeping two fields.

The one I would fix first

Split 3 from 4. Whoever ran it does not judge it.

This is cheap. It does not need tooling, a process document, or anyone's permission - it needs a second name on the item. And it converts an agent's self-report from a verdict into a claim, which is the single highest-leverage change available in this whole area.

The objection is always throughput: we are two people, we do not have a spare reviewer. Fair. Then swap - you review mine, I review yours - and accept that the review is shallow. A shallow review by someone who did not do the work still catches the confident-wrong-answer failure mode, which is the one you actually have.

The account question is its own decision

Decision 3 has a part people skip: whose subscription does the run consume?

The tempting answer is to share a login so anyone can run anything. Do not. It destroys attribution, makes usage limits collective, and generally violates the terms you agreed to. The alternative that works is moving the task to whoever has capacity, and letting them run it on their own account under their own identity. Same flexibility, no shared credentials.

Wagglet's how-it-works page lays this out as five decisions with one owner each, which is where I got the framing, and it is blunt about the account rule - move the prepared task, keep every account with its owner. It also has a short section on work that should not be handed off at all, which I appreciated, because most tools in this space write as though everything should be.

You do not need five people. You need the five decisions to be visibly separate, so that when one of them is being skipped, somebody notices.