The Person Who Still Has to Say Yes

I run a government AI delivery team of 40 people. We use AI to save millions of public dollars, and to make each government employee about 10x more effective at the work that still needs a human.

That only holds if the last mile stays human.

On this team, AI can draft the change, open the branch, and write the tests. It still does not get to production.

Production has a name on it.

That is not a slogan we printed after a workshop. It is the gate we actually run. Staging first. Then a named human. Then production. No model, no agent, and no "it looked good" gets to skip that last step.

If you want to know how we think about AI in government work, start there. Not with the demo. With who is allowed to say yes.

What the agents are actually good at

We let automation run where the path is known and the cost of a miss is low.

1. First drafts, not final ones. A model can propose a patch, a reply, a test. That is useful. It is also unfinished work.
2. Speed on the routine. Sorting, tagging, scaffolding, the work that used to eat a morning before the real judgment started.
3. Consistency on a known path. A well-specified workflow does not have a bad Monday. That is a feature. It is not a release manager.

If we refuse those gains, we are not protecting people. We are protecting the line.

What no agent is allowed to own

A model can recommend. It cannot be accountable.

1. The merge. A change can be generated in minutes. It does not become real until a person with a name puts it on the path to production.
2. The exception. Real work arrives sideways: a missing page, a case that does not fit the happy path, a result that is confident and wrong. Exceptions are not edge cases. They are the job.
3. The apology. When we get it wrong, nobody wants a log file. They want a person who can say what happened, what we will do, and who is responsible.

"Human-in-the-loop" is the last mile of every decision we are willing to defend.

The job did not vanish. It moved.

Here is the part we owe people honesty about.

When an agent takes the first draft, the person who used to type it is not doing that work anymore. Pretending otherwise is how trust dies inside the building before it dies outside it.

What we are designing for:

1. The work goes upstream. Fewer hours on keystrokes. More hours on judgment, exceptions, and the cases that do not fit.
2. Oversight is a skill, not a vibe. Reviewing a model's draft is a different muscle than writing the first one. Teams have to be trained for it, staffed for it, and measured for it. Otherwise the loop is a checkbox with a name next to it.
3. The high-leverage seat changed. It is not the person who types faster than the model. It is the person who knows when to stop the ship.

We will not talk about efficiency without talking about the people whose Tuesday just changed. Both things can be true: the old task went away, and the new responsibility is more important.

What the gate looks like here

1. A human names the decision, not the tool.
2. Routine drafts can move fast. Anything that can affect a real user waits for the gate.
3. We stage before we ship. Production is not a place we think out loud.
4. If we cannot explain a change in plain language, it does not go forward. Opacity is a stop sign.

The loop is the product. The model is a component inside it.

What we are asking of you

If you are watching AI change your queue: the valuable seat is the one that can still say no.

If you are building these systems: design the stop. Design the name on the decision. Design the path for the case the form never asked about.

The takeaway: Speed is a public good. So is a person who can still say no.

We want this to be a conversation, not a broadcast. Where should a system be allowed to move without a human, and where should a human still have to say yes? Reply with a real example from your work. We read them.