The AI Site Supervisor: The Role of the COO

When you design a system with multiple roles working side by side, the first problem you run into isn't technical. It's a plain but unavoidable question: who approves what, and in what order?

In our early design, we routed all approvals through a single window. Proposals and status updates from every role came together in one place for review. As a system, it was easy to understand, and it had the advantage of keeping records in one spot.

But once we actually ran it, a different problem showed up. The more cases came in, the more approval requests piled up at that one point. While a case waited for approval, the other roles connected to it stalled too. Urgent and non-urgent cases lined up in the same queue, and prioritizing turned into whatever felt right in the moment. No one made a mistake, yet the whole system kept slowing down, bit by bit. It wasn't a visible accident — it was a slow buildup of delay. In this series, we call this state "silent collapse."

Carving Out the Role of Site Supervisor

To keep this state from taking hold, we set up the COO (short for Chief Operating Officer — the role that runs an organization's day-to-day operations). Using a construction site as an analogy, this isn't the person who draws the blueprints or the person who gives final sign-off. It's the site supervisor who decides the order the day's work gets done in.

The COO's job breaks down into two parts. The first is cross-team prioritization — deciding which case from which role gets handled first. The second is progress tracking — checking whether things are actually moving in the order that was decided, and spotting where they've stalled.

The COO doesn't make the heavy decisions. Those go through the advisory process we've covered in earlier chapters. What the COO handles is the traffic control that comes before that: deciding which case goes to the advisory process first, and which case can be cleared with a quick check instead. That sorting is the job.

Not Rebuilding a Single Point of Approval

There's a catch worth watching for here. If routing every approval through a single window is what caused the silent collapse, simply replacing that window with the COO just reproduces the same problem in a new location. The COO could easily become the new bottleneck (a point where the flow piles up and gets stuck) — now brokering every single case.

To avoid that, we put two constraints on the COO. First, the COO doesn't hold final judgment. It sorts cases and tracks progress, but leaves the actual approve-or-not decision to each role's advisory process. Second, the sorting criteria are set as rules ahead of time. Instead of the COO judging each case on the spot, the routing — which path a case takes — is decided in advance based on its weight and how far its impact reaches. With the criteria fixed, the system is less likely to jam even when the COO isn't available.

A Role You Barely Notice — Until It Stops

The site supervisor's job is easy to forget precisely when it's working well. But once we ran the system, we saw what happens without this plain traffic control: no matter how good a decision any other role makes, the time it takes for that decision to reach the ground keeps stretching out, bit by bit.

The COO role isn't there to improve the quality of decisions. It's there to keep the distance short between a decision being made and it actually moving forward. Don't route every approval through one point, but without someone directing traffic, nothing moves either. This role is carved out to sit in the middle of those two extremes.