The AI Coordination Tax

Start with the work the human still has to do

That leftover burden is what we call the AI coordination tax.

It is not an established industry metric. It is a GO Beyond frame for a simple question:

How much human effort still sits between an objective and an accepted result?

Where the work hides

The data already shows why this matters

Controlled studies show real AI productivity gains. In the current NBER record for Generative AI at Work, access to an AI assistant increased customer-support productivity by 14% on average across 5,179 agents, including a 34% improvement for novice and lower-skilled workers. In a separate six-month randomized field experiment with about 6,000 knowledge workers, users who actually used the tool spent roughly 25% less time on email, about three fewer hours per week, while meeting time did not significantly change. Those results support both sides of the claim: AI can make individual work faster while coordination-heavy work can remain stubbornly unchanged. Sources: NBER · Microsoft Research

That gap—faster individual work without equivalent reduction in coordination burden—is what the coordination-tax frame is meant to measure.

Verification can become its own job

As AI takes on more generation, human work often shifts toward verification, integration, and stewardship.

That can be a good trade when the accepted outcome improves.

It becomes a bad trade when the person spends so much time checking, routing, re-explaining, and recovering that the nominal time savings disappear.

Measure the burden directly

Tokens and generations do not tell us whether the system reduced work.

Better measures include:

• manager minutes per accepted outcome;
• number of human interventions;
• duplicate work and repeated briefing;
• handoff defects;
• verification time;
• recovery time and cleanup cost;
• accepted outcomes per human hour.

What should remain human

Judgment, permission, preference, and accountability should remain human where they matter.

The coordination tax is the avoidable work created because the system cannot carry enough of the machinery itself.

A better system should absorb more routine coordination while keeping the decisions that actually require judgment visible.

A one-week audit

Pick one recurring AI workflow and track it for a week.

How many times did you restate context? How many outputs did you reformat? How many handoffs did you perform manually? How much time did verification take? How many times did you rescue a failed path?

Then compare that total human effort with the accepted outcome.

That is much closer to the value people actually experience.

What the metric is for

AI capability is improving quickly.

That is useful progress.

The metric matters because raw model speed does not tell us how much work the person still has to do before the result can be accepted.

The next productivity question is not only “How fast did the AI generate?” It is “How much human effort still stands between the objective and the result?”