AI Did the Work. Now It’s Waiting on You.

AI was supposed to remove the bottleneck. In some workflows, it is moving the bottleneck from production to supervision.

By DeQuan · Predicai · October 31, 2026

I had the shape before I had the name

Signal 018 documented an operating run where more than a hundred pieces of work were staged and nothing was sent.

The important lesson was not the count. It was that execution could move toward AI while consequential responsibility remained human.

But another lesson was hiding underneath it.

If AI can create work faster than a person can inspect it, the bottleneck can simply move.

Generation gets cheap. Review becomes expensive.

I saw that happen in another run when approved copy regenerated three times in one afternoon.

One version drifted. Each new version forced another compliance check.

The system was creating new review work faster than the last review could expire.

Nothing catastrophic happened. That is exactly why the example matters.

Each generation looked inexpensive. Each generation reopened a settled decision.

That is review debt.

Anthropic is measuring the same shift

Anthropic’s internal study is useful precisely because the company is unusually far along in its use of AI.

That also means its percentages should not be generalized to every workplace.

But as a frontier case, the direction is difficult to ignore.

Between fall 2025 and spring 2026, the share of interactive Claude sessions involving meta-work rose from roughly:

10% → 22%

Meta-work includes writing specifications, monitoring agents and reviewing agent output.

Anthropic also reports that approximately 80% of interactive session time consists of the agent waiting for human input.

The company’s own conclusion is unusually direct:

“This acceleration makes human attention the bottleneck for agent output.”

AI output increased too. This is not a story about the technology failing.

It is a story about what becomes scarce when machine execution scales faster than human judgment.

The agent may have capacity. The human still has a calendar.

Source: Anthropic, October 16, 2026

Production can scale faster than judgment

An agent can research twenty accounts in parallel. I cannot make twenty consequential account decisions in parallel.

An agent can prepare a hundred messages. I cannot deeply evaluate a hundred ambiguous messages at once.

An agent can generate five versions of a document almost instantly. Every meaningful version can create another decision.

AI capacity scales. Human attention remains constrained.

A system that ignores the second fact eventually creates queues.

What good review design actually looked like

A real September run made the difference visible.

The wide pull began with 129 records.

Before the final pre-send gate, 26 opt-outs had already been removed upstream.

By the time the live set reached the final gate:

• 119 rows were statused
• 115 were verified
• 4 required verification
• 0 of 119 were opt-outs

There had also been eight correction events earlier in the workflow, before the final gate.

That distinction matters. The eight corrections were absorbed upstream. The four unresolved exceptions were what actually reached me.

Now imagine the alternative.

The AI processes 119 rows and hands all 119 to the human with a request to “review.”

Technically, a human is in the loop. Operationally, we created a review queue.

The real system did something different. Suppression happened upstream. Verification happened upstream. Corrections happened upstream. State traveled with the work.

At the gate, I saw the four exceptions, not 119 undifferentiated outputs.

The human was not removed. The human was better placed.

“Human in the loop” tells me almost nothing

Microsoft researchers interviewed 17 experienced developers working with autonomous software agents.

They found that meaningful oversight occurred in at least four forms:

That is much closer to real AI oversight than the familiar diagram where the system runs until someone clicks Approve.

Some of the most valuable human decisions happen before execution.

Which source can the system use? What does good look like? What should stop the workflow? Which action deserves escalation?

Answer those questions well enough upstream and fewer downstream outputs require judgment.

Good governance should not maximize human review. It should maximize the value of the review that remains.

Source: Microsoft Research / FAccT 2026

Review debt accumulates quietly

Review debt is easy to miss because each individual request seems small.

One more draft. One more verification. One more version. One more approval. One more exception.

The AI is fast enough that another artifact feels nearly free.

But generation cost and review cost live on different ledgers.

The model can produce another version in seconds. The reviewer still has to determine what changed and whether that change matters.

That is why the three regenerations in one afternoon mattered.

The first generation created useful work. The next two reopened settled work. One drifted. All three required human attention.

One of my operating rules now follows directly from that experience:

Approved copy is reproduced, not regenerated.

Creativity is valuable before approval. Consistency is valuable afterward.

A system that cannot distinguish those two states repeatedly converts settled work back into review work.

Push judgment upstream

The best way I have found to reduce review debt is not to review faster. It is to make fewer things require review.

These are operating practices, not the full architecture. That comes next.

The human should not become the hidden API

There is a failure mode I increasingly watch for.

Every time the AI encounters uncertainty, it asks the human. Every time two sources conflict, it asks the human. Every time a rule is missing, it asks the human. Every time another agent becomes confused, it asks the human.

Technically, the workflow still works.

In practice, the human has become the hidden API that every component calls when the architecture runs out.

Some escalation is necessary. Constant escalation is a signal.

The system has not actually absorbed the routine. It has only moved the routine into questions for the operator.

The goal is not zero human involvement. The goal is intentional human involvement.

Sometimes the answer is not to delegate

There is another useful clue in Microsoft’s research on its most advanced AI users.

Frontier Professionals were not simply more aggressive about using AI. They were also more deliberate about deciding what should remain human.

The important principle is simple: every delegated task creates some combination of context work, supervision, verification and failure surface.

Sometimes the trade is excellent. Sometimes you are simply replacing execution with auditing. Sometimes the work itself is the practice you need to preserve.

The sophisticated question is not Can AI do this?

Does delegating this create more useful capacity than supervision cost?

The scarce resource may be attention

Accessible machine intelligence is becoming abundant. Human attention is not.

That changes the optimization target.

A system that produces enormous amounts of work while consuming enormous amounts of human review may still be poorly designed.

A system that produces less work but surfaces exactly the four decisions a person needs to make may create far more leverage.

So alongside every productivity question, I now want another measurement:

How much human judgment does this system consume?

Signal 020 argued that operating AI is becoming a discipline.

This is one of the first costs that discipline has to manage.

AI did the work. Now it is waiting on you.

The question is whether you designed what happens when it comes back.