The Coordination Tax | Eliran Keren

A UC Berkeley study tracked 2,000 workers through 2025. Workers who adopted AI tools. Workers who used them every day. Workers doing exactly what the technology promised.

By the end of the year, 67% of them were working more hours, not fewer.

That finding is making rounds now, attached to the usual hand-wringing about AI and burnout. HBR confirmed it in February. Forrester tracked teams whose output expectations increased 35% within 90 days of AI adoption. Cortisol levels up. Micro-breaks down. Cognitive fatigue arriving earlier in the day.

The obvious take: AI failed them.

That’s the wrong level of analysis.

What the data actually shows

The tools worked. Task completion times dropped dramatically. Reports that took ten hours took two. Code that took a week took a day. Analyses that required an afternoon were done by lunch.

The tasks got faster.

The work got heavier.

That’s the pattern nobody’s naming. And until someone names it, the proposed fixes will keep missing the real problem.

Because when a task that used to take six hours takes forty minutes, that time doesn’t go back to the worker. It gets absorbed — by the organization as higher output expectations, or by the worker as overhead nobody accounted for. Forrester tracked exactly how those savings were actually spent: 42% converted directly into more deliverables per sprint. 27% absorbed as expanded project scope. 18% consumed by managing, prompting, and reviewing AI outputs.

8% returned to the human.

You ran faster. The track got longer.

The overhead nobody measured

There’s a specific tax built into every AI-assisted workflow. It doesn’t show up in the time-savings analysis. It shows up in the fatigue.

Every AI output requires verification. Not because the tools are bad — they’re not — but because they’re wrong in ways that require expert knowledge to catch. You can’t just accept the output. You have to hold your own understanding and critically compare it against what the system produced. That’s not a passive read. It’s sustained high-intensity cognitive work. With no recovery breaks.

Before AI, knowledge work contained natural cognitive pauses. Formatting a spreadsheet. Waiting for a report to compile. Searching through documents. These felt like inefficiencies. They were actually recovery time — the pauses that allowed the next hard problem to be approached with a clear head.

AI eliminated them. Every task that used to take twenty minutes now takes twenty seconds. So the worker moves immediately to the next cognitively demanding task. No pause. No reset. Eight hours of uninterrupted high-level cognitive work.

And then there’s the coordination overhead. The outputs that need to be routed. The handoffs that need to be managed. The signals that need to be triaged against everything else sitting open. AI produced more information, more drafts, more analysis, more outputs. Humans became the integration layer — deciding what the output means, routing it to the right place, closing the loop.

Nobody assigned that job to a tool. Nobody built a system for it. It fell to the person.

That’s the coordination tax.

The invisible infrastructure gap

This is an infrastructure problem, not a fatigue problem.

The distinction matters because it points to a completely different solution.

A fatigue problem gets fixed with better time management, clearer boundaries, work-life balance programs. That’s the conversation happening right now across HR departments and productivity newsletters. It’s not wrong. It’s just addressing symptoms.

An infrastructure problem gets fixed by building the missing layer.

Think about what actually creates the coordination tax. It’s not the volume of work. It’s the overhead of managing the handoffs — deciding what each AI output means for what’s still open, routing outputs to the right threads, surfacing which signals actually require human attention versus which ones can be handled below the waterline.

That is exactly the category of work that has no tool.

You have tools that execute tasks. You have tools that monitor information. You have tools that generate outputs. You don’t have a tool that knows what’s open, understands what matters, and absorbs the coordination overhead of connecting all of it.

So humans absorb it instead.

The AI stack got more powerful. The coordination layer got more burdened. The tools handed off their work to people who were already at capacity.

The question the data is actually asking

The Berkeley finding isn’t “AI is exhausting.” It’s: when you speed up individual tasks without building a layer that coordinates the system, where does the overhead go?

It goes to the human.

That’s true of every infrastructure transition. When enterprise software moved from paper to digital, individual transactions got faster. The overhead of managing, reconciling, and routing those transactions didn’t disappear — it moved to whoever was responsible for the system. The solution wasn’t fewer transactions. It was ERP: a coordination layer above the tools.

We’re at the same inflection point in knowledge work.

The tools are fine. The individual task layer is getting very good. The missing piece is what sits above it — a system that knows what’s open, understands what each new output means for the decisions and commitments in flight, and handles the coordination overhead that humans are currently absorbing by default.

Not a smarter assistant. Not a better task automator. A coordination layer.

The kind that closes loops instead of opening them.

What gets measured

There’s a reason this pattern has been invisible until recently. The metrics that organizations track went up. Output volume increased. Sprint velocity increased. Report counts increased. The dashboards looked good.

What deteriorated was harder to measure. Quality of decisions made under sustained cognitive load. Creative problem-solving capacity in the second half of the day. Judgment on novel cases where the AI output couldn’t be trusted on its face.

By Q3 2025, companies that had aggressively ratcheted up AI-driven output expectations were seeing elevated turnover. The cost of replacing exhausted workers exceeded the short-term productivity gains. Now the data is starting to surface in earnest.

67% worked more hours. Only 8% of the time savings came back to them.

That’s not a behavioral problem. That’s a design problem.

The tools were built. The coordination layer wasn’t.

That’s the gap. That’s the fix. And it’s not another productivity app.

It’s infrastructure.