The Oversight Tax
Harvard Business Review published something worth reading slowly in February.
The title: “AI Doesn’t Reduce Work — It Intensifies It.”
Not a fringe take. Not a contrarian hot take. HBR. The same publication that ran the 2024 cover story on AI as the productivity revolution. Now running the correction.
The reactions were predictable: “The tools need to improve.” “We haven’t figured out prompting yet.” “Give it time.”
All of it misses the point.
What the data actually shows
Employees now experience approximately 275 interruptions daily from meetings, emails, and notifications. After each interruption — including the ones generated by AI tools — it takes 23 minutes and 15 seconds to fully regain deep focus.
Meanwhile, AI saves an average of 1.5 to 2.5 hours per week per worker.
Do the math. You’re not winning.
But that’s not even the interesting part. The interesting part is where the work is going. Because AI did reduce execution time on defined tasks. The hours aren’t being lost to the old work. They’re being absorbed by something new.
The work AI creates
Every AI action generates a management event.
AI drafted an email. Someone has to decide whether to send it, edit it, or redo it — and that decision carries more weight than it used to, because the draft was fast and therefore likely imprecise.
AI captured the meeting. Someone has to determine which action items are real commitments, which were off-the-cuff ideas, which were already resolved.
AI completed the research. Someone has to verify which parts are accurate, which are confidently wrong, and which need a second source before they can be used.
None of this shows up in the productivity savings calculation. The hours saved on execution are real. The hours created by oversight are real too. One gets measured. The other doesn’t.
The HBR data isn’t describing bad AI. It’s describing a gap in the architecture.
The pattern nobody’s naming
There’s a phrase from the CHI 2025 research on generative AI and knowledge workers that’s stuck with me: “self-reported reductions in cognitive effort with corresponding reductions in confidence.”
Less effort. Less confidence. That’s not a productivity gain. That’s a trade.
AI compressed the execution phase. In exchange, it expanded the verification phase. And verification is cognitively expensive — not because it’s hard, but because you have to maintain the original context, hold the AI output alongside it, and make a judgment call about the gap between them.
Multiply that by every AI action in your stack. Every document. Every draft. Every summary. Every suggestion. Every automated task that touched something you’re responsible for.
That’s where the hours went. Into a new category of work we don’t have a name for yet.
Let’s call it oversight work.
Why oversight work is harder than execution work
When you execute a task yourself, the context is in your head. You know what you were trying to accomplish, what you cut, what you deferred, and why.
When AI executes it, you inherit the output with no context. You have to reconstruct the reasoning, evaluate the judgment calls the AI made invisibly, and verify that what was produced matches what you actually needed.
That’s structurally harder than doing the work. Not because AI is unreliable — it often isn’t. But because the management of outputs requires more cognitive overhead than the management of your own work.
And there’s a second problem: loops.
Every AI action opens a loop. “Is this draft sent?” “Did this task get completed correctly?” “Was this action item tracked?” When you do the work yourself, closing the loop is automatic. When AI does it, closing the loop requires a deliberate check. A review step. A verification event.
Most people are running dozens of open loops from AI actions they haven’t gotten back to. The loops don’t disappear. They sit in the background, consuming attention, surfacing at wrong moments, adding to the 275 interruptions per day that are already making deep focus structurally impossible.
The wrong fix
The reflex is to use AI to manage AI. Get a better system prompt. Add a review agent. Build a checklist.
That’s adding more AI to the stack — which adds more loops, more oversight events, more management overhead. The problem compounds.
The other reflex is to reduce AI usage. Fewer tools. Less automation.
That works, but it surrenders the upside. The execution savings are real. Giving them back to reclaim focus is a bad trade.
Neither fix addresses the actual gap: there is no layer managing what AI produces on your behalf. No system that knows which AI actions completed, which loops are open, which commitments came from a draft that was never finalized. No infrastructure that takes the oversight work off the human and routes it correctly.
What’s actually missing
The productivity promise of AI is real. The productivity delivery is structurally incomplete — not because the tools are bad, but because the coordination layer above them doesn’t exist.
You need a system that knows: what AI did today, what it created that needs a human decision, what it started that hasn’t been closed. A layer that surfaces the right oversight at the right moment, instead of letting oversight accumulate into a second invisible job.
We are spending billions on AI execution. We are spending almost nothing on AI management infrastructure.
HBR saw the symptom. The cause is a missing layer. Until that layer exists, every AI capability gain will come with a hidden management tax — paid in focus time, paid in accumulated loops, paid in the cognitive overhead of overseeing a system you built to save you time.
The tools aren’t failing. The architecture is incomplete.
Eliran Keren — Founder of Deeplica, building the coordination layer for the person at the center.
Sources: AI Doesn’t Reduce Work—It Intensifies It — HBR, Feb 2026 · Knowledge Worker Productivity Statistics 2026 — Speakwise · The Impact of Generative AI on Critical Thinking — CHI 2025 · Gartner: AI Projects Stall Before ROI — April 7, 2026
