The Verification Tax | Eliran Keren

Harvard Business Review published research this spring with a name I hadn’t heard before: AI brain fry.

Julie Bedard and colleagues at Boston Consulting Group surveyed nearly 1,500 full-time employees. The finding: a meaningful share reported acute cognitive fatigue linked to heavy AI use — mental fog, slower decision-making, headaches. Not from working more hours. From using AI.

The mechanism they identified is something they’re calling the Verification Tax.

Here’s the definition: the cognitive overhead of checking every AI-generated output for accuracy. The drain of holding your own knowledge while critically comparing it against the machine’s output — simultaneously, continuously, across every task.

They’ve got the symptoms right.

They got the cause wrong.

What the research says

The Bedard study documents something real and worth naming precisely.

Before AI tools, knowledge work contained natural cognitive breaks embedded in workflow. Waiting for a report to compile. Searching through documents. Writing a first draft from scratch. These pauses weren’t waste — they were recovery. They let the brain reset between bursts of high-intensity thinking.

AI eliminated them. When a task that used to take twenty minutes now takes twenty seconds, you don’t recover. You move immediately to the next cognitively demanding task. Then the next. Then the next.

The result is what the researchers call uninterrupted streams of high-intensity cognitive work with no natural recovery time.

And then there’s the verification tax itself. AI output looks correct. That’s what makes checking it expensive. If the AI wrote something obviously wrong, you’d catch it fast and move on. But it writes something plausible — well-structured, confident, mostly right. So you have to hold the original question in your head while reading the output, comparing it against what you know, running a quiet background process that asks: is this actually right?

That’s not just reading. That’s dual-track cognition. And it’s happening every time.

HBR confirmed it: AI tool adoption correlates with increased work intensity, not decreased workload. Workers produce more. Monitor more. Manage more information in the same time. They feel responsible for all of it.

More AI. More exhaustion. Less time returned.

The misdiagnosis

The research frames this as a workflow problem. A human behavior problem. A training and adaptation problem.

The proposed solutions follow logically from that framing: take AI breaks, set verification checkpoints, don’t use AI for high-stakes tasks without careful review, establish organizational norms around AI workload.

These are sensible suggestions. They’re also treating the symptom.

The actual structural failure here is not that people are overusing AI. It’s that the AI is executing while the human is still coordinating.

The verification tax exists because the human has to track what the AI did, whether it was right, whether it aligned with what actually matters today, whether it contradicted something committed to yesterday, whether it fits into the broader context that nobody wrote down.

The agent handles the execution. The human handles everything else.

That is the cognitive load. Not the tasks themselves. The coordination overhead of managing systems that can do anything but don’t understand what matters.

What work looks like without a coordination layer

Picture a knowledge worker using five AI tools on a given day.

Each tool executes well. The draft is good. The summary is accurate. The research is solid. The email is well-written. The analysis runs.

Now picture what that person carries.

They know that the draft assumed a scope that changed two days ago. They know the summary was from a source that contradicts something in the investor deck. They know the email was sent before they’d heard back from legal. They know the analysis was run on last quarter’s data, not the updated set.

None of the AI tools know any of that. None of them were designed to.

The human is the coordination layer. They hold the context. They track the commitments. They remember what changed and what it means. They close the loops — or, more accurately, they try to close them, while carrying them.

This is what HBR is measuring. Not overuse of AI. Not insufficient human judgment. The structural reality that execution has been delegated but coordination has not.

The invisible cost

Here’s the thing about the verification tax: it’s invisible on every productivity dashboard.

If you measure outputs — documents created, emails sent, reports generated — the AI worker looks more productive. They are more productive, in the narrow sense. They produce more.

What the dashboard doesn’t capture: the growing cognitive overhead of managing what’s been produced. The open loops. The context that hasn’t transferred. The commitments that need tracking. The gaps between what the agent executed and what the situation actually required.

The more execution you delegate, the larger this invisible tax becomes.

And right now, everyone is delegating more execution. Every major platform shipped an agent execution layer this year. OpenAI. Microsoft. Salesforce. Google. They are all, simultaneously, increasing the execution capacity of every knowledge worker — without touching the coordination problem at all.

The verification tax will grow in direct proportion to the number of agents deployed.

This is not a prediction. It’s arithmetic.

What a coordination layer actually solves

The answer is not to use less AI. The answer is to stop using humans as the coordination layer.

A system that understands context — your open commitments, your current priorities, what changed and why, what matters today versus what can wait — removes the cognitive overhead of tracking all of that yourself.

You’re not checking the AI’s output against what you know because you’re holding the context. The system holds the context. The verification burden shifts from “is this right given everything I know” to “is this right given what the system knows.”

That’s not a small distinction. It’s the entire difference between cognitive load that scales with AI deployment and cognitive load that doesn’t.

The Bedard study names the problem clearly: the verification tax is the cost of cognitive overhead that wasn’t there before AI tools. What they didn’t surface is the structural reason it exists: the overhead is coordination work that has no infrastructure.

Every piece of coordination work a human does manually — tracking context, managing commitments, noticing what’s open, closing loops, routing tasks to the right moment — is overhead that scales with the number of agents you deploy.

Until someone builds the layer that takes it over.

Where this lands

The sequence that’s been unfolding in the enterprise AI space this year follows a logic worth naming.

We built execution. We deployed it everywhere. Now the humans running the execution are exhausted from the coordination burden that didn’t disappear — it multiplied.

HBR is measuring the cognitive tax of manual coordination at scale. What they’re calling AI brain fry is what the absence of a coordination layer feels like from the inside.

The problem is not that AI is making work harder. The problem is that AI is doing the easy part — the execution — while the hard part — understanding what to execute, why, for whom, and whether the result actually closed the loop — still lives in human heads.

That’s an infrastructure problem. Not a wellness problem.

And it has a structural solution.

Deeplica is building the coordination layer. The system that understands what matters, tracks what’s open, and decides what the agents work on — so the human doesn’t have to.