The Verification Tax

Foxit commissioned a study of 1,400 professionals across the US and UK. The sample included 1,000 desk-based workers and 400 senior executives. Both groups were asked to estimate how much time AI saves them and how much time they spend verifying AI output.

The executives: 4 hours 36 minutes saved per week, 4 hours 20 minutes spent verifying. Net gain: 16 minutes.

The workers: 3 hours 36 minutes saved per week, 3 hours 50 minutes spent reviewing. Net loss: 14 minutes.

Sixteen minutes of executive gain. Fourteen minutes of worker loss. That is the real productivity story of AI in 2026, compressed into two numbers that almost nobody is discussing.

The Hierarchy of Illusion

The executives in the Foxit study are not lying. They perceive real gains, and from where they sit, the perception makes sense. Executives interact with AI output after it has been cleaned, verified, and presented by their teams. They see the finished product. They do not see the verification loop that produced it.

Workers see the loop. They live in it.

As we explored in McKinsey Measured the Wrong Thing, controlled trials consistently contradict executive perception of AI productivity. The METR trial showed developers 19% slower with AI tools while believing they were 24% faster. The Foxit data adds a new dimension: it quantifies where the time goes. Not into production. Into verification.

The pattern holds across levels but inverts in direction. Executives are farther from the output, so they are more optimistic. Workers are closer to it, so they report what they actually experience. Distance from the work correlates with confidence in the tool.

60% of executives report “high confidence” in AI accuracy. Among workers, 33% say the same. Only 10% of workers say they are “extremely confident.” The people who use the output trust it least. The people who approve budgets trust it most.

The Verification Tax Is Real, but the Exact Numbers Are Soft

A necessary caveat. The Foxit study relies on self-reported time estimates, not time-tracked data. Workers estimating “3 hours 50 minutes of review time” are approximating, not logging. The precise figure of 16 minutes net gain or 14 minutes net loss should be read as directional, not decimal-precise.

The directional finding, however, is strong: the time spent verifying AI output nearly cancels the time AI saves. This aligns with every controlled study published in the past year. The Foxit data is valuable not because it proves a 16-minute number but because it names the mechanism. The cost is not in AI failing. It is in humans checking whether AI failed.

Foxit’s SVP of Product, Evan Reiss, put it directly: “AI accelerates creation, but it introduces new layers of review, fact-checking and correction.” That framing, from a company that sells document AI, is more candid than most vendor research manages to be.

The Cognitive Threshold

BCG published a separate study in Harvard Business Review the same week. The sample: 1,488 US full-time workers. The finding: AI productivity gains exist, but only up to a threshold.

Workers using three or fewer AI tools reported genuine productivity improvements. Workers using four or more tools saw productivity collapse. They reported 14% more mental effort, 12% greater fatigue, and 19% more information overload. BCG calls this “AI brain fry.”

Julie Bedard, BCG’s managing director, described it plainly: “Things were moving too fast, and they didn’t have the cognitive ability to process all information.”

The threshold matters because it contradicts the prevailing enterprise strategy of layering AI tools across every workflow. More tools does not equal more output. After three tools, the cognitive overhead of managing AI across multiple surfaces erodes the productivity that any single tool provides. The verification tax is not just about checking one tool’s output. It is about maintaining coherent judgment across a growing stack of tools, each producing output that requires independent evaluation.

34% of workers experiencing “AI brain fry” reported intent to quit, compared to 25% without it. The productivity tool is becoming a retention risk.

The Restructuring Paradox

Here is where the Foxit data turns from uncomfortable to alarming.

68% of executives report that AI has already triggered organizational restructuring or headcount changes. Not planned. Already happened. 72% say they prioritize retraining programs. 93% say they track return on AI engagement metrics.

Reread those numbers alongside the productivity findings. Organizations that are gaining 16 minutes per week (at best) from AI have already restructured around the assumption that AI delivers transformational productivity. They cut headcount. They redesigned roles. They built measurement systems to track returns. All of this was based on executive perception of value, not measured verification of it.

The 93% tracking ROE is especially telling. They are measuring return on engagement (how much AI is being used) rather than return on verification (how much AI output survives scrutiny). Adoption is the metric. Nobody is tracking the tax.

The Macro Confirms the Micro

If AI were delivering the productivity gains executives perceive, the effect would show up in macroeconomic data by now. It does not.

The Federal Reserve’s February 2025 analysis attributed a 1.1% productivity increase to generative AI. Goldman Sachs, publishing in March 2026, found no meaningful economy-wide relationship between AI adoption and productivity. These are not pessimistic projections. They are measurements of what has actually happened after two years of aggressive enterprise AI deployment.

The macro story and the Foxit micro story tell the same thing from different altitudes. Workers spend their time gains on verification. Organizations spend their efficiency gains on managing the complexity AI introduces. The net contribution, measured at any level above the individual task, approaches zero.

UC Berkeley’s longitudinal research on AI work intensification adds the human cost. AI does not reduce work. It intensifies it. Workers absorb more tasks, switch context more frequently, and rest less, because AI makes each individual task feel lighter even as the aggregate load increases. The productivity shows up in output metrics. The burnout shows up six months later.

What Would Change If We Measured the Tax

The verification tax is not a reason to abandon AI. It is a reason to measure honestly.

An organization that knows its workers spend 3.8 hours per week verifying AI output can make informed decisions about that cost. It can invest in better verification tooling. It can set cognitive load thresholds. It can budget review time into project plans instead of treating it as invisible overhead. It can stop restructuring around perceived gains and start restructuring around measured ones.

An organization that measures only adoption and executive satisfaction will continue to cut headcount, layer on tools, and celebrate productivity gains that exist primarily in survey responses.

Three things would need to change.

Measure the full cycle, not the generation. Time-to-output is half the equation. Time-to-verified-output is the whole equation. If your measurement system cannot distinguish between “AI produced a draft” and “a human confirmed the draft was correct,” you are measuring the wrong thing.

Set tool thresholds deliberately. The BCG data gives a clear signal: three tools, gains; four tools, losses. Treat this as a starting hypothesis, not a universal law. But test it. Measure cognitive load as you add AI surfaces. Stop assuming more tools equal more productivity when the data says the opposite.

Track verification cost as a line item. If 93% of organizations track return on AI engagement, the infrastructure for measurement exists. Add one metric: hours spent verifying AI output per team per week. That single number would transform every AI investment decision in the organization, because it would make the tax visible.

The 16-Minute Question

The Foxit study will be cited by AI optimists as proof that AI delivers net positive results (16 minutes!). It will be cited by AI skeptics as proof that AI delivers nothing meaningful (only 16 minutes?). Both readings miss the point.

The point is that organizations are restructuring, cutting staff, and investing billions around a technology whose measured net productivity contribution, at the executive level, is 16 minutes per week. At the worker level, it is negative.

This is not an argument against AI. It is an argument against making irreversible organizational decisions based on unverified assumptions about AI value. The restructuring has happened. 68% already made changes. The measurement that should have preceded those changes has not.

The verification tax is real. It is measurable. And until organizations start measuring it, every AI productivity claim is an estimate from someone who does not pay the tax.

---

This analysis synthesizes Foxit’s State of Document Intelligence report (March 2026), BCG’s AI workplace study published in Harvard Business Review (March 2026), and Goldman Sachs macro productivity analysis (March 2026).

Victorino Group helps organizations build measurement infrastructure that distinguishes AI enthusiasm from AI value. Let’s talk.