The Verification Tax: Why AI Time Savings Shrink Once Someone Checks the Work

By Victor Hoang · Updated Oct 10, 2026

Ask a team how much time AI saves and you'll get a confident number. Ask who reads the output before it goes anywhere, how long that takes, and how often it comes back wrong, and the number gets quieter. That gap is the question every executive funding AI should be asking: how much of the time we think we're saving survives once a person has to check the work?

Our view is simple. Every AI output carries a verification tax: the time someone spends reviewing, correcting, and sometimes redoing it. Headline savings are gross figures. What the business keeps is the net, and the net depends far more on the kind of work than on the quality of the tool. Leaders who measure net savings, and who point AI at work where checking is cheaper than doing, will see real returns. Leaders who accept gross numbers will fund a lot of activity that nets out near zero.

Key Facts

• In a January 2026 survey of 3,200 AI-using employees commissioned by Workday, nearly 40% of AI time savings were lost to rework, meaning correcting errors, rewriting content, and verifying outputs. Workday sells AI products, so treat it as a vendor survey.
• In the same survey, only 14% of employees consistently got clear, positive net outcomes from AI, and 77% of daily users reviewed AI work as carefully as human work, if not more.
• In a randomized trial, 16 experienced open-source developers took 19% longer on 246 tasks when AI was allowed, though they had forecast a 24% speedup and afterward believed they'd been 20% faster.
• In a BCG experiment with 758 consultants, AI users on a task outside the model's capability were 19 percentage points less likely to produce a correct solution than those without AI.
• A 2025 BetterUp and Stanford Social Media Lab survey of 1,150 US desk workers found 40% had received "workslop" in the past month, at an average of two hours per incident and $186 per employee per month.

Where the Savings Go

The pattern across very different studies is consistent enough to take seriously. Time moves from creating to checking, and checking is not free.

Start with the survey evidence. Workday's research, fielded by Hanover Research in November 2025, found that 85% of employees report saving one to seven hours a week with AI. Those are self-reported. The same study says nearly 40% of the savings leak back out through rework. That is a vendor survey, and the exact percentage will vary by company, but the direction fits the controlled studies below: the first draft is fast, and the editing is where the afternoon goes.

Now the controlled evidence. The METR trial is the cleanest look at what happens when you measure rather than ask. Sixteen experienced developers working on their own mature codebases completed 246 tasks, each randomly assigned to allow or disallow AI. AI-assisted tasks took 19% longer. The more useful finding for executives is the perception gap: developers expected a 24% speedup and, after finishing, still believed they'd gained 20%. People were wrong about their own time, in the flattering direction.

Two cautions belong here. The study used early-2025 tools, and METR has since said it is likely that developers are more sped up now, while calling its own newer data only very weak evidence because developers increasingly declined to work without AI. So don't read the 19% as a permanent law. Read it as proof that felt speed and measured speed can diverge, and that nobody should trust the felt number alone.

The Noy and Zhang experiment points at the mechanism. In a preregistered study of 444 college-educated professionals on writing tasks, ChatGPT cut time taken and raised quality, and it restructured tasks toward idea generation and editing and away from rough drafting. That's the tax in miniature. The drafting cost fell, and the editing share of the job rose.

When Checking Costs More Than Doing

The most important result for deciding where to deploy AI is the one that looks like bad news. In the Harvard and BCG experiment, consultants on tasks inside the AI's capability completed 12.2% more tasks, 25.1% faster, at more than 40% higher quality. On a task deliberately placed outside that frontier, AI users were 19 percentage points less likely to be correct. The authors call this a jagged frontier: some tasks are easy for the model, and others that look equally hard are not.

The consultants who failed outside the frontier weren't careless. The output was fluent, and fluent output is cheap to produce and expensive to doubt. Verification cost isn't uniform. It depends on whether the checker can tell a right answer from a wrong one quickly. That is a property of the task, not the tool.

This is why we think of verification as a ratio. For any task, compare the cost of checking an output against the cost of producing it yourself. If checking takes a fifth of the time doing would, AI wins even when it's wrong some of the time. If checking requires redoing the work to know whether it's right, you've paid for the draft and for the job, and you've added the risk of a plausible error slipping through.

The BetterUp and Stanford survey shows what happens when the tax is pushed onto someone else. It defines workslop as AI-generated work that looks polished but lacks substance, and 40% of respondents had received some in the past month. The sender saved time. The receiver paid it back, with interest. If your measurement only looks at the sender, the savings are real on paper and negative on the P&L.

A Worked Example (Illustrative)

Take a hypothetical 40-person revenue team. Each rep says AI saves them five hours a week on account research and follow-up drafting. Those numbers are invented to show the arithmetic, not to benchmark anyone.

On these made-up numbers, the net is 40% of the headline, so a forecast that multiplies five hours by 40 reps overstates the benefit by 2.5 times. The point isn't the specific figures. It's that three of the four rows never show up in a tool-usage dashboard or a vendor survey, and each of them is cheap to estimate with a one-week time log.

The categories also differ. Checking account research against the CRM is quick, because the facts are sitting in one place. Checking a drafted renewal-risk analysis is slow, because you'd need to rebuild the reasoning to trust it. Same team, same tool, very different tax.

A Decision Framework: Is Checking Cheaper Than Doing?

Score each candidate workflow on two things. How expensive is it to verify an output, and how expensive is an error that gets through? Then act on the combination.

The pattern is that AI pays where verification is mechanical and fails to pay where verification is a judgment call. This lines up with what we've argued about who in your workforce can actually catch AI's mistakes: the scarce skill is evaluation, and it sits with people who know the domain.

Why Verification Cost Hides

Three things keep the tax out of sight.

It's distributed. One person's saved hour becomes three people's twenty-minute reviews. No single manager sees the whole bill, so every local report looks positive. The same dynamic shows up in coordination cost: work that moves between people carries a cost that isn't on anyone's budget.

It's unmeasured by design. Usage dashboards count prompts, sessions, and seats. Survey questions ask how much time AI saves, not how much time you spend checking it. The METR result suggests we aren't even good at estimating our own net time when asked directly.

It shows up as quality drift, not hours. If reviewers get tired and start approving by glance, the tax doesn't vanish. It converts into errors that reach customers. Workday found that 77% of daily users review AI work as carefully as human work, if not more, which is a healthy instinct, and also a clear sign of how much review time is being spent.

What to Do Next

For a CEO, COO, or CFO who has approved AI spend on the strength of time-saved estimates, here is a sequence that fits a quarter.

1. Change the metric. Report net hours saved, defined as gross hours saved minus review, correction, redo, and time spent handling colleagues' AI output. Our piece on measuring AI ROI beyond time saved covers the broader scorecard.
2. Run a one-week time log on three workflows. Pick one that's easy to verify, one that's ambiguous, and one that's high-stakes. Ask people to record time to draft, time to check, and time to fix. A rough log beats the most polished estimate.
3. Sort every AI use case with the table above. Fund the ones where checking is mechanical. Put the ones where checking requires an expert behind a pilot with a defined exit.
4. Make checking cheaper. Require sources with claims, keep source documents next to the output, and build tests wherever the work allows them. Every minute you take off verification is a minute of savings you keep.
5. Set a no-workslop norm. The sender owns the quality of what they pass on. If you govern AI use only at the data layer and not the output layer, the cost lands on the receiving team.
6. Re-measure when the model changes. The frontier moves, and so does the tax. A workflow that was too risky six months ago may be cheap to verify now, and the reverse can be true too. Build the re-measure habit into a strategy that survives model churn. And remember that the cheapest check is one made against rules that are written down, which is why undocumented know-how is so often the hidden source of the tax.

None of this argues against AI. Where output can be checked cheaply, the evidence of real gains is strong, and the BCG results on in-frontier tasks are large. The argument is for counting honestly, so the savings you report are the ones you actually keep.

Learn More

• Measuring AI ROI Beyond 'Time Saved'
• AI Copilots vs. AI Agents: Understanding the Difference Matters
• The Governance Gap: What Leaders Get Wrong About AI at Work
• Move Now or Wait for AI to Mature?
• Does AI Favor Generalists or Specialists?