A new Workday survey of thousands of employees landed this week, and one number in it is going to bother you the rest of the day.
85% of people say AI saves them 1–7 hours a week. That part you've heard before. That's the headline. That's the brochure.
Here's what's buried in paragraph four: 37% of that time is immediately spent correcting AI errors.
Run the math. If AI saves you 5 hours a week, you're spending about 2 of them fixing what it got wrong. Your real gain is 3 hours — not 5. And nobody told you about the other 2 when they sold you the subscription.
There's a word for this. Researchers are starting to call it the AI Tax.
You're Not Imagining It
If you've ever thought I spend more time managing this tool than it saves me, you weren't being pessimistic. You were doing the math correctly when no one else was.
The AI Tax isn't a bug in one tool. It's a structural reality of how generative AI works. The output is impressive. The output is also imprecise, context-blind, occasionally confidently wrong, and in constant need of a human who actually knows the business to review it.
That human is you.
Every AI draft needs a read. Every AI customer response needs a check. Every AI-generated summary needs to be compared against what actually happened in the meeting. Every AI recommendation needs to be measured against the specific constraints of your specific situation — the ones the AI doesn't know about and can't know about without you telling it.
That's not laziness on the AI's part. That's the nature of the technology. Generative AI is a probability engine producing its best statistical guess. Your job is to catch the guesses that are wrong.
And that catching? That's the AI Tax.
The Problem With the ROI Calculation
Most small business owners calculate AI ROI the same way: How much time did this save me?
That's the right question, but it's only half of it.
The full question is: How much time did this save me, minus the time I spent correcting, verifying, reviewing, and fixing what it produced?
Those two numbers are different. Often significantly different.
Worse: the correction cost is invisible. When you write the first draft, you know exactly how long it took. When you review an AI draft, fix three things, and move on, the time doesn't feel like a real cost — it feels like editing. But editing time is work time. If you're spending an hour a day reviewing AI outputs, that's 5 hours a week, 250 hours a year. That's a real number. It belongs in the calculation.
The Tax Is Worst in the Middle
Here's the pattern most business owners eventually notice: the AI Tax hits hardest when you're partway in.
At the beginning — you're learning the tool, you're impressed, you're finding things it does well. The tax feels manageable because the wins are obvious.
At the end — once you've trained a custom workflow, given the tool good context, built prompts that reliably produce usable output — the tax shrinks dramatically. The good prompts get the right outputs most of the time.
The brutal zone is the middle. You're past the honeymoon. You've discovered the failure modes. You haven't yet built the workflows that mitigate them. The AI is producing inconsistent output and you're spending significant time sorting through it.
Most people quit in the middle. They cancel the subscription at week 3 or week 6, right before the tool would have started paying off. They leave with a vague sense that it didn't work, when what actually happened is they stopped before the tax got cheaper.
How to Think About This
None of this means AI isn't worth it. The net gains are real for most use cases, even accounting for the tax.
But you need to calculate net, not gross.
A few honest questions to run through your stack right now:
What's my actual error rate on this tool's output? If you're correcting 1 in 3 outputs, that's a 33% overhead. If you're correcting 1 in 10, that's 10%. Those are very different tools.
Am I in the learning curve, or have I plateaued? If your error rate has been roughly the same for 2 months, the tax isn't going to get cheaper with more time. Time to rethink the workflow.
Which tasks have I actually stopped doing because of AI, versus which tasks now have an AI step added? There's a meaningful difference between AI replacing your work and AI becoming another layer of your work. The latter often increases the total.
If AI disappeared tomorrow, would I have more free time or less? The answer is usually more revealing than people expect.
The Businesses Winning This
Here's what's different about the small businesses that are genuinely ahead on AI: they don't measure AI by how impressive it is. They measure it by what they stopped doing.
Automation that removed a full task from their week — something they no longer touch at all, not something they review and approve. That's where the real gains are. Not "AI drafts it, I check it" but "AI does it and I never see it."
The path there is harder. You need reliable inputs, clear processes, and enough trust in the output to actually let go. Most businesses aren't there yet. But the ones chasing that — not the impressive demo, but the thing they can actually stop doing — are building an AI stack that earns its keep.
The AI Tax isn't going away. The technology produces impressive outputs at scale and imprecise outputs at the margins, and humans will be reviewing those margins for a while.
The best thing you can do is count it. Don't measure AI by time saved — measure it by time freed. The difference is whether you ended up doing less work, or just doing different work.
Different work, at the same pace, at $20 a month, is a subscription. It's not a transformation.
Know which one you bought.