AI arrives with a second bill. It doesn't show up in the business case, and in most teams it lands on the same few people.
85% of employees say AI saves them between one and seven hours a week. Then roughly 37% of that time goes straight back into correcting, clarifying, and rewriting what it produced. Workday surveyed 3,200 leaders and employees in November, with Hanover Research running the fieldwork1.
Workday puts it as a ratio. For every ten hours of efficiency gained through AI, nearly four are lost to fixing the output.
The license was in the business case. The four hours weren't.
Because the four hours aren't spread evenly. Only 14% of employees say they consistently get clear net-positive outcomes from using AI. The heaviest users carry the most: 77% of daily users review AI-generated work at least as carefully as work done by a colleague, and the most engaged employees lose around a week and a half a year to rework1.
Put names to that. The people paying the second bill are the ones who notice when something is wrong, so the cost concentrates on whoever you can least afford to slow down. It never appears as a line item. It appears as those people being slightly less available than they used to be.
The same pattern shows up in observed behavior rather than self-report. ActivTrak's Productivity Lab looked at 443 million hours of workplace activity across 1,111 companies, then compared a subset of 10,584 people for 180 days before and after they started using AI. Every category of work went up. Email time rose 104%, messaging 145%, business management tools 94%. Average daily focused time fell by 23 minutes2.
AI didn't reduce the work. It changed the shape of it, and the new shape has more coordination in it.
Only the first one looks like AI work. The other two look like a busy week, so they seldom get traced back to the tool that created them.
| In The Business Case | On The Second Bill |
|---|---|
| Licenses and seats | Hours spent checking output |
| Time saved per user | Time returned to rework |
| Training rolled out at launch | Ongoing coaching as the work changes |
| Faster output | More coordination around that output |
| Gains spread across the team | Cost concentrated on the heaviest users |
Three things, and the order matters.
Protected time for the checking. If verification is real work, put it on the calendar as real work rather than expecting it to happen in the gaps. A team that has time to check produces less rework downstream.
Coaching for the people carrying the most. Not tool training. Judgment: what to trust, what to verify closely, and when the faster path isn't worth it. Workday found two thirds of leaders calling skills training a top priority, and fewer than four in ten of the people carrying the most rework saying any of it has reached them1. That gap is fixable and it's cheap.
A place to put what gets learned. When someone works out that a particular task needs heavy checking and another doesn't, that finding should outlive the person who found it. Otherwise every new user pays the same tuition.
Find the person on your team who uses AI the most, and ask them for three numbers. Roughly how many hours a week it saves them. Roughly how many hours go into checking and fixing. And what they've stopped doing to make room for the second number.
The third answer is the one worth writing down. Something always gets dropped, and it's usually the work with no deadline attached: the coaching conversation, the account research, the thing that pays off next quarter.
The second bill isn't a reason to slow down on AI. The time savings are real and most employees can feel them. But a gain you haven't budgeted for gets paid out of somewhere, and right now it's being paid by your strongest people out of hours nobody assigned.
Fund the checking. Coach the judgment. Keep what gets learned so the next person starts further along.
Up Next: Why the Job Description Never Changed
The second bill has a structural cause, and it has nothing to do with the software.
In 89% of organizations, fewer than half of roles have been updated to reflect what AI can now do. People are running current tools inside job structures written years ago — the output got faster, but the role, the expectations, and the review process all stayed where they were. The next post covers what to change in the role rather than in the software.
Read the Next Post →Sources
Numbered citations throughout this article draw on the sources above.
