A team can fill a shared drive with AI-written memos, drafts, and summaries by lunchtime and still leave the business exactly where it was on Monday.
The OECD’s 2026 productivity compendium points to analysis of 12,000 firms in 27 European Union countries and finds that AI adoption was tied to a four percent lift in labor productivity over the short run. The report then puts the brakes on the celebration: the gains were uneven by sector, hard to measure cleanly, and in some tasks came with weaker skill growth or worse quality. Faster output is not better output; it is often just faster output.
The dashboard likes the wrong numbers
Most workplaces still reward the easiest thing to count. If a team produces 80 client notes instead of 40, closes more support tickets, or ships more first drafts, the chart goes up and somebody gets to say the word productivity without blushing. AI is perfect for this sort of theater. It can turn one person into a small factory of text, slides, and polite-looking nonsense.
The problem is that the dashboard usually stops at the first pass. It counts the document, not the hour spent checking whether the document says something false. It counts the ticket closed, not the follow-up message from the customer who found the mistake. It counts the draft, not the two rounds of cleanup needed to make the draft usable. AI raises the volume, then quietly hands the bill to the same humans who were supposed to be getting more efficient.
Two teams can look equally busy and be doing completely different kinds of work. Team A uses AI to push out more marketing emails, reports, or client updates. The numbers are handsome. The folder is full. The trouble starts when people downstream have to verify every claim, fix every dodgy reference, and rewrite every paragraph that sounded fluent and was wrong. Team B ships fewer pieces, but it cuts error rates, trims delay, and brings down the number of complaints. One team looks productive on a chart. The other team actually makes the machine behind the chart behave better.
A legal team makes the contrast brutal. A high-volume AI workflow might draft 50 contracts a day, yet 15 percent of them come back for serious repair because clauses are off, details are missing, or the language is simply unsafe. A slower team might draft 20 contracts, but fewer than 1 percent need major rework. The first group looks busier. The second group costs less to clean up.
AI writes faster than humans can trust it
The hidden tax is not just in fixing errors after the fact. It also shows up in the time people spend checking work before they can trust it at all. A 2023 working paper by Erik Brynjolfsson, Danielle Li, and Gabriel Raymond found that human editors spent 40 percent more time editing AI-generated text than writing from scratch, mostly because they had to fact-check and correct the machine’s output. The excited demo never mentions this part.
Once you see that pattern, a lot of office AI hype starts to look thin. The first draft arrives quickly, then the human reader has to interrogate it line by line. Is the number right? Is the tone appropriate? Did it invent a policy? Did it miss the one detail that makes the whole thing useless? The machine saves time on composition and spends it on supervision. Sometimes that is still a net win. Sometimes it is just a more glamorous version of extra work.
The damage is not limited to the document in front of you. Junior staff learn from repetition. If AI keeps doing the first pass, the junior writer never learns how to build an argument, the support agent never learns how to handle a messy complaint, and the analyst never learns how to tell a shaky conclusion from a solid one. The organization gets more output today and thinner judgment tomorrow. This is a bad trade if you need people who can think when the template breaks.
AI makes it easy to confuse movement with progress. A person who produces five polished drafts in the time they used to produce two looks like a gain. But if the five drafts need heavy correction, and the old two were good enough to send, the business has not become more productive. It has become more active. These are different things.
Measure the mess, not the shimmer
If you want a better test, stop asking how much AI output you got and start asking what it cost to make it usable. The useful metrics are unglamorous, which is usually a sign that they are honest.
Track these instead:
- error rate per document
- hours spent on rework
- number of corrections before approval
- customer complaints tied to AI-generated material
- compliance failures or legal escalations
- time from first draft to final sign-off
If those numbers rise, the shiny output count is a trap. If they fall while quality holds steady, you have something worth keeping. That sounds obvious, but a lot of organizations are still congratulating themselves for speed while paying the hidden invoice somewhere else in the workflow.
The best teams will probably not be the ones that produce the most AI-generated material. They will be the ones that learn where AI really saves labor and where it just moves labor into a darker room. Some tasks will tolerate the trade. Others will not. A support script that slightly improves response time is one thing. A contract, a medical note, or a client-facing financial summary that forces three people to clean up the wreckage is something else entirely.
The real question is not whether AI can make staff look busier. It clearly can. The real question is whether your organization has started measuring the time, trust, and reputation it burns while everyone admires the speed.
