Does AI actually make knowledge workers more productive?

AI and Knowledge Work: What 2026 Evidence Shows

Sometimes, and far less reliably than the marketing claims. A 7,000-workplace Danish study found near-zero effect on earnings or hours, while BCG consultants gained sharply on suited tasks. Ethan Mollick's jagged frontier explains it: AI roughly triples output on about a third of tasks and quietly degrades others. The net gain depends on whether you can tell which is which.

Whenever someone asks me this expecting a clean yes, I tell them the data refuses to cooperate, and that refusal is the most useful thing about it. The honest 2026 answer is conditional: AI makes some knowledge workers much more productive on some tasks, makes others measurably slower, and at the level of whole organizations the effect so far is close to invisible. Three findings, taken together, draw the real shape.

Start with the optimistic one. The Boston Consulting Group and Harvard study, published in Organization Science in March 2026, randomized 758 consultants and found that those using GPT-4 on tasks inside the model's competence finished faster and produced higher-rated work, with the largest lift going to lower performers. AI compressed the gap between weak and strong. That's a genuine effect, and it's the headline most people stop at.

Now the sobering one. Economists Anders Humlum and Emilie Vestergaard tracked AI chatbot adoption across roughly 7,000 Danish workplaces and titled their paper, pointedly, "Large Language Models, Small Labor Market Effects." Despite heavy adoption, they found no significant impact on earnings or recorded hours in any occupation, with confidence intervals ruling out effects above one percent. Measured time savings landed near three percent, and workers reabsorbed more than eighty percent of those reclaimed minutes straight back into other work, including new work the AI itself created, like editing its output. The micro-gains were real and the macro-result was nothing.

Then the cautionary tale about measurement itself. METR's randomized trial on experienced open-source developers reported in 2025 that AI made them nineteen percent slower, even as those same developers were convinced it had sped them up by twenty. That perception gap is the story. But by February 2026 METR walked the number back: their follow-up was crippled by selection bias, because so many developers flatly refused to work without AI even at fifty dollars an hour, and many avoided exactly the tasks where AI helped most. Their current read is that AI probably does help in early 2026, though they no longer trust their own clean estimate. When the best measurers in the field publicly say they can't measure this reliably yet, treat any confident percentage in a vendor deck as marketing.

What ties these together is the idea Ethan Mollick named in Co-Intelligence: the jagged frontier. AI's abilities don't form a smooth line where everything gets a little easier. They're jagged, brilliant at drafting, summarizing, and first-pass research, abruptly poor at tasks needing real-world context, judgment under ambiguity, or accountability. The evidence backs this precisely: AI roughly triples output on about a third of knowledge tasks and adds almost nothing, sometimes subtracting, on the rest. Productivity therefore isn't a property of the tool. It's a property of how well a given worker maps that frontier, sensing when to lean in and when the cleanup will cost more than the draft saved.

This is also why the developer perception gap matters so much to me as a coach. Kahneman's Thinking, Fast and Slow predicts it exactly: a fluent AI response feels like progress, and that feeling is generated by System 1 before you've verified a thing. Workers who feel twenty percent faster while being slower aren't lying; they're trusting fluency over outcome. The high performers I work with close that gap by tracking results, not vibes, and that discipline is part of why the gains compound for some and evaporate for others.

So my practical answer is this. Don't ask whether AI makes knowledge workers productive; ask which tasks, which worker, and measured how. For yourself, audit where it genuinely helps over a few weeks rather than trusting the buzz, and protect the deep, judgment-heavy work it can't touch, the kind Cal Newport built Deep Work around. The deliberate practice of finding where you create real value, which I explore in why exploration drives success, matters more now, not less, because AI floods the easy half of the work and leaves the hard, defining half entirely to you. The productivity is real. It just isn't automatic, and it isn't evenly distributed.