In 1865, the economist William Stanley Jevons noticed something odd about coal. As steam engines became more efficient, using less coal per unit of work, Britain’s total coal consumption went up, not down. Efficiency didn’t reduce demand. It unlocked it.

“Jevons paradox strikes again!” Microsoft CEO Satya Nadella wrote on LinkedIn after DeepSeek showed the world that AI could be made dramatically cheaper. “As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can’t get enough of.”

Nadella was talking about demand for AI itself. But there’s a version of this paradox he didn’t mention — one playing out inside every company that’s adopted AI tools. Not demand for AI. Demand for work.

The study that nobody wanted

Researchers from UC Berkeley spent eight months embedded at a 200-person tech company, studying how AI tools changed work patterns. They expected to document productivity gains. Instead, they documented a trap.

AI didn’t reduce work. It intensified it.

The pattern they found was self-reinforcing: AI accelerated certain tasks, which raised expectations for speed. Higher speed made workers more reliant on AI. Increased reliance widened the scope of what they attempted. And wider scope expanded the quantity and density of work.

They called it “workload creep.” It had no natural stopping point.

The researchers identified three mechanisms driving it. First, task expansion: roles blurred as people discovered AI could help them do things outside their job description. Product managers started writing code. Researchers took on engineering tasks. If AI could help you do it, why not add it to your plate?

Second, temporal expansion: work bled into non-work hours. Prompting AI felt easy, lightweight enough that evenings and weekends didn’t quite count as “working.” Unpaid labour disguised as curiosity.

Third, thread multiplication: workers ran parallel AI tasks simultaneously, managing multiple projects where they’d previously focused on one. It felt like momentum. It was actually cognitive fragmentation.

"I shipped more code than ever. I also felt more drained than ever."

Siddhant Khare is a software engineer who builds AI infrastructure. His essay "AI Fatigue Is Real" hit the top of Hacker News — not because it was surprising, but because every developer recognised themselves in it.

“We used to call it an engineer,” Khare told Business Insider. “Now it is like a reviewer. Every time it feels like you are a judge at an assembly line and that assembly line is never-ending.”

Before AI, he might spend a full day in deep focus on one problem. After AI: six problems a day, each one “only takes an hour.” But context-switching between six problems is brutally expensive for the human brain. The AI doesn’t get tired between problems. He does.

This is the shift nobody talks about. AI didn’t just change the speed of work — it changed the nature of work. Writing code gives you flow states. Reviewing AI-generated code gives you decision fatigue. The generative work that people found satisfying got replaced by evaluative work that drains them.

“By Wednesday, I couldn’t make simple decisions anymore,” Khare wrote. “My brain was full. Not from writing code — from judging code.”

The numbers confirm it

An Upwork survey of 2,500 workers found that 77% of AI-using employees said the tools had increased their workload. Not reduced it. Increased it. Nearly half didn't know how to achieve the productivity gains their employers expected.

A study by METR found that experienced developers using AI tools took 19% longer to complete tasks while believing they were 20% faster. People felt more productive while being measurably less productive.

Here’s the gap that matters: 96% of C-suite executives expect AI to boost productivity. 77% of workers say it’s making them do more work. Both can’t be right. The data says it’s the workers.

Why this keeps happening

Jevons had it right in 1865. When you make something more efficient, people don't use less of it. They use more.

Build more roads, you get more traffic. Make flights cheaper, more people fly. Give workers AI tools that halve the time to draft a document, and they don’t go home at 3pm. They draft twice as many documents.

The acceleration trap isn’t a technology problem. The technology works exactly as advertised. It genuinely makes individual tasks faster. The problem is what happens next: the gains get absorbed by expanding expectations rather than protected as genuine time savings.

It’s induced demand, applied to knowledge work. And like induced demand on roads, no amount of extra capacity fixes it. The system just fills up.

Does agentic AI break the pattern?

Most of this research describes AI as an assistant — chat-based tools that speed up individual tasks while keeping the human in the loop for every step. Agentic AI changes the model. Instead of doing the work faster with AI help, you delegate whole tasks to agents and review the outcomes. You're an orchestrator, not a reviewer on an assembly line.

That might break the creator-to-reviewer problem. But it could also raise the ceiling. If chat-based AI let people attempt six problems a day instead of one, agentic AI might let them attempt thirty. The trap doesn’t disappear — it scales. Whether it breaks or intensifies depends entirely on whether management redesigns expectations to match.

This is a management problem

The research all says the same thing: AI intensification is a management and organisational design failure, not a technology failure.

People got faster tools and the same job descriptions, the same meeting schedules, the same performance metrics. Of course work expanded. Nobody redesigned the system around the new capabilities.

If AI makes someone’s core work take four hours instead of eight, the answer isn’t “fill the other four hours.” It’s “use the other four hours for the deep thinking that makes the four hours of output better.” The moment you measure productivity by volume, Jevons wins. Measure outcomes instead: decisions made, problems solved, quality shipped.

Khare’s burnout came from context-switching, not from working hard. Six problems a day is six shallow passes. One problem a day is one deep solution. AI should enable depth, not breadth. And if your output is up 40% but your team is burning out, you haven’t gained anything. You’ve borrowed from the future.

Ninety-six percent of executives expect productivity gains. Seventy-seven percent of workers say their workload increased. Someone is wrong. The first step is admitting who.

The takeaway

AI makes tasks faster. But "faster tasks" and "less work" are not the same thing — and the gap between them is where burnout lives.

Jevons figured this out about coal in 1865. We’re learning it about knowledge work in 2026. The efficiency gains are genuine. The question is whether your organisation captures them as genuine improvements in people’s working lives, or whether they just become fuel for the next acceleration cycle.

The tools aren’t the problem. The system around them is.