The assumption embedded in every AI coding tool pitch is that faster equals less work. Write a function in 3 minutes instead of 20, and those 17 minutes belong to you. Harder problems, rest, deeper thinking. That assumption has been tested empirically now, and it is wrong in a specific, measurable way.
In February, Harvard Business Review published findings from an eight-month embedded study at a 200-person tech firm. Researchers Aruna Ranganathan and Xingqi Maggie Ye spent two days a week inside the company while employees adopted AI tools, tracking internal communications and running over 40 interviews. Their finding: AI tools consistently intensified work rather than reducing it. Employees worked at a faster pace, took on broader scope, and extended their hours — often without anyone asking them to.
The researchers called it workload creep.
How It Actually Works
The mechanism is not complicated, but it is easy to miss from inside it.
When AI helps you finish a task faster, the time does not disappear into rest. It gets filled. Sometimes by management: the system just absorbed one hour of your capacity and a new ticket appears. Sometimes by you: the backlog you had been ignoring for months is suddenly tractable because AI can "handle it." Sometimes by scope expansion across roles — product managers in the study started writing their own code, and user researchers picked up engineering tickets. The democratization of skill that AI enables felt empowering initially. What it actually meant was that employees absorbed adjacent work that previously belonged to other people.
The result was not developers doing the same work in less time. It was developers doing more work in the same time, and sustaining that pace until they could not.
A Separate Study Measured What That Does to Your Brain
BCG published research in March tracking 1,488 full-time U.S. workers on the cognitive effects of AI tool use. Their term for the outcome: "AI brain fry" — a specific form of mental fatigue from overseeing AI tools beyond your cognitive capacity.
The numbers are worth sitting with. When workers' AI-related work required high levels of oversight — checking output, managing parallel agents, validating generated code — they expended 14% more mental effort, which correlated with 12% greater mental fatigue and 19% greater information overload. Workers using four or more AI tools saw self-reported productivity drop sharply. Workers using three or fewer reported gains. The marginal tool past three was not adding capability; it was fragmenting attention and compounding the oversight burden.
Among workers who reported AI brain fry: 33% more decision fatigue, 39% more major errors, and 34% actively intending to leave. Those last two travel together badly. The people most cognitively degraded by AI tool overuse are also the ones making the most mistakes.
The Oversight Cost Is Structural
Something the BCG finding captures that most AI productivity discussions miss: AI-assisted work is not just faster work. It is a different kind of work.
When you write code without AI, your attention is on the problem. When you use AI, attention splits between the problem and the output — reading what it produced, checking what it missed, catching the logic error, integrating the snippet into the existing architecture. That oversight cost is real, continuous, and invisible to every metric that measures lines of code or tasks closed.
Developers doing the most AI-assisted work are often not doing less thinking. They are managing multiple parallel threads: their own code, the AI's output, validation of that output, and the expanded scope that the AI's speed has created. The cognitive load shifts and often grows.
This helps explain several things that otherwise look like contradictions. Why developers using AI report feeling busy and exhausted rather than free. Why Faros.ai's analysis of 22,000 developers found that higher AI adoption correlated with 91% longer PR review times and a 9% increase in bugs per developer, even as raw output metrics climbed. Why the HBR researchers found the initial productivity surge consistently gave way to lower quality work and higher turnover.
Individual throughput goes up. Sustainable capacity goes down.
What the Usage Data Shows
We built xeve partly because we were skeptical of self-reported productivity narratives. When you track app usage, coding time, and context switches automatically, patterns emerge that surveys miss.
One of them: developers using AI coding tools heavily tend to have more total active computer time per day, not less. Their focus session lengths are often shorter. The pattern looks like higher activity alongside higher fragmentation — more windows open, more context switches per hour, shorter uninterrupted blocks. That is exactly what the HBR and BCG findings would predict. Faster individual tasks, but more task threads, more oversight, more cognitive load.
A developer who finishes a feature in one hour instead of three does not typically code for one hour and stop. They start another feature, review more PRs, or get handed something the backlog never had capacity for. The speed creates the demand. The demand fills the hours.
The Fix Is Not Fewer AI Tools
Neither study argues for abandoning AI. The BCG researchers are explicit: three or fewer tools, used with intentional norms, produced productivity gains. The problem is treating AI as a throughput maximizer without any corresponding discipline about capacity.
The HBR prescription is what they call an "AI practice" — structured pauses before major decisions, sequencing work to reduce context switching, protecting time for actual rest rather than immediately refilling it. That is not fundamentally different from what you need to protect deep work generally: the absence of norms is what gets you, not the presence of tools.
The measurement implication is the same. If you are using AI coding tools and your self-reported productivity is high but your focus block lengths are shrinking, your daily active hours are growing, and your bug rate is up — those signals matter more than the feeling. The feeling of productivity is the last thing to degrade when workload creep is setting in. You are moving fast right up until the point you are not.
Track output, not activity. Committed code per coding hour. Focus block length over time. Defect rate per feature shipped. These are harder to game than tasks closed or prompts per day. They also tell you when you are in a sustainable state versus a sprint that will eventually collapse.
The developers who will get the most from AI long-term are not the ones using the most tools. They are the ones treating their cognitive capacity as a finite resource — measuring it, protecting it, and refusing to let faster task completion automatically become an expanding workload.
AI did not promise you more work. Your response to it did.