Why AI Can Leave You More Tired: Decision Density and Cognitive Load

Why does working with AI leave people more tired?

Working with AI can feel more exhausting not because there is more to do, but because there are more judgment calls to make in a day. Several friends asked me the same question independently: why does using AI leave me more tired? I felt the same way. Work was getting done faster and output was up, yet by the end of the day my mind never seemed to have stopped. I came to think that fatigue may come less from how much you do and more from how many times a day you have to decide something.

Writing a document used to mean spending most of the time gathering, organizing and producing. Now AI returns a first draft in minutes, sometimes several versions at once. That looks like time saved, but every output comes with a string of questions. Is it accurate? Is it usable? Which version is better? How much needs to change? Should we move to the next step? Each one is a small decision. Across a full working day, they arrive one after another. AI can hand over the next result almost instantly; a person still has to review, choose and commit.

A slide shared by a senior peer in an HR and AI discussion group laid out the chain reaction: higher decision density, shorter recovery windows, accumulating cognitive fatigue, and judgment that becomes harder to exercise. One line from that slide stayed with me, roughly: manage the inflow and outflow instead of blaming the reservoir. It led me to a sharper question about AI adoption. Organizations count hours saved and output gained. Does anyone count the work that still requires a human to review it, or look at the pace at which one person is making those calls?

Making the rhythm visible

To explore that question, I built Decision Pulse. It analyzes the day's AI conversation logs stored on your computer and produces a single-page dashboard. You can see when you were responding to AI most intensively, how many conversations you were switching between, how long you went without a break, and whether you tended to fire off the next instruction shortly after a long AI response.

I tested it on my own records for September 22. That day I spent about three hours actively working with AI and sent 144 messages, an average of roughly 47 per active hour. The busiest stretch was 2 to 3 p.m., with 49 messages in the hour. Between 2:15 and 2:28 p.m., a 13-minute window, I switched between five conversations eleven times. Before seeing the data, I simply felt I had done a lot that day. Afterward I realized the work had not been sequential. I had been carrying several unfinished threads at once, circling back to check one, then making the next call on another.

The numbers have limits. One message is not one decision, and a quick reply does not mean a poor judgment. I cannot tell from the data whether my decision quality declined. What the tool does is surface patterns of dense responding and switching that are normally hard to notice, which was enough to make me ask whether I had been tracking how much review work I was absorbing as AI sped up.

How should workload be measured after AI adoption?

Workload should include not just output, but the time spent checking, comparing, revising and owning the result. As someone working in HR, I want this question in every AI adoption discussion. If someone used to produce ten pieces of content and AI now helps generate more versions, production time may drop. The checking, comparing, editing and accountability for the result still sit with that person. Looking only at output volume misses that part of the job.

I see Decision Pulse as a starting point, beginning with my own data, for individuals and teams to talk more concretely about two things: which judgments genuinely need a human, and whether time spent reviewing AI output is recognized as part of the work. The tool currently reads Claude Code and Codex conversation logs; the skill can be installed in Claude Code, Codex, GitHub Copilot, Gemini CLI and Cursor, and the data stays on your own machine. Once installed, asking your AI tool to check on your day generates the dashboard.

Three things to take away

1. Think back to a recent draining day and count how many decisions you made after AI handed you a result, not just how many tasks you finished.
2. When assessing the value of AI for a team, include the time spent checking, comparing, revising and owning the outcome as part of the workload.
3. Review your own response density and conversation switching once with a tool like Decision Pulse, and use what you find to open a team conversation about which judgments must stay with people, not to grade individual performance.