You had a great day. Cleared the inbox. Responded to 40 messages. Used AI to draft three reports. Pushed through five meetings. Closed 12 tasks.
You also can't remember the last hour, your head hurts, and making one more decision — even "what should I eat?" — feels impossible.
Here's the uncomfortable finding: the days that feel most productive are often the most cognitively expensive. And the gap between "felt productive" and "was effective" is wider than most people realize.
The 39-point gap
Researchers at METR studied developers using AI coding tools and found a striking pattern: afterwards, participants estimated AI had made them 20% faster. Measured, they took 19% longer (METR, July 2025, 16 experienced developers).
That's a gap of roughly 39 points between perception and reality. Not a small discrepancy — a fundamental mismatch between the felt experience of productivity and the measured output.
The reason is instructive. AI tools remove certain types of work — boilerplate, formatting, lookup tasks. These are visible, tangible completions. You can see the output. Your brain registers each one as a "got it done" signal.
What AI adds is less visible: evaluation decisions. For every AI suggestion, you make a micro-judgment — accept, reject, modify, or verify. BCG's 2026 survey found that workers with high AI-oversight demands reported 12% more mental fatigue than workers with low oversight demands. Each evaluation is individually small. Cumulatively, they're enormous.
So a day with AI assistance produces more visible output (feels productive) while consuming more cognitive capacity (is actually draining). The output counter goes up. The decision battery goes down.
Why slow days feel better
Compare that to a day where you worked on one thing. Maybe you spent three hours on a single document. You produced less visible output. But the cognitive profile is different:
- Fewer decisions per hour. You're in one context, making judgment calls about one problem.
- Lower ambiguity load. You understand the problem deeply. Each decision is clearer.
- Less evaluation overhead. You're generating, not reviewing. Creating is cognitively cheaper than evaluating someone else's creation (or an AI's).
- More flow state. Flow (Csikszentmihalyi, 1990) depends on sustained, uninterrupted attention; the often-quoted figure of 15–20 minutes to get there is a popular rule of thumb, not a measured finding. Multi-task "productive" days break this constantly.
The result: you feel less productive (because the output counter is lower) but are actually in better cognitive shape at the end of the day. Your decision-making capacity isn't depleted. You can still decide what to have for dinner.
The decision-load explanation
This isn't just perception bias. It's a measurable phenomenon with a specific mechanism.
Decision load — the cognitive weight of all your unresolved judgment calls — has five dimensions:
- Volume — how many decisions are queued
- Ambiguity — how unclear the options are
- Reversibility — how permanent the consequences feel
- Interdependence — how many other decisions depend on this one
- Time pressure — deadline proximity vs bandwidth
A "productive" day with AI typically scores high on Volume (many micro-decisions) and low on Ambiguity (each one is small and clear). The total load is high because the volume is extreme — even though each individual decision is easy.
A "slow" day working on one complex problem scores lower on Volume but may be higher on Ambiguity or Reversibility. The total load is often lower because there are simply fewer decisions, even though each one is harder.
This is the intuition behind decision-fatigue research (Baumeister et al., 1998): volume matters, not just difficulty. Treat it as a working idea rather than settled science — a large 2016 multi-lab replication (Hagger et al.) failed to reproduce the core ego-depletion effect.
Three things that help
1. Track decisions, not tasks.
At the end of the day, don't just count what you produced. Count the decisions you made. A day with 200 micro-decisions (AI review, email triage, meeting responses) is more depleting than a day with 10 substantive decisions, even if the productivity app says you "did less."
The shift: when you see a high-decision day coming (meetings, AI-heavy work, inbox processing), deliberately schedule fewer decisions for the following morning.
2. Batch AI review into blocks.
Don't context-switch between creating and evaluating AI output. Set a dedicated block for "AI review" — process all suggestions, drafts, and outputs at once. This reduces the per-decision setup cost and preserves creative capacity for your own work.
The research behind this: Leroy (2009) showed that each task switch leaves "attention residue" — cognitive fragments from the previous task that impair the next one. Switching between "create" mode and "evaluate AI" mode every few minutes maximizes this residue.
3. Protect one "slow day" per week.
Block one day (or even a half-day) with no meetings, no AI tools, no inbox processing. Work on one thing. The output will look low on any dashboard. The cognitive recovery will make the rest of your week measurably better.
This is not a productivity hack. It's cognitive maintenance — like sleep, but for your decision-making apparatus.
The measurement problem
The core issue is that we measure the wrong thing. Productivity tools measure output: tasks completed, messages sent, tickets closed. Decision load is invisible to every tool in your stack.
A day that looks great in Jira, Notion, or your time tracker may be the day that depleted your capacity for the rest of the week. A day that looks "unproductive" may be the day that preserved it.
Until we measure decision load directly — not as a proxy for stress or workload, but as its own dimension — we'll keep optimizing for the wrong metric and wondering why "productive" days feel terrible.
This assessment tool has been retired
It was part of our original decision-load research. Our current work applies the same governance thinking to how organizations run autonomous AI agents.
Take the Governance Stress TestSources
- METR (July 2025) — "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (arXiv:2507.09089). Participants believed AI made them 20% faster; measured, they were 19% slower.
- BCG (2026) — "When Using AI Leads to Brain Fry." Harvard Business Review. Workers with high AI-oversight demands reported 12% more mental fatigue than those with low oversight demands (survey of 1,488 US workers).
- Baumeister, R. F., et al. (1998) — Ego depletion: Is the active self a limited resource? Journal of Personality and Social Psychology, 74(5), 1252–1265. Foundational ego-depletion study; the effect has not held up well in large replications (Hagger et al., 2016).
- Leroy, S. (2009) — Why is it so hard to do my work? The challenge of attention residue when switching between work tasks. Organizational Behavior and Human Decision Processes, 109(2), 168–181.
- Csikszentmihalyi, M. (1990) — Flow: The Psychology of Optimal Experience. Foundational flow-state research.
- Kotler, S. (2014) — The Rise of Superman. Modern flow-state mechanism research.
AI-assisted and human-reviewed. Research cited from peer-reviewed and industry-published sources. Measurement, not treatment.
Corrections, 28 September 2026: An earlier meta description said AI users “feel 75% faster”; METR found developers believed AI made them 20% faster while measured they were 19% slower. The BCG 12% figure compares workers with high and low AI-oversight demands, not AI-assisted with manual work. We relabeled the 15–20 minute flow figure as a rule of thumb and noted the replication debate over ego depletion.