Decision Fatigue Is Dead. The Fourth Hour Is Not.
Vishal Sachar
Co-Founder & CEO of CLRT
There is a fair chance your calendar is built on findings that no longer exist. Willpower as a tank that drains through the day. Important decisions before breakfast, because judgment runs on glucose. Work in 90 minute blocks, because the brain cycles that way. Attention spans shorter than a goldfish's. Each of these was sold to executives as settled science, and each has quietly failed replication or been traced back to nothing at all. The collapse is a decade old and it is still barely known outside the journals. What deserves attention is the one finding that survived, because it is more useful than the folklore it replaces, and because the way most companies now work is making it more expensive to ignore.
Start with the centrepiece. Ego depletion, the claim that self-control draws down a single finite resource, sat underneath two decades of productivity advice and several bestselling books. In 2016, 23 laboratories ran the same preregistered experiment on 2,141 participants and measured the effect at d = 0.04, with a confidence interval that comfortably includes zero. That is not a small effect. It is statistically indistinguishable from no effect at all. A second multi-lab attempt in 2021, this time 36 laboratories and 3,531 participants, returned d = 0.06. The glucose version of the story, willpower as blood sugar, fell separately: experiments showed that merely rinsing the mouth with a sugar solution and spitting it out lifts self-control as effectively as drinking it, which no metabolic account can explain. Whatever those original studies were measuring, it was not fuel.
The supporting cast went the same way. The famous hungry judges study, in which Israeli parole boards supposedly granted 65 percent of applications after a meal and almost none before the next one, is now best understood as largely an artifact of scheduling: the order of cases was not random, and simulation work published in Judgment and Decision Making showed the reported magnitude is far larger than hunger could plausibly produce. The 90 minute ultradian work block is sleep architecture stretched onto the waking day; when researchers went looking for a 90 minute rhythm in waking cognitive performance, they did not find one. And the eight second goldfish attention span was never research at all: the BBC traced it in 2017 to a sourceless statistic in a marketing deck. None of this is hidden. It is simply unread, because the folklore is more quotable than the corrections.
What survives is narrower and more useful. Fatigue that accumulates across a working session degrades real decisions, measurably. The cleanest evidence is clinical. Across 204 clinicians and 21,867 visits for respiratory infections, a JAMA Internal Medicine study found the adjusted odds that a clinician prescribes an antibiotic, the easy yes in that consulting room, climb steadily through a clinic session: up 26 percent by the fourth hour relative to the first. Notice what this is not. It is not a tank running dry, and it does not present as stopping. The clinicians kept working at full speed. They drifted toward whatever closed the encounter fastest, interrogating less and defaulting more. There is converging evidence elsewhere, EEG work in 2023 measured videoconference fatigue as a real neurophysiological state rather than a mood, but the clinical result is the one to build on, because it puts a number on judgment decaying by the hour.
The reason this matters more now than it did in 2014 is arithmetic. An executive running an agent-heavy operation does not make fewer decisions than before. They make more. Agents compress the production of work, and every unit of produced work returns as a decision: approve, reject, escalate. Drafting and analysis that used to absorb a team's week now lands on one person's afternoon as a queue of things that look finished. If judgment decays across a session, and the volume reaching the session is multiplying, the binding constraint on an agent-run company is not model capability and it is not headcount. It is the number of genuinely good judgment hours available in a day, and what is allowed to consume them. Most calendars spend those hours by accident, in whatever order the queue arrived.
The folklore was wrong about the mechanism, not about the decay. Judgment fades by the hour, and agents are multiplying what reaches it.
A deeper dive
It is worth being precise about how the old canon failed, because the failure pattern is a usable instrument. The mouth rinse result is the elegant one: if self-control ran on blood glucose, a sugar solution swished and spat could not restore it, since nothing meaningful is metabolised in the seconds involved. It restores it anyway, which points to motivation rather than metabolism, a signal that effort is worth spending rather than a tank being refilled. The hungry judges collapse teaches something different. The reported swing was so extreme that hunger would need to be one of the most powerful forces ever measured in behavioural science, and the mundane explanation was sitting in the scheduling data: prisoners without representation tended to be heard later in each session, and favourable cases take longer, so the pattern assembles itself without any biscuit. The general rule is that when a finding is both extremely convenient for a keynote and extremely large, those two properties are usually related. An executive does not need to referee replication debates. They do need to notice that a calendar is a theory of where judgment comes from, and that most executive calendars are still running on a theory that died in 2016.
The clinical finding also tells you exactly what to defend against, because it shows how fatigue presents in a competent professional. The clinicians did not slow down or feel themselves failing. They shifted toward the option that ended the encounter: prescribe, reassure, next patient. Translate that into an executive reviewing agent output and the shape becomes uncomfortable. Agent work product is optimised to look finished. It is fluent, formatted, confident, and wrong in ways only interrogation reveals. An approval requested in your fourth consecutive hour meets precisely the state the antibiotic data describes: a bias toward the easy yes, applied to material engineered to invite it. The defensible design has two layers, and neither is a willpower programme. First, reduce the count. Most decisions that reach a fatigued executive should never have reached a human, because a test, a rule, or a cheap reversal could have settled them, and a verification layer can absorb them wholesale. Second, schedule what remains. The genuinely irreversible calls, the ones where a wrong yes compounds, belong in protected early-session hours, and anything arriving outside that window should default to deferred rather than approved. That is queueing discipline applied to judgment, and it is the part that does not ship with the agent.
Key terms
- Ego depletion
- The claim that self-control draws down one finite resource, so exerting willpower on a task leaves less for the next.
- Preregistered replication
- A study whose method and analysis are locked publicly before any data is collected, removing the flexibility that let the original effects appear.
Work with CLRT
This is a design problem, and it is the kind CLRT takes on. When we build agentic systems for a business, the deliverable is not only the agents. It is the decision architecture around them: which calls the system absorbs, which it stages for a human, and how the irreversible ones are routed to the hours when judgment is actually present. If your operation is growing an approval queue faster than it grows judgment hours, that gap compounds quietly, in fourth-hour yeses nobody remembers making. Talk to us before it does.
Vishal Sachar
Vishal Sachar is the Co-Founder and CEO of CLRT, where he helps UAE businesses make sense of applied agentic AI and put it to work. He writes on agentic systems, AI governance, and the economics of automation. Reach him at vishal@clrtstudio.com or on LinkedIn.
