Judges rule differently late in a session, according to the most famous study in this literature. Doctors order fewer cancer screenings at 5pm than at 8am, and write more antibiotic prescriptions in the fourth hour of a clinic than in the first. The clinical measurements hold up. The judicial one has been contested for fifteen years.
The explanation attached to them is that willpower works like a battery and runs down as you spend it. That explanation has a name, ego depletion, and it has mostly failed replication. The largest test of it ran across 23 laboratories and found essentially nothing.
So “decision fatigue” is two claims wearing one label, and only one has survived. This post pulls them apart quickly, then spends most of its length on the part you can act on: one lever if you own a calendar, a different one if you only get to book time on someone else’s.
What decision fatigue actually means
Two things get folded into the phrase, and keeping them apart is most of the work.
| The claim | What it says | Where it stands |
|---|---|---|
| The pattern | Decisions shift in direction as a session wears on, or as more decisions stack up behind you | Measured repeatedly in clinics. Measured once in courtrooms, and disputed ever since |
| The mechanism | A finite mental resource gets spent down, and that depletion is what causes the shift | Failed the largest preregistered test ever run on it |
You can accept the first and reject the second without contradicting yourself.
Be precise about “worse,” too. In none of the studies below did researchers grade individual decisions and mark them wrong. What moved was the direction of the choice: more of one option early in a session, more of another late. That the late choices were worse is an inference, reasonable in the clinical cases and still an inference. If you want a read on how you decide under load rather than a rule about clocks, the decision-making assessment is the better instrument.
The pattern that holds up
Antibiotics rise across the clinic day
“Antibiotic prescribing increased throughout the morning and afternoon clinic sessions for antibiotics sometimes indicated and antibiotics never indicated ARIs.” That sentence comes from Linder and colleagues in JAMA Internal Medicine, reporting on primary care visits for acute respiratory infections, the category where antibiotics most often do nothing.
Comparing the fourth hour of a session with the first, the adjusted odds of a prescription were 1.26 times higher, 95% confidence interval 1.13 to 1.41, P below .001 for the trend. The comparison sits inside one clinic session: same clinicians, same complaint category, different hour. The drift appeared in both categories the researchers separated out, including the one where guidelines say an antibiotic should never be written.
Notice which way it points. Prescribing ends the appointment. Declining takes a conversation about why the patient is going home empty-handed. The later the hour, the more often the clinician took the exit.
Cancer screening falls from morning to evening
Five years later, Hsiang and colleagues in JAMA Network Open ran the clock against a different decision. Breast cancer screening was ordered in 63.7% of eligible appointments at 8am and 47.8% at 5pm. Colorectal screening fell from 36.5% to 23.4% across the same day. Adjusted odds dropped roughly 6% per hour on both orders (odds ratio 0.94 each, confidence interval 0.93 to 0.96 on breast screening). Per hour that sounds tiny. End to end it is 16 percentage points, across appointments all eligible for the same order.
Both studies are observational, and their authors say so. Nobody randomized patients into a 4pm slot, and a clinic running an hour behind at 5pm differs from the 8am version in ways that go past tiredness.
Hold the two results next to each other, because on the surface they disagree. Prescriptions go up as the day runs on. Screening orders go down. If a drained clinician simply made worse calls you would expect one consistent direction of error, and there isn’t one. Park that. It matters later.
The parole study nobody has fully settled
The famous one. Writing in PNAS in 2011, Danziger, Levav and Avnaim-Pesso analyzed 1,112 parole rulings by Israeli judges: “The percentage of favorable rulings drops gradually from ≈65% to nearly zero within each decision session and returns abruptly to ≈65% after a break.”
“This conclusion depends on the order of cases being random or at least exogenous to the timing of meal breaks.” That is Keren Weinshall-Margel and John Shapard in the same journal the same year. Their objection is about scheduling, not psychology. Boards finish one prison’s cases before breaking, and prisoners without counsel tend to be heard last: in their figures, unrepresented prisoners succeeded 15% of the time against 35% for those with a lawyer. The dataset recorded neither representation nor in-prison behavior, which parole law requires the board to weigh.
Then Andreas Glöckner ran simulations in Judgment and Decision Making in 2016. He modeled “a (hypothetical) rational judge who aims to avoid starting work on cases that could not be completed in the time that remains in the current session,” and a decline of similar magnitude fell out on its own. No depletion required.
Danziger, Levav and Avnaim-Pesso replied under the title “Reply to Weinshall-Margel and Shapard: Extraneous factors in judicial decisions persist,” and held their ground. (It sits behind a bot challenge on PubMed Central, so it is named here rather than linked.)
Three arguments over one dataset, unsettled fifteen years on. Glöckner’s title says the magnitude is overestimated, a narrower charge than “wrong.” The reply says the effect persists, which is not the same as saying the explanation does. Nobody has shown the curve comes from depletion rather than from how a docket gets built.
Why the battery metaphor failed the test
Ego depletion got tested properly in 2016. A multilab preregistered replication coordinated by Martin Hagger and Nikos Chatzisarantis ran one standardized protocol across 23 laboratories with 2,141 participants, registered in advance so nobody could tune the analysis after seeing the data. The pooled effect came out at d = 0.04, 95% confidence interval -0.07 to 0.15. An interval that spans zero means the data are consistent with no effect at all.
There is a field test too. Andersson and colleagues went through 231,076 medical triage calls handled by 174 nurses, looking for the signature of decision fatigue across a shift. Their conclusion, in their own words: “We thus found no evidence for decision fatigue.” Bayesian evidence ran above 22 to 1 in favor of no effect on their main tests, a stronger statement than simply failing to find one.
The metaphor survives anyway. It matches what a long day feels like from the inside, and it turns a bad call at 5pm into a resource problem rather than a judgment problem. Before the replication there were hundreds of small studies supporting depletion, which is also what a literature looks like when underpowered studies get published only when they find something.
Any productivity system built on the willpower tank is built on the part that broke.
Where this breaks
Every study above involves someone making dozens of similar, consequential decisions back to back: a full clinic session, a docket of parole hearings, a shift of triage calls. That is not what a manager’s Tuesday looks like. Four meetings, a handful of approvals and one genuinely hard call about a person is a different workload, and there is no reason to assume a curve fitted to those settings transfers onto it.
Nobody has run this study on managers. Not the good version, with shift-level data and a preregistration. So the honest status of “your 4pm people-decision is worse than your 9am one” is: plausible by analogy, unmeasured in your setting.
There is a second gap. The clinical findings show direction shifting, not accuracy dropping. If your 4pm self reliably picks the option that requires less conversation, that only costs you when the low-conversation option happens to be the wrong one. Sometimes it is the right one, and you spent the morning overthinking.
Moving a hard conversation from 4pm to 9am still costs nothing if the effect turns out not to exist in your job.
What managers can do: control the calendar
When managers first try to protect time for decisions, we see the same pattern: they block the calendar but not the sequence. Two hours on Wednesday, no meetings, labeled deep work, and inside it the same order as every other day: email, then approvals, then the performance conversation that has been sliding for three weeks, now at the tail with a 4pm standup pressing on it.
Guarding the quality of your attention inside that block is a separate job, and the harder one. This is the cheaper question of which decision goes first once you are inside it.
Split the decisions on your week into two piles.
One pile closes when you decide. Expense claims, leave requests, small tool spend, sign-off on a doc that already carries a recommendation. Nothing follows, so these can sit anywhere.
The other pile opens something. Putting a report on a performance plan. Choosing between two finalists after the loop. Telling an engineer their promotion case is not going up this cycle. Extending an offer above band. Splitting a team in a reorg. Killing a project someone has spent a quarter on. Each one starts a conversation, a written follow-up, and usually a second meeting with someone who did not like the answer.
Whatever causes the clinical pattern, it lands on the second pile: the option taken late was the one that closed the encounter.
Anything in the second pile takes the first slot of a day, or the first half hour after lunch. Not the tail of a block, not after a run of approvals, not the fourth agenda item. If you have one call this week about someone’s role, promotion or exit, it goes first thing.
Batch the first pile into one window. Expenses, leave, small spend, sign-offs: one block, one pass. Trickling them across the day parks a queue of small decisions between you and the large one.
Some second-pile decisions cannot move. A candidate flies out Thursday evening. The skip-level who has to be in the room is only free at 4:30. Three moves for a fixed slot:
- Write the criteria in the morning, decide in the afternoon. Two lines: what makes this a yes, what makes it a no. The 4:30 meeting then applies a judgment you made at 9.
- Price the effortful option in writing before the meeting. “Option B means I write the plan, tell the team, and hold two follow-ups.” Named in advance, it stops quietly losing for costing more.
- Split the decision at the reversible seam. Settle the half you can undo now, book the half you cannot for tomorrow’s first slot. Naming an interim owner is reversible. Telling someone their role is going away is not.
If you have said “let’s pick this up next week” twice in one afternoon, you are deferring rather than gathering information. Whatever you deferred goes at the top of tomorrow.
The lever here is order, not hours. Same week, same meetings, the one decision that opens something moved to a different position in the run, which is what prioritization looks like on a calendar that is already full. Our work with new managers hits the sequencing question directly.
“I moved it to nine and it went exactly the same way. Twice.” Claire, an illustrative composite rather than a real client, had moved her hardest conversation to the top of the day for a month. It kept going badly because she had not decided what outcome she wanted before walking in. The clock was never her constraint.
Sequence is a narrow fix. It pays on decisions you have already thought through, which is what the rest of a manager’s decision-making practice is for.
What ICs can do: time the ask, not the calendar
Making an ask legible to your manager is its own craft, and we cover it separately. This is the narrower question of where in the day the ask lands.
Go back to the two clinical studies that seemed to disagree. Prescriptions rose across the day. Screening orders fell. Both are the same move. Writing the prescription ends the appointment; ordering a colonoscopy starts a conversation. Late in a session the option requiring less work won in both, and in one that meant doing more, in the other doing less.
Neither study tested that. It is the explanation that fits both results, which is weaker than the explanation being right. What drifts late may not be the answer’s quality so much as which answer counts as the default. That puts it in the same family as the biases that bend a call before anyone notices, with one difference in your favor: this one runs on a clock, so you can see it coming.
You do not control your manager’s calendar. You control which hour your question reaches it, and how much that hour has left to give.
- Read the day before you book into it. Look at what already sits on their Tuesday and count only the meetings that end with someone deciding something: a skip-level, a budget review, a candidate debrief. Status meetings do not count. If your architecture question would be the fourth decision in five hours, ask for Wednesday at 9 instead. If their calendar is private, one message gets you the same read: “what does Thursday look like? I need twenty minutes where you have to make a call, not just hear one.”
- Ask for the position in the run, not the hour on the clock. When someone says “any time Thursday,” that is a real choice, and most people spend it on the first opening they see. An 11:40 slot sits at the end of a four-meeting morning. A 2pm slot sits at the front of a shorter stretch. Same person, same day, different place in the sequence. “First thing after lunch” also survives a day that moves, which “2pm” does not.
- When the only slot is a bad one, cut what the slot has to produce. A 5:30 meeting will carry one job, not two. It will not both build a shared picture of the problem and produce a call on it. So take one of those out of the room: send the picture in writing that morning and leave the late slot only the yes or no, or accept that the late slot is where the picture gets built and book the call itself into tomorrow’s first opening. Asking a tired half hour to do both is how you get the answer that ends the meeting.
- Do not stack two asks into one slot. With a big ask and a small one, the small one is the expensive mistake, because leading with it spends the fresh attention on the cheap decision and hands the big one a manager who has already decided something for you today. Big ask first, small one in writing afterwards. If both are big, they belong on separate days. Two hard calls in one thirty-minute meeting means the second gets whichever answer closes it fastest.
- Batch small approvals into one message. Three pings across an afternoon are three separate decisions, each landing later in the run than the one before it. One message with three lines and a recommended answer on each is a single decision, and it stops your small stuff from queuing up in front of whatever else that person has to decide today.
- Let the calendar pick the hour, not your deadline. Most late-day asks are late because the asker waited, not because the afternoon was the only opening. An ask that lands at four on Thursday for a Friday morning ship has already removed every answer except a rubber stamp or a no. Pick the slot you want first, then work backwards two days and start there.
There is also the case where the answer came back no and you think the hour is the reason. Before you re-book it into a better one, work out which no you got.
Only one of the two moves with the clock. A no that came out of disagreement is a question about how to reopen an argument, and pushing back on your manager is where that goes. A no that came out of the hour is the one a better slot can fix.
- No to the decision. Your manager thinks the other option is better. A fresh 9am slot changes nothing, because the hour was never what you were arguing about. Ask what evidence would change their mind, go get it, or drop it.
- No to the work. They agree with you and cannot absorb what a yes creates this month: the conversation with finance, the replanning, the two people who will ask why they weren’t consulted. Come back with the version that costs them less. “I’ll write the doc and the comms, you approve them” is a different ask from “we should do this.”
Telling them apart takes one question, asked in the room rather than a week later: “Is this a no to the idea, or a no to doing it now?”
The time management discipline that protects your own week applies to how you spend other people’s attention too, and it is the part of the job most individual contributors get no training in.
None of this has been measured on a manager’s calendar and it may never be. What you can check is your own record: the last three things you took to someone at 4pm, and what came back.
Merlin can nudge you before the ask, which is roughly when the timing decision is still yours to make.