Decision Fatigue: The Hidden Tax of AI Output
The phrase going around this year is “AI brain fry.” A 2026 study from Boston Consulting Group, published in the Harvard Business Review, surveyed nearly 1,500 full-time workers and found that the single most mentally taxing thing they did all day was not writing, not analysis, and not meetings. It was watching over AI. People who had to closely oversee AI agents reported 14% more mental effort, 12% more mental fatigue, and 19% more information overload than people whose AI use was lighter. Same tools that were sold to save their brains were quietly frying them.
That is the headline. Here is the durable thing underneath it, the part that will still be true when the phrase has changed.
Everyone assumed AI would give you back time. Buy the tool, kill the busywork, get your hours back. For the making part, that is exactly what happened. Drafting, coding, researching, designing, all of it got faster and cheaper, close to free. But making was never the whole job. The other half of the job is deciding, and AI did not shrink that half. It blew it up. Every draft it writes is now a choice about which draft to keep. Every ten options it hands you is a choice about which one to run. The tool that removed the friction from doing poured all of that friction straight onto deciding, and deciding is the one thing that gets worse the more of it you do.
I have felt this in my own week, and I have watched it flatten founders far sharper than me. It does not feel like fatigue at first. It feels like productivity. That is what makes it dangerous.
Table of contents
- The bottleneck just moved
- The Oversight Tax: one task, ten forks
- Deciding is a depleting resource
- The approval treadmill
- The 3 p.m. founder
- The compounding cost of one tired decision
- Why more AI can’t fix it
- The Delegation Map: where each decision goes
- The founders who win decide less
- The contrarian take: the high-value-decisions myth
- What to do Monday morning
- FAQ
The bottleneck just moved
Every system has one binding constraint, the single spot that caps the whole thing. Speed up anything that is not the constraint and you get nothing, because the constraint still holds the line. This is old factory wisdom, and it maps cleanly onto a person’s day.
For most of the history of building things, the binding constraint on a founder was execution. You could think of ten things worth doing and get to two of them, because doing was slow. So the whole culture of productivity grew up around that constraint. Work faster. Ship more. Reduce the time between idea and thing. Every tool, every method, every book aimed at the same wall: make execution cheaper.
AI knocked that wall down. Execution is no longer the thing you run out of. And the moment a constraint stops binding, it stops mattering, and the bottleneck jumps somewhere else. It jumped onto the step that AI cannot do for you, which is choosing. Which draft. Which direction. Which of the fifteen plausible-looking things to actually commit to. I wrote a whole piece on one version of this, the gap between how much you can now build and how big you dare to aim. Decision fatigue is the sibling problem. Ambition is about the size of what you attempt. Decision fatigue is about the wear on the muscle that picks.
The proof that the bottleneck moved is sitting in the productivity numbers, and they are strange. By 2026, 91% of companies were using AI in at least one function, yet 89% of managers reported no change in productivity, and only about 39% saw any profit impact. MIT’s NANDA research found that roughly 95% of enterprise generative-AI pilots delivered no measurable return. Read that next to the “AI makes me 10x faster” claims and it looks like a contradiction. It is not. Individuals got faster at making. Organizations did not get better at deciding, sorting, and committing, and that is where the value actually sits. The making got cheap. The bottleneck did not disappear. It relocated, and most people are still optimizing the wall that already fell.
The Oversight Tax: one task, ten forks
Here is the core mechanism, and it is almost arithmetic. Call it the Oversight Tax.
Before AI, one task was roughly one decision. You decided to write the email, then you wrote it. The deciding and the doing were fused, and the doing was the expensive part. AI splits them apart and inverts the cost. Now you ask for the email and get three versions. That is not one decision anymore. It is a choice between three, plus a judgment on whether any of them is good, plus a call on what to fix, plus a decision on whether to ask for a fourth. One task fanned out into a small cloud of forks, and every fork has to be resolved by the one part of the process AI did not replace. You.
Multiply that across a day. The AI writes your outreach, so now you are picking messages instead of writing them. It drafts your spec, so now you are judging a spec instead of thinking one through. It generates ten designs, ten names, ten subject lines, and each batch of ten is a new pile of small verdicts dropped on your desk. The volume of things produced went up ten times. The volume of things you must decide went up with it. That is the tax: cheap execution does not come free, it comes with a surcharge paid in decisions, and the bill lands entirely on the person at the top of the funnel.
It helps to see the two halves of the job as two separate ledgers. AI is generous on one and quietly brutal on the other. Here is what it takes off your plate, and what it drops back on.
| The task | Execution friction AI removes | Decision friction AI adds |
|---|---|---|
| Writing | Drafts the email in seconds | Now you pick between three versions and judge the tone of each |
| Research | Summarizes twenty sources instantly | Now you decide what to trust and what it quietly missed |
| Building | Ships the feature in an afternoon | Now you decide whether to ship it, and you decide it faster than you can think |
| Ideas | Generates ten directions on demand | Now you kill nine good-enough options to keep one, ten times a day |
| Hiring, ops, comms | Drafts the JD, the policy, the reply | Now every one of them waits for your yes, and there are more of them |
The left column is the reason you bought the tool. The right column is the bill, and almost nobody reads it before signing.
Deciding is a depleting resource
If deciding were free, none of this would matter. You would just make more decisions, faster, forever. But deciding is not free, and it is not renewable within a day. It runs on a budget, and the budget drains.
This is one of the older findings in behavioral science. The psychologist Roy Baumeister spent years documenting what he called ego depletion: making decisions and exercising self-control draw on the same limited pool, and each act of choosing leaves less in the tank for the next one. Later work has argued about the size and the mechanism of the effect, and it is fair to say the strongest early claims were oversold. But the plain, load-bearing version has held up in ordinary experience for anyone who has ever tried to make a hard call at the end of a brutal day. The tenth decision is worse than the first. Not because you got dumber. Because the resource that quality decisions run on is lower than it was at nine.
The most cited illustration is uncomfortable. In a 2011 study, Shai Danziger and colleagues looked at 1,112 parole decisions made by eight experienced judges in Israel over ten months. At the start of a session, the judges granted parole about 65% of the time. Right before a food break, that rate fell close to zero. After the break, it snapped back to 65%. Same judges, same law, same kinds of cases. The variable that moved the outcome was how many decisions the judge had already made since last resting. The magnitude of that specific effect has been debated since, and case ordering probably explains part of it, so I would not stake a life on the exact numbers. I would stake real money on the direction. A tired decider is a worse decider, and the tiredness comes from the deciding itself.
Powerful people have quietly organized their lives around this for years. Barack Obama told Vanity Fair, “You’ll see I wear only gray or blue suits. I’m trying to pare down decisions. I don’t want to make decisions about what I’m eating or wearing. Because I have too many other decisions to make.” That is not a fashion quirk. It is a president treating his decision budget as the scarce input it is, and refusing to spend a single unit of it on a necklace of trivial choices. He understood the thing most founders have not yet internalized: your capacity to decide well is finite, it depletes through the day, and where you spend it is itself the most important decision you make.
Put the Oversight Tax and the decision budget together and you have the whole trap in one sentence. AI multiplies the number of decisions you face. Your capacity to make them well is fixed and falling. So the more AI output you pull toward yourself, the sooner you hit the cliff, and the worse every decision after it becomes, including the ones that actually matter.
The approval treadmill
The BCG study I opened with points at the exact spot where this hurts most. It was not using AI that fried people. It was overseeing it. Watching, checking, approving. And there is a clean reason overseeing is heavier than doing: when you do a thing, you are inside it, you have context, the next move is obvious. When you oversee a thing someone else produced, you have to reconstruct the context from the outside, judge whether the output is right, and decide what to do about it, cold, over and over. That is pure decision load with none of the flow that doing gives you.
I call the state this creates the approval treadmill. Your AI agents produce, and you approve. They produce again, and you approve again. The work feels like it is moving, and technically it is, but you have quietly turned yourself into a full-time reviewer of machine output. Your day is now a queue of yes-or-no calls on things you did not make, with no end and no flow, and every one of them costs a unit of the budget. The BCG researchers found something that should stop every founder mid-stride: productivity rose going from one AI tool to two, still rose with a third but slower, and then declined past three tools. The oversight load crossed over and started eating the gains. More tools did not mean more output. Past a point it meant less, because a human can only stand at the end of so many conveyor belts.
This is a cousin of a risk I have written about directly, automation bias, where a team quietly stops checking AI output because it has been right for so long. The two failures are twins with opposite symptoms. Automation bias is what happens when you stop paying the oversight cost and let things through unchecked. The approval treadmill is what happens when you pay it in full and it grinds you down. Both come from the same root: a flood of machine output arriving at a single human who has to make a real decision about each piece of it. You either drown checking it or you stop checking it. Neither is a good place to run a company from.
The 3 p.m. founder
Here is what the trap looks like from the inside, because it does not look like fatigue. It looks like a normal, busy, productive day, right up until the decisions start going bad.
You start the morning sharp. The first few calls are good, considered, the kind of judgment you are proud of. Then the AI-assisted day does its thing. A stream of drafts to approve, options to sort, outputs to sanity-check, each one small, none of them the big strategic call you imagined AI would free you up for. By early afternoon the budget is low and you cannot feel it draining, because that is the cruel part, depletion does not announce itself. It shows up as a subtle shift in how you decide. You start reaching for the fastest acceptable answer instead of the best one. You approve things to make the queue shorter rather than because they are right. You defer the hard call to tomorrow, again. This is the 3 p.m. founder, making the day’s most consequential decision with the day’s most depleted mind, because the trivial ones ate the morning.
The attention data makes it worse. ActivTrak’s 2026 workplace report found the average focused work session had fallen to 13 minutes and 7 seconds, down 9% in three years. A Fortune report the same year found deep-focus sessions dropping while time spent on email roughly doubled. So the founder is not just depleted, they are fragmented, ricocheting between an inbox, three AI tools, and a dozen open decisions, never in one long enough to think a hard problem all the way through. Decision quality does not just need budget. It needs uninterrupted runway to actually engage, and the AI-assisted day shredded the runway into 13-minute strips.
None of this shows on a dashboard. Your output metrics look great. More shipped, more sent, more done. The cost is invisible because it lives in the counterfactual, in the sharper decision you would have made with a full tank and a clear hour, the strategic bet you rushed, the wrong hire you rubber-stamped at 4 p.m. This is the same illusion I unpacked in the velocity illusion, where the feeling of shipping fast hides whether you are shipping the right things. Decision fatigue is the engine underneath it. You feel fast because the making is fast. You cannot feel that the deciding got worse.
The compounding cost of one tired decision
It would be easier to shrug all this off if a fatigued decision were just a slightly weaker version of a good one, a rounding error you make up tomorrow. It is not, and the reason is that founder decisions do not sit still after you make them. They compound. A single depleted call at the top of a company propagates outward and forward, and it keeps costing long after the afternoon it was made.
Think about what actually lands in the Protect quadrant. The hire you rubber-stamp at 4 p.m. because the interview was your ninth decision of the hour is not a one-day mistake, it is six to twelve months of a wrong person shaping a team, and the cost of unwinding it dwarfs the minutes you saved by deciding fast. The pricing number you set while depleted becomes the ceiling every future quarter lives under. The strategic bet you approved because the queue was long, not because you had thought it through, becomes the thing an entire team spends the next two quarters executing against, at full speed, in possibly the wrong direction. Small daily depletion, enormous downstream bill, because the decisions that deplete you least in the moment are often the ones that compound the most over time.
AI sharpens this asymmetry in a way worth naming. It made execution mistakes cheap to reverse. Ship the wrong feature and a model can help you rebuild it by Friday. But it did nothing to make decision mistakes cheap to reverse, and the highest-stakes decisions, the ones about direction and people and what not to do, were never reversible on a weekend. So the cost of a bad call did not fall with the cost of building. If anything it rose in relative terms, because now the thing that separates companies is not who can execute, almost everyone can, but who pointed their cheap execution at the right target. That pointing is a decision, made by a depletable human, on a budget that the AI-flooded day drains faster than ever.
This is why protecting the decision budget is not a wellness nicety or a productivity hack. It is the highest-return resource-allocation choice a founder makes, and it hides in plain sight because the cost of getting it wrong never shows up as a line item. It shows up as a hire that did not work, a strategy that drifted, a year spent climbing the wrong hill efficiently. The founders who compound in the right direction are not the ones who made the most decisions. They are the ones who arrived at the few that mattered with something left in the tank.
Why more AI can’t fix it
The natural founder reflex, once you see the pile of decisions, is to reach for the tool that got you here. Buy the AI that filters your AI. Automate the sorting. Let a model triage the outputs so you only see the ones that matter. Sometimes this helps at the edges. As the main strategy it fails, for a reason worth being precise about.
Adding AI to manage your AI does not remove decisions. It adds a layer, and the layer has its own decisions. Now you have to decide whether to trust the triage, tune the filter, catch what it wrongly killed, and judge whether its sorting matches your actual priorities, which change weekly. You did not escape the deciding. You moved it up a level and made it more abstract, which is often harder, not easier. There is a documented pattern behind this that researchers in 2026 started calling the delegation feedback loop: the better AI gets at handling tasks, the lower the bar drops for handing the next task over, so you keep delegating more, which keeps generating more output, which keeps generating more decisions. The tool that causes the flood cannot be trusted to drain it, because draining it is itself a judgment call, and judgment is exactly the scarce thing under attack.
The deeper issue is a mismatch of currencies. AI scales like compute. You add more, you get more, cheaply, on demand. Decision quality does not scale like compute. It scales like sleep, or attention, or trust: it is finite, it depletes, and you cannot buy a bigger daily supply by spending money. This is why the founders who are struggling most are often the ones using the most tools. They keep applying a compute solution to a human-capacity problem, and the two do not trade. The answer to too many decisions was never going to be a machine that generates more of everything. I made a related argument in the AI efficiency trap: optimizing the cheap thing harder does not help when the cheap thing was never your constraint.
The Delegation Map: where each decision goes
The way out is not to use less AI. It is to stop treating every decision as if it deserves the same slice of you. Most of them do not. The skill that matters now is sorting: deciding, in advance, which decisions get your depleting budget and which ones never should have reached you at all.
Two questions sort almost everything. How often does this kind of decision come up, its volume? And how much does getting it wrong cost, its stakes? Cross those two and you get four boxes, and each box has a different correct answer. This is the Delegation Map, and its whole job is to keep your fresh-mind hours away from the low-stakes flood.
Each quadrant is a different instruction, and getting the instruction right is most of the battle.
| Quadrant | Example | What you do | What it protects |
|---|---|---|---|
| Automate | Categorizing expenses, formatting notes, first-pass replies | Hand it fully to AI. Do not open the output at all. | Your attention, from being nibbled to death |
| Batch | Approving routine copy, choosing between AI drafts, small spend | Write one rule, or review a batch weekly, never one at a time | Your budget, from the death by a thousand yeses |
| Protect | Strategy, key hires, pricing, what not to build | Reserve a fresh, uninterrupted block. Decide these first. | The decisions that actually move the company |
| Fatigue Trap | A flood of consequential, one-off calls with no system | Attack the volume: systematize the repeatable parts out of it | You, from being the single point of failure |
The map is not about doing less. It is about noticing that a huge share of what now lands on you is low-stakes volume masquerading as work that needs your judgment. It does not. Choosing between three AI-written subject lines is an Automate or Batch decision that got dressed up as a Protect decision because it arrived in your inbox with your name on it. Send it back down where it belongs and the flood drops by most of its size.
The founders who win decide less
The counterintuitive move, once you accept a fixed budget, is to aim for fewer decisions, not faster ones. Speed is a trap here, because a faster bad decision is still a bad decision, made sooner. Fewer is the real goal, and there are three durable ways to get there.
The first is pre-commitment: decide the rule once, in a calm moment, so the individual cases stop being decisions at all. Obama’s suits are pre-commitment. So is “we do not take meetings before noon,” or “any refund under fifty dollars is auto-approved,” or “we only build features that ten paying customers have asked for.” Each rule you set converts a recurring drip of small choices into a single decision you already made when your budget was full. This is the entire logic of the Batch quadrant, and it is the highest-return habit a busy founder can build.
The second is real delegation, including to AI, without the treadmill. The trap is not delegating the doing. It is keeping the deciding by insisting on approving everything. If a decision genuinely belongs in Automate, delegating it means you do not look at the output, full stop. A yes on every single item is not oversight, it is anxiety with extra steps, and it costs the same budget as a real decision for none of the value. Pick the handful of workflows where checking actually catches costly errors, check those hard, and let the rest run. This connects to something I have argued about the skill of judgment itself, in the reps problem: you keep sharp judgment by spending it on decisions that deserve it, not by spreading it thin across everything.
The third is taste, which is the ultimate decision-saving device. A person with strong taste in a domain decides faster and better with less depletion, because much of the deciding has been pre-compiled into instinct. They do not deliberate over the three AI drafts. They glance and know. Building that instinct is slow and it is exactly the thing AI tempts you to skip, which is why I wrote a full guide to building taste on purpose, and why the fluency trap, where AI lets you produce good work in a field you never actually learned, is so dangerous. Taste is a decision budget you deposited years ago and get to spend cheaply now. AI cannot lend it to you.
The contrarian take: the high-value-decisions myth
The standard advice about all this is comforting and wrong. It goes: let AI handle the busywork so you are freed up for the big, high-value decisions. Every productivity influencer says a version of it. The picture is a founder relieved of drudgery, sitting back to make three important calls a day with a clear head.
That is not what happens, and the reason is structural. AI does not just remove the busywork, it generates a flood of new, plausible, medium-stakes options, and that flood buries the high-value decisions rather than surfacing them. When AI hands you ten strategies instead of forcing you to think one through, you have not been freed to decide well, you have been handed a sorting job on top of the deciding job. The big decision is now sitting somewhere in a pile of ten good-looking ones, and you have to spend budget just to find it before you can spend budget to make it. More output does not clarify the important choice. It camouflages it. The founders who actually get the freedom the influencers promise are the ones who ruthlessly refuse most of the output, so the few decisions that matter can stand alone with room around them.
Let me argue the honest other side, because there is one. Some decisions genuinely should be handed to AI completely, and founders who reflexively distrust every model output create their own fatigue by re-deciding things a machine could own outright. The goal is not blanket resistance to delegation, and it is not heroic manual control of everything. It is discernment about which quadrant a decision lives in. A founder who automates the low-stakes flood and a founder who insists on personally approving every AI output are both wrong, in opposite directions. The skill is the sorting, and the sorting is itself the highest-value decision you make all day, which is a small irony worth sitting with. The most important use of your judgment is deciding where not to use it.
What to do Monday morning
This is fixable, and it does not require a new tool. It requires reclaiming the deciding as a thing you manage on purpose. Here is the concrete version.
Run a decision audit for one day. Keep a simple tally. Every time you make a call, mark it, and note whether it was triggered by AI output, an approval, a draft, an option to pick. Most founders are shocked to find that more than half their decisions are low-stakes reactions to machine output, not the strategic work they think fills their day. You cannot fix a load you have never measured.
Map last week’s decisions onto the four quadrants. Take the recurring ones and sort them: Automate, Batch, Protect, or Fatigue Trap. Anything sitting in Automate or Batch that you are still touching one at a time is a leak. That is where the budget is bleeding, and it is the cheapest thing to fix.
Write three pre-commitment rules this week. Pick your three most repetitive low-stakes decisions and make a standing rule for each, so they stop being decisions. Auto-approve under a threshold. A default answer for a common request. A hard filter on what you will even consider. You are converting a daily drip into one decision you already made.
Protect one fresh block for the decisions that matter. Put a real, uninterrupted hour early in your day, before the flood, and spend it only on Protect-quadrant calls. Make the consequential decision with a full tank, not with what is left at 3 p.m. If the strategic call is the most important thing you do, it should get your best budget, not your last scraps.
Kill one treadmill. Find one workflow where you are approving AI output item by item out of habit, not because your checking catches real errors. Either automate it fully or move it to a weekly batch review. Measure whether anything breaks. Almost always, nothing does, and you get a piece of your mind back. For the larger operating system this fits into, start with the founder operating system and its AI-age update, and if you are still deciding which skills are even worth keeping sharp in this era, this is where I would start.
FAQ
What is decision fatigue?
Decision fatigue is the well-documented decline in the quality of your decisions as you make more of them. Choosing draws on a finite mental resource that depletes through the day, so late decisions tend to be worse than early ones, not because you got less capable but because the budget that good decisions run on is lower. It was studied under the banner of ego depletion by Roy Baumeister and illustrated by the 2011 parole-judge research, and while the exact size of the effect is debated, the direction is a common, lived experience.
How does AI cause decision fatigue if it is supposed to save time?
AI saves time on execution, the making, but the job also has a deciding half, and AI multiplies that half. Every draft it writes becomes a choice about which draft to keep, every batch of ten options becomes nine rejections, and every output becomes an approve-or-edit call. The tool removes the friction from doing and pours it onto deciding. Since deciding is the part that depletes and does not scale, you can feel busier and more tired even as your raw output goes up.
Is decision fatigue the same thing as AI brain fry or burnout?
They overlap but they are not identical. AI brain fry, the term from the 2026 BCG study, is the broad mental exhaustion from overseeing AI, and decision fatigue is the specific mechanism inside it: the depletion of your capacity to choose well. Burnout is a longer-term, more total collapse. Decision fatigue is a daily, recoverable drain, but if you never manage it, the chronic version of it is one of the roads that leads toward burnout.
Why can’t I just use more AI tools to handle the extra decisions?
Because a tool that filters your AI output adds its own decisions: whether to trust it, how to tune it, what it wrongly killed, whether its priorities match yours this week. You move the deciding up a level and often make it harder. There is a pattern researchers in 2026 named the delegation feedback loop, where better AI keeps lowering the bar for handing over the next task, generating more output and more decisions. Decision quality scales like sleep, not like compute, so you cannot buy more of it by adding tools.
How do I know if decision fatigue is hurting my decisions?
Watch the shape of your day. If your morning calls are sharp and your late-afternoon calls are rushed, deferred, or rubber-stamped, that is the budget draining. Other signs: you approve things to shorten the queue rather than because they are right, you keep pushing the hard decision to tomorrow, and your important choices are happening late in the day rather than first. A one-day decision audit, tallying every call and its trigger, makes it visible fast.
What is the fastest way to reduce decision fatigue as a founder?
Write pre-commitment rules for your most repetitive low-stakes decisions so they stop being decisions at all. Auto-approve under a threshold, set a default answer for a common request, define a hard filter on what you will even consider. Each rule converts a daily drip of small choices into one decision you already made when your budget was full. Pair that with protecting one fresh, uninterrupted block early in the day for the decisions that actually matter.
Should I let AI make important decisions for me to save my budget?
Only the low-stakes, reversible ones, and for those you should not even review the output. High-stakes decisions are exactly the ones to protect, not delegate, because that is where your judgment earns its keep. The skill is sorting decisions by volume and stakes and sending each to the right place: automate the cheap flood, batch the routine, and reserve your fresh-mind hours for the few calls that shape the company. Delegating the wrong tier of decision to AI is how automation bias and expensive mistakes get in.
Does decision fatigue really change decision quality, or is that a myth?
The strongest early lab claims about ego depletion were oversold, and the precise magnitude of the parole-judge effect has been challenged, so healthy skepticism about specific numbers is warranted. What has held up is the everyday, load-bearing version: making many decisions is tiring, tired people make worse and lazier decisions, and reducing trivial choices frees capacity for important ones. You do not need the strong lab claim to be true to benefit from treating your daily decision capacity as finite, because in practice it behaves that way.
The one line to remember
AI did not remove your decisions. It multiplied them. Execution got cheap and deciding got expensive, and deciding is the one thing that gets worse the more of it you do. The founders who win the next few years will not be the ones who use AI to make more decisions faster. They will be the ones who use it to make fewer decisions, on purpose, so the handful that matter get a full tank and a clear hour instead of the scraps left at 3 p.m. Guard the budget like the scarce input it is, because it is the one thing you cannot buy more of, and it is the thing everything else runs on.