Von Neumann and Morgenstern (1944) formalised rational choice under uncertainty: maximise expected utility, not expected value. The formula is Σ (probability × utility). A 50% chance of $100 and 50% chance of $0 has expected value $50. But the rational decision depends on utility — how much satisfaction each outcome delivers. $100 means more to someone with $10K than to someone with $10M. Diminishing marginal utility: each additional dollar adds less satisfaction than the one before. Expected utility theory (EUT) says the rational actor multiplies each outcome by its probability, applies a utility function to each outcome, and chooses the option with the highest sum.
The distinction from expected value matters. A gamble with positive expected value can have negative expected utility if the downside is catastrophic relative to your wealth. A gamble with negative expected value can have positive expected utility if the upside delivers outsized satisfaction. EUT is normative — it describes how a rational agent should decide. Kahneman and Tversky's Prospect Theory (1979) showed we systematically violate it: we're loss-averse, we overweight small probabilities, we evaluate gains and losses asymmetrically. EUT remains the benchmark.
The strategic use: when outcomes are probabilistic, model decisions as expected utility. Estimate probabilities, assign utilities, multiply, sum. The discipline forces explicit treatment of risk and preference. Applied to investing, the Kelly Criterion is expected utility with logarithmic utility — bet size proportional to edge over odds. Applied to business, Amazon's "two-way door" is expected utility for reversible decisions: low downside, high upside, so expected utility favours action over delay. The model does not tell you what to do. It gives you a structure for deciding. The probabilities and utilities are yours to supply. The value is in making them explicit.
Diminishing marginal utility is central. $100 to someone with $10K delivers more utility than $100 to someone with $10M. The same dollar, different satisfaction. This is why expected value fails: it treats all dollars equally. Expected utility corrects for the fact that the first million changes your life and the tenth changes almost nothing. The utility function captures this. Without it, you're optimising the wrong thing.
The model is prescriptive, not descriptive. It tells you what to do, not what people do. Kahneman's Prospect Theory overturned EUT as a description of human behaviour — we violate the axioms routinely. But EUT remains the right standard for rational choice. When you catch yourself deviating from it, ask why. Sometimes the deviation is a bug (sunk cost fallacy, loss aversion). Sometimes it's a feature (you value something the model doesn't capture). The discipline is to know the difference.
Section 2
How to See It
Expected utility appears wherever decisions involve uncertain outcomes and the decision-maker must weigh probability against consequence. The diagnostic: would the "rational" choice differ from the choice people actually make? If yes, either people are miscalculating expected utility or they're applying a different calculus (Prospect Theory, mental accounting, etc.). Train yourself to spot the structure: multiple possible outcomes, each with a probability and a consequence. The decision is a function of both. Most people fixate on one or the other — the upside or the downside, the probability or the payoff — and ignore the product. The four domains below — investing, business, personal life, strategy — are where the structure appears most often. The same logic applies wherever uncertainty meets consequence.
Investing
You're seeing Expected Utility when an investor evaluates a startup: 10% chance of 100x return, 90% chance of total loss. Expected value is positive. Expected utility depends on the investor's wealth, risk tolerance, and the shape of their utility function. A pension fund may reject it; a billionaire may accept it. Same expected value, different expected utility. The Kelly Criterion formalises this: optimal bet size depends on edge, odds, and the investor's utility function. Most investors never write down the calculation. The best ones do — or have internalised it through experience.
Business
You're seeing Expected Utility when a founder decides whether to launch a new product line. The outcome is probabilistic — market response, competitive reaction, execution risk. The rational framework: assign probabilities to each scenario, estimate utility (or disutility) of each, multiply, sum. Most founders skip the formalisation and rely on intuition. The ones who don't often make better calls. The discipline of writing down scenarios forces clarity on what "success" and "failure" mean — and whether the probabilities are realistic or wishful.
Personal life
You're seeing Expected Utility when someone chooses between a stable job and a risky startup. The stable job has high probability of moderate outcome. The startup has low probability of extreme upside, high probability of moderate downside. Expected value may favour the startup. Expected utility depends entirely on the person's utility curve — how they weight security vs upside, how loss-averse they are. Two people with identical information can make opposite choices. Both can be rational. The utility function is the differentiator.
Strategy
You're seeing Expected Utility when a company evaluates a "two-way door" decision — reversible, low downside, high upside. Amazon's framework: these decisions should be made quickly because the expected utility calculation favours action. The downside is bounded; the upside is unbounded. Delay has negative expected utility. The signal: when a team spends weeks debating a decision that could be reversed in days, they're misapplying expected utility. The cost of delay exceeds the cost of being wrong.
Section 3
How to Use It
When facing probabilistic outcomes, make the expected utility calculation explicit. Assign probabilities. Assign utilities (or use a proxy: money, time, satisfaction). Multiply. Sum. Compare. The discipline is simple. The execution is hard — we resist assigning numbers to uncertain outcomes, and we confuse expected value with expected utility. The payoff: decisions that are consistent with your actual preferences rather than your intuitions in the moment.
Decision filter
"When outcomes are uncertain, ask: what is the expected utility of each option? Probability × utility for each outcome, summed. If you cannot assign utilities, use a proxy — but be explicit about what you're optimising for. The act of formalising often reveals that you're optimising the wrong thing."
As a founder
Model major decisions as expected utility. A pivot: what's the probability it works? What's the utility if it does? What's the utility if it doesn't? A hire: probability they succeed, utility of success, utility of failure. The two-way door principle — reversible decisions with bounded downside — is expected utility in disguise. If reversal costs little and upside is large, expected utility favours action. Bezos built Amazon's decision culture around this: type 2 decisions get made fast because delay has negative expected utility. The discipline: classify decisions by reversibility and downside. Type 2 gets speed. Type 1 gets deliberation. Most decisions are type 2; most organisations treat them as type 1. A/B tests are expected utility in practice: you're explicitly estimating the probability that variant B beats A, and the utility of the improvement. Run the test. The expected utility of learning exceeds the expected utility of guessing.
As an investor
Portfolio construction is expected utility applied. Kelly Criterion — bet size proportional to edge over odds — maximises long-run growth of wealth under specific assumptions. It's expected utility with a logarithmic utility function. The key: your utility function shapes optimal bet size. Risk-neutral investors bet more. Risk-averse investors bet less. Most people are risk-averse in gains, risk-seeking in losses — Prospect Theory — which produces suboptimal expected utility. The discipline: know your utility function, then size positions accordingly. When evaluating a startup, the expected value may be positive. The expected utility depends on your wealth, your other positions, and how you weight the tail outcomes. Same deal, different utility functions, different optimal decisions.
As a decision-maker
Before any decision with uncertain outcomes, write down the scenarios, probabilities, and utilities. The exercise forces clarity. Often you discover that the "obvious" choice has lower expected utility than the alternative — or that the probabilities you're implicitly using don't match the ones you'd endorse if made explicit. Expected utility is a forcing function for honest probability assessment. The second benefit: it surfaces when you're optimising the wrong thing. Revenue? Profit? Market share? Survival? The utility function defines the objective. Get it wrong and the whole calculation is wrong.
Common misapplication: Confusing expected value with expected utility. A lottery ticket has negative expected value — the house edge. It can still have positive expected utility for someone who values the thrill or the dream. The reverse: a positive-EV bet can have negative expected utility if the downside is ruinous. Expected value is one number. Expected utility requires a utility function. Use the right one.
Second misapplication: Treating EUT as a prediction of behaviour rather than a normative standard. People don't maximise expected utility. They approximate it, violate it, ignore it. The value of EUT is not in predicting what people will do. It's in clarifying what the rational choice would be — and then asking whether your actual choice differs, and why. The gap between EUT and behaviour is where Prospect Theory, mental accounting, and other models earn their keep.
Third misapplication: Using expected value when the stakes are asymmetric. A bet with positive expected value can be ruinous if the downside is total loss. A bet with negative expected value can be rational if the upside delivers outsized utility (e.g., a lottery ticket for someone who values the dream). The expected value calculation ignores this. Expected utility doesn't. When the outcomes are asymmetric — small chance of huge gain, or small chance of catastrophic loss — expected value is the wrong metric. Use expected utility.
Section 4
The Mechanism
Section 5
Founders & Leaders in Action
The founders below applied expected utility — explicitly or implicitly — to capital allocation, organisational design, and strategic bets. The common thread: they formalised the probability-utility calculus rather than relying on intuition alone. Bezos operationalised it for decision speed. Buffett applied it to portfolio construction. Musk and Thiel applied it to contrarian bets where the market's probability estimates were wrong. Each understood that the rational choice under uncertainty depends on both the probabilities and the shape of the utility function. None of them would describe their process in these terms. The framework is a reconstruction of what they did — and a guide for what you might do.
Bezos's "two-way door" framework is expected utility applied to organisational speed. Type 2 decisions are reversible, low-cost to undo, high upside if right. Expected utility favours action: the downside is bounded, the upside is not. Bezos pushed these decisions down and demanded speed. Type 1 decisions — irreversible, consequential — get more deliberation. The distinction is expected utility: when the cost of being wrong is low and the cost of delay is high, expected utility says act. Amazon's willingness to experiment at scale — Prime, AWS, Kindle — reflects this. Each bet had bounded downside and unbounded upside. Expected utility said yes. The "disagree and commit" principle extends it: once a decision is made, further debate has negative expected utility. The cost of delay and second-guessing exceeds the benefit of marginal improvement.
Buffett's capital allocation is expected utility with a conservative utility function. He avoids permanent loss of capital — the downside has near-infinite disutility for him. He concentrates in high-conviction positions — the upside, when he's right, compounds for decades. His utility function is effectively logarithmic: he cares about percentage returns, not absolute dollars. The Kelly Criterion — which Buffett has cited — is expected utility maximisation with log utility. Buffett's "circle of competence" is a constraint that improves his probability estimates. Better probabilities, better expected utility. His refusal to invest in tech for decades was not ignorance; it was expected utility. He couldn't estimate the probabilities. Without reliable probabilities, expected utility is undefined. He stayed in his circle where probabilities were estimable.
Musk's bets on SpaceX and Tesla were expected utility calculations with extreme outcomes. Low probability of success, but if successful, civilisation-scale upside. His utility function appears to weight existential outcomes heavily — making humanity multi-planetary has utility that dwarfs conventional financial returns. The same logic: he accepts high variance because the upside, in his utility calculus, justifies it. The "first principles" framing often masks an expected utility calculation — he's explicit about probabilities and utilities that others don't formalise. SpaceX's bet that reusable rockets were possible when the industry said they weren't: he assigned higher probability to success than the consensus. Different probability estimates, different expected utility, different decision.
Peter ThielCo-founder, PayPal & Palantir; Partner, Founders Fund
Thiel's "zero to one" thesis is expected utility for contrarian bets. When the consensus assigns very low probability to an outcome, the expected value can be positive even if the outcome is unlikely — because the payoff is enormous. Thiel looks for opportunities where the market's probability estimate is wrong. His venture strategy: asymmetric upside, bounded downside. The expected utility of a portfolio of such bets — when you have conviction the probabilities are mispriced — is positive. The key is having better probability estimates than the market. "The best venture investments seem to involve companies that are good at hiding their secrets" — the market underprices the probability of success. Thiel's edge is probability estimation, not utility; he shares the conventional utility for financial returns.
Section 6
Visual Explanation
Section 7
Connected Models
Expected utility gains explanatory power when connected to adjacent frameworks. Prospect Theory describes how we deviate from it. Kelly Criterion operationalises it for repeated bets. Risk tolerance defines the utility function. The connections below show how EUT fits into the broader lattice of decision models — and how to use it in combination with models that describe actual (rather than ideal) behaviour.
Reinforces
Expected Value
Expected value is the special case of expected utility when utility equals money. When outcomes are monetary and the decision-maker is risk-neutral, EUT reduces to maximising expected value. The connection: EUT generalises expected value to account for risk preference and diminishing marginal utility. In practice, expected value is a useful first pass — but it fails whenever the decision-maker is not risk-neutral or whenever outcomes have non-monetary dimensions. Expected value answers "what's the average outcome?" Expected utility answers "what should I do?"
Tension
Prospect Theory
Prospect Theory describes how people actually decide under risk — and it systematically violates EUT. We're loss-averse. We overweight small probabilities. We evaluate gains and losses relative to a reference point. EUT is normative; Prospect Theory is descriptive. The tension: use EUT as the benchmark; expect Prospect Theory in practice. When designing choices for others — pricing, incentives, nudges — Prospect Theory tells you how they'll actually respond. When making your own decisions, EUT tells you what the rational choice would be. The gap between the two is where most decision errors live.
Reinforces
Kelly Criterion
Kelly is expected utility maximisation with logarithmic utility. Bet size = edge / odds. It maximises long-run growth of wealth. The connection: Kelly is the optimal strategy when your utility function is log and you have repeated opportunities. EUT provides the framework; Kelly provides the formula. The Kelly fraction tells you what proportion of capital to risk on each bet. Full Kelly is aggressive; half-Kelly is more conservative. Both are expected utility maximising under different risk tolerances. The criterion is widely cited by investors; few follow it precisely. The discipline of sizing bets by edge and odds is the main takeaway.
Section 8
One Key Quote
Von Neumann and Morgenstern's formulation is the cleanest statement of the core principle. The rational agent does not maximise expected value. They maximise expected utility — and the distinction is everything.
"The individual maximizes expected utility — the sum of probabilities times utilities for each outcome."
— John von Neumann & Oskar Morgenstern, Theory of Games and Economic Behavior (1944)
The formula is simple. The implications are not. The quote captures the entire framework in one sentence. Every extension — Kelly, two-way doors, risk tolerance — follows from it. The rational agent does not sum probabilities times dollars. They sum probabilities times utilities. Get that right and the rest follows. Get it wrong and you're optimising the wrong thing. Once you accept that rational choice maximises expected utility rather than expected value, every decision under uncertainty changes. Risk preference matters. Wealth level matters. The shape of your utility function matters. Von Neumann gave us the framework. Applying it is the work.
Section 9
Analyst's Take
Faster Than Normal — Editorial View
The most underused discipline in decision-making is making the expected utility calculation explicit. Most decisions with uncertain outcomes are made intuitively. The intuition may be good. It may be terrible. You don't know until you write down the scenarios, probabilities, and utilities. The act of formalising forces clarity. Often you discover that the probabilities you're implicitly using are wrong — or that you're optimising for the wrong thing entirely.
Expected value is a trap. Founders and investors default to expected value when they should use expected utility. A startup with 10% chance of 100x and 90% chance of zero has high expected value. Whether you should invest depends on your utility function. A pension fund correctly rejects it. A billionaire correctly accepts it. The expected value is the same. The expected utility is not. The same logic applies to career decisions, product bets, and strategic pivots. Know what you're optimising for.
The two-way door principle is EUT in disguise. Bezos didn't invent expected utility. He applied it to organisational design. Reversible decisions with bounded downside and unbounded upside: expected utility favours speed. The cost of delay exceeds the cost of being wrong. Amazon's culture of "disagree and commit" and "bias for action" is expected utility operationalised. The lesson: when the decision structure matches the two-way door pattern, act. Deliberation has negative expected utility.
Prospect Theory doesn't invalidate EUT. It describes how we deviate from it. Use EUT as the normative benchmark; expect Prospect Theory in practice. When you're loss-averse, you'll reject positive-EV bets. When you overweight small probabilities, you'll buy lottery tickets and overpay for insurance. The strategic use: calibrate. Know when you're deviating from EUT and whether the deviation serves you or not.
The utility function is the hidden variable. Most decision analysis focuses on probabilities. The utility function is equally important. Two people with identical probability estimates can make opposite decisions because their utility functions differ. The founder who values legacy over liquidity will take risks the founder who values security will not. The investor with logarithmic utility will size positions differently than the investor with linear utility. Make your utility function explicit. It determines everything.
Reversibility changes the calculus. The two-way door insight generalises: when you can undo a decision, the expected utility of acting often exceeds the expected utility of waiting. The option value of delay is low when reversal is cheap. Most organisational delay comes from treating reversible decisions as irreversible — escalating, consensus-seeking, analysis-paralysing. The fix: classify first. Is this reversible? If yes, bias toward action. The expected utility math favours it.
Section 10
Test Yourself
Expected utility is claimed more often than it's applied. These scenarios test whether you can distinguish expected value from expected utility, apply the two-way door logic, and avoid the sunk cost fallacy. The key diagnostic: are you optimising the right metric? And are you using the right utility function? Each scenario has a correct answer that follows from the EUT framework. The wrong answers reflect common confusions — expected value vs expected utility, sunk costs vs future outcomes, delay vs deliberation.
Expected utility in practice
Scenario 1
A founder has two options: (A) Pivot to a new market with 30% chance of 10x growth, 70% chance of failure; (B) Stay the course with 80% chance of 2x growth, 20% chance of flat. Expected value of A is higher. The founder chooses B.
Scenario 2
Amazon's 'two-way door' framework says reversible decisions should be made quickly. A team delays a decision for six weeks to gather more data. The decision is reversible; the cost of reversal is low.
Scenario 3
An investor is offered a bet: 50% chance of doubling their money, 50% chance of losing half. Expected value is positive (1.25x). They decline.
Scenario 4
A company is deciding whether to launch a new product. The team estimates 40% chance of success (moderate revenue), 60% chance of failure (write-off of development cost). The CEO says: 'We've already spent $2M. We can't walk away now.'
Section 11
Top Resources
The literature on expected utility spans three centuries — from Bernoulli's 1738 paper to Von Neumann's axiomatisation to Kahneman's demolition of its descriptive validity. Read in order: Bernoulli for the intuition, Von Neumann for the rigour, Kahneman for the reality check. The applied material — Kelly, Bezos, Thiel — shows how the framework translates to capital allocation and organisational design. The primary sources are dense; the secondary material (Poundstone, Kahneman) is more accessible. Start with the applications if the theory feels abstract. The five resources below cover the full arc: origin (Bernoulli), formalisation (Von Neumann), descriptive reality (Kahneman), applied investing (Poundstone), and applied strategy (Bezos).
The foundational axiomatisation of expected utility. Dense but definitive. The proof that consistent preferences under uncertainty imply a utility function is the core result. Read the appendix on utility for the cleanest presentation. The axioms — completeness, transitivity, continuity, independence — are worth understanding; violations of independence explain many paradoxes. The book is primarily about game theory; the utility theory was developed as a foundation for analysing strategic interaction. It has since become the standard framework for decision theory.
Kahneman's Prospect Theory — how we violate EUT. Loss aversion, probability weighting, reference dependence. EUT is the benchmark; Kahneman shows the systematic deviations. Essential for understanding why expected utility is normative and how we actually behave. The chapters on prospect theory and loss aversion are the most directly relevant. Kahneman and Tversky's original 1979 paper "Prospect Theory: An Analysis of Decision Under Risk" is the academic source; the book makes it accessible. Read both if you want the full picture.
The Kelly Criterion — expected utility with log utility. Poundstone traces the history from Shannon to Thorp to Buffett. The connection between Kelly and expected utility is explicit. Readable and applied. Kelly maximises long-run growth; it's the optimal strategy when you have repeated opportunities and log utility.
Bernoulli's resolution of the St. Petersburg paradox introduced expected utility. The insight that utility diminishes with wealth predates Von Neumann by two centuries. Short and accessible. The paradox: a game with infinite expected value that no one would pay much to play. Bernoulli's resolution: maximise expected utility, not expected value.
Bezos's letters and essays on two-way doors, type 1 vs type 2 decisions, and bias for action. Expected utility applied to organisational design. The practical application of EUT to decision speed. The 2015 letter on "disagree and commit" and the 2016 letter on "day one" are particularly relevant. The two-way door concept appears in the 2015 letter: "Some decisions are consequential and irreversible or nearly irreversible — one-way doors — and these decisions must be made methodically, carefully, slowly. But most decisions aren't like that — they are changeable, reversible — they're two-way doors. And those decisions can be made by a high-judgment individual or small group." That's expected utility in plain language.
Expected Utility — Rational choice under uncertainty: multiply each outcome by its probability, apply utility, sum. Not expected value.
Reinforces
Risk [Tolerance](/mental-models/tolerance)
Risk tolerance is the shape of your utility function. Risk-averse: concave (diminishing marginal utility). Risk-neutral: linear. Risk-seeking: convex. EUT says optimal decisions depend on your risk tolerance. The two are inseparable — you cannot apply EUT without specifying risk tolerance. Most people don't know their risk tolerance explicitly. They discover it when they face a real trade-off. The exercise of applying EUT forces the question: how much would I pay to avoid a 50% chance of losing X? The answer reveals the shape of your utility function.
Tension
[Loss Aversion](/mental-models/loss-aversion)
Loss aversion — losses hurt roughly twice as much as equivalent gains feel good — violates EUT's assumption of consistent evaluation. EUT would treat gains and losses symmetrically (in utility space). Loss aversion makes us reject positive-EV gambles when the downside is framed as a loss. The tension: EUT says take the bet; loss aversion says no. The strategic implication: reframe. Present the same choice as avoiding a loss rather than securing a gain (or vice versa) and people's choices shift. EUT is frame-invariant. People are not.
Leads-to
Utility
Utility is the satisfaction or value derived from an outcome. EUT requires a utility function. The concept of utility — subjective, not directly observable — is what separates EUT from expected value. Understanding utility shapes how you apply EUT: know what you're optimising for. The shape of the utility function (concave, linear, convex) determines risk preference. Concave: risk-averse. Linear: risk-neutral. Convex: risk-seeking. Most people are concave for gains, convex for losses — the combination that produces Prospect Theory's predictions.
Probability estimation is the bottleneck. Most people can articulate utilities — what they want, what they'd pay to avoid. Probability estimation is harder. We're overconfident in our predictions. We anchor on recent experience. We neglect base rates. The expected utility framework assumes you have probabilities. In practice, the probabilities are the weak link. Improve your probability calibration — track predictions, update on outcomes — and your expected utility calculations improve. The framework is only as good as its inputs. Buffett's circle of competence is a probability calibration strategy: stay where you can estimate probabilities reliably. Outside the circle, expected utility is undefined.
The framework scales. Expected utility applies to individual decisions and to portfolio-level decisions. A single bet has an expected utility. A portfolio of bets has an expected utility that depends on the correlation structure — diversified bets have different portfolio utility than correlated bets. The same logic applies to a company's project portfolio: the expected utility of the portfolio is not the sum of individual project expected utilities when projects are correlated. This is why concentration can be rational: if your highest-conviction bets are uncorrelated, the portfolio expected utility may favour concentration over diversification. The math gets complex. The principle doesn't.
