What decision theory is, and why leaders go wrong even with enough data
The theory studies the reasoning behind a choice, from picking a bus or a taxi to a powerful person deciding whether to pursue major social change. That is how the Stanford Encyclopedia of Philosophy defines it [6]. Its entry was first published on 16 December 2015 and revised on 20 August 2025.
The theory has two parts, and the normative one says how a rational agent should choose. Under uncertainty, expected utility theory says that “one should prefer the option with greatest expected desirability or value” [6]. The descriptive part shows how people actually choose. The Nobel committee’s 2002 summary records the gap: real decision makers often don’t judge uncertain events by the laws of probability [2].
The field also has Russian roots. In 1939 the Soviet mathematician Leonid Kantorovich wrote an essay on the efficient use of resources in individual enterprises. He applied a mathematical method, linear programming, to planning early on. On 14 October 1975 he shared the Nobel memorial prize in economics with Tjalling Koopmans for work on the optimum allocation of resources [5].
In the introduction to my book How Not to Go Bankrupt!? [12], I described how a leader’s work has changed. Advice turns into a dialogue with artificial intelligence, and market analysis into checks against countless algorithms. Decision making becomes a balance between logic and intuition, between reason and the heart. I believe the theory exists for exactly this balance: it shows where habits of thought quietly replace logic.
Certainty, risk and uncertainty: where the line runs
The line runs where the odds can be measured: Frank Knight’s Risk, Uncertainty and Profit (1921) separated measurable risk from unmeasurable uncertainty [7]. In his words, “Uncertainty must be taken in a sense radically distinct from the familiar notion of Risk.”
The Stanford Encyclopedia traces two working terms back to Knight [6]. Choice under risk means the outcomes come with a known probability distribution. Choice under uncertainty means the decision maker relies on their own probability beliefs.
| Situation | What you know | Example in a company | What to do |
|---|---|---|---|
| Certainty | The outcome of each option | A purchase at an agreed price from a trusted supplier | Compare the options and choose |
| Risk | The options and their odds from your own history | Product returns, late payments across your customer base | Calculate the expected value |
| Uncertainty | The options, but not the odds | A new market, a new product, a first AI project | Turn it into risk with a pilot |
Mistakes start when a leader confuses these situations. Some plug invented probabilities into a model and get a precise number with false confidence. Others refuse to calculate even where sales history already gives the frequencies.
A pilot turns uncertainty into risk, because it replaces guesses with your own frequencies. Vernon Smith, who shared the 2002 prize with the psychologist Daniel Kahneman, ran wind-tunnel tests: new market designs tried in the lab before real use [1]. The committee’s 2002 summary says his experiments shaped energy markets in Australia and New Zealand [2]. In a company, a pilot on one process plays this role; our glossary explains what an AI pilot projectAI pilot project. An AI pilot project is a short launch of an AI solution on one process, with a metric set in advance, to check the effect before a large investment. is.
In the introduction to my book I wrote that there are no guarantees, and every worthy path is always a risk. Knight adds one thing to that thought: risk can be calculated. In my practice, uncertainty must first be turned into risk.
Expected value: the calculation to make before the argument
The expected value of an option is the sum of its dollar outcomes weighted by their probabilities. It is a simplified form of expected utility, honest as long as the company can bear the possible lossloss function. A loss function in machine learning is a formula that turns a model’s error into one number: the smaller it is, the closer the model’s answers are to the right ones.. The Nobel committee’s 2002 summary calls expected utility theory the dominant one [2].
Amazon’s letter to shareholders for 2015, filed with the US Securities and Exchange Commission (SEC) in April 2016, gives a short example [8]. A bet with a 10% chance of a hundredfold payoff is worth taking every time, even though it loses nine times out of ten. Its expected return is 0.1 × 100 = 10 stakes, nine more than went in.
Now my own example in dollars, an illustration rather than market data: a company is deciding whether to launch a new product line. A full launch costs $300,000: success brings $600,000 net, and failure loses the whole stake. A two-month pilot costs $30,000 and, to keep it simple, shows the outcome honestly; a launch after it brings only $570,000, since the market won’t wait. The team puts the chance of success at 40%.
| Option | If it works (40%) | If it fails (60%) | Expected value |
|---|---|---|---|
| Full launch now | +$600,000 | −$300,000 | 0.4 × 600,000 − 0.6 × 300,000 = +$60,000 |
| Pilot first, then launch | 570,000 − 30,000 = +$540,000 | −$30,000 (pilot only) | 0.4 × 540,000 − 0.6 × 30,000 = +$198,000 |
| Walk away | 0 | 0 | 0 |
The pilot wins by more than three times, though it brings less in the good case: it cuts the loss in the bad case from $300,000 to $30,000. The answer depends on the chance of success. This table is a decision tree in its smallest form: three branches, two outcomes each, weighed by one probability.
| Chance of success | Full launch | Pilot first | Choose |
|---|---|---|---|
| 20% | −$120,000 | +$84,000 | Pilot |
| 40% | +$60,000 | +$198,000 | Pilot |
| 60% | +$240,000 | +$312,000 | Pilot |
| 90% | +$510,000 | +$483,000 | Full launch |
Below a 33% chance, a full launch loses money on average; up to about 82%, a pilot first pays more; above that, launch at once. The threshold is the point: the team stops arguing about faith in the project and argues about whether the chance is above or below 82%. If the worst outcome would leave the company short of cash, the average no longer counts: I choose the option whose worst outcome the company survives.
The calculation has two limits. It is honest only where the odds can be measured or have been tested by a pilot. And it doesn’t protect you from biases: in the Allais paradox, which the encyclopedia covers, many people choose inconsistently between lotteries [6]. A number won’t choose for you, but it makes the argument concrete.
Expected utility, as the Stanford Encyclopedia describes it, weighs each outcome by its utility, what it is worth to the decision maker, not by its sum alone [6]. At 90%, a full launch beats the pilot on average: +$510,000 against +$483,000. Yet a 10% chance of losing $300,000 can sink a small company, and then the pilot is wiser.
Writing about choosing investors, I said in my book that entrepreneurship is risk, but I wanted my mistakes not to be fatal [12].
Cause one: we count from a reference point, not from the outcome
People reactReAct. A ReAct agent is an AI agent that works in turns: it reasons, calls a tool, looks at the result and reasons again. more to a change from a reference point, often the status quo, than to the final outcome, and fear a loss more than an equal gain. That is how the Nobel committee’s 2002 summary describes the findings of Kahneman [2]. With Amos Tversky he built prospect theory as an alternative to expected utility theory [1]. This is where cognitive biases begin.
The summary lists everyday effects: costly insurance for small appliances, and a long drive for a few dollars off a cheap item but not an expensive one [2]. Richard Thaler, who won the 2017 prize, added mental accounting: people sort money into separate accounts in their heads. He also described the endowment effect, where people value an item more once they own it [4].
I see this in companies all the time. A project stays alive because money has already gone into it, though that money won’t come back whatever you choose. A department spends its budget by year end because it is its own account. In my book I wrote that behavioral economics, built in the late 20th century by scientists such as Kahneman, showed that emotions sway money decisions and often bias them [12].
The fix for this cause: judge each option by where the company will stand in a year, not by yesterday’s spending. The question for the meeting is simple: if we had spent nothing yet, would we start this project today?
Cause two: heuristics stand in for probability
Under uncertainty, judgment takes shortcuts, called heuristics, that systematically depart from the basic rules of probability. This is Kahneman’s second finding, according to the Royal Swedish Academy of Sciences’ 2002 press release [1].
The committee’s 2002 summary describes three experiments [2]. People thought a day with over 60% boys among newborns equally likely in a small hospital and a large one, though small samples vary far more. An investor who sees a fund manager beat the index two years running believes in their skill, though the evidence is weak. A person keen on politics was labeled a member of parliament rather than a salesperson, though salespeople are far more common.
In a company, these mistakes look ordinary. A verdict on a new channel is based on two good months, and a candidate is hired for resembling the star of the team. A survey is trusted as if it were the customer. Amazon’s 2016 letter (SEC filing, April 2017) warns that 55% satisfied beta testers, up from 47% in the first survey, is a hard figure to read [9].
I tested my own startup ideas in the wrong place too. In my book I admit that I tried to test ideas on the reaction of people around me, and it was pointless: their judgment didn’t reflect the real potential [12]. The fix: before you conclude, ask how often this happens in general and how many cases the conclusion rests on. Two good months and five reviews from friends are not a sample.
Cause three: rationality is bounded, so a company settles for acceptable

A company picks a satisfactory option, not the optimal one, as Herbert Simon showed [3]. On 16 October 1978 the Royal Swedish Academy of Sciences awarded him the Nobel memorial prize in economics for his research into the decision-making process in economic organizations.
Per the 1978 press release, Simon replaced the all-knowing entrepreneur of classical theory with several people [3]. Their rationality is limited by gaps in knowledge and by personal and social ties. In its words, “Individual companies, therefore, strive not to maximize profits but to find acceptable solutions to acute problems.” His main book, Administrative Behavior, came out in 1947.
Kantorovich reached a similar conclusion from another side. According to the 1975 press release, authority can be pushed down only with a rational price system [5]. That system includes one standard interest rate for investment. For a company, I read this as follows: units decide well on their own when they share one criterion and one cost of money.
My practical conclusion: write down what counts as good enough, and the cost of waiting, before you look for options. Otherwise the satisfactory option gets picked out of fatigue. In my book I wrote that leaders who think ahead soften the impact of uncertainty, make fewer spontaneous decisions and build a solid base for steady growth [12]. Our analysis of risk-based inspection shows where a model helps decide what to check first.
Cause four: one heavy process for reversible and irreversible decisions
A company loses speed when it applies one slow process to both irreversible and reversible decisions [8]. Amazon’s 2015 letter to shareholders (SEC filing, April 2016) explains why. The letter calls irreversible decisions one-way doors and says to make them slowly and carefully. Most decisions can be undone; these are two-way doors, and small groups should make them fast.
According to the letter, a growing organization tends to use the heavy process for most decisions. The result is slowness, unthoughtful risk aversion, too little experimentation and fewer inventions [8].
The 2016 letter (April 2017) sets a threshold: “most decisions should probably be made with somewhere around 70% of the information you wish you had” [9]. If you wait for 90%, you are usually slow. You need to spot and fix bad decisions quickly; then a mistake costs less than it seems.
My own threshold: if you can undo a decision within a week, at a cost the team can write off alone, decide in a small group at 70%. If you can’t undo it, or undoing is costly, run the full process: calculation, premortem and a written record. In the introduction to my book I wrote that mistakes are not enemies but teachers [12]. A reversible choice gives that lesson fast and cheap.
How to test a decision before it gets expensive
The first test happens before launch, in a premortem, and costs only the team’s time [10]. Gary Klein described it in Harvard Business Review in September 2007: the team assumes the project has already failed and looks for the reasons. Gather the people who will carry out the decision and say that a year has passed and the project failed. Give everyone ten minutes to list reasons, then match each with a step before the start.
I described the second method in my book for talks with investors [12]. Model several scenarios, from a sudden change of terms to unexpected demands, and prepare an answer for each so nothing takes you by surprise. Inside a company, that means three or four scenarios, from rising prices to losing a key customer.
The third method is for offers that look too good. In my book I recallprecision and recall. Precision and recall are two quality metrics of a model. Precision shows what share of the found items is correct. Recall shows what share of the needed items the model found. the Dutch tulip mania, which burst in 1637 [12]. My advice there: if an opportunity seems too good to be true, step back, check the documents, and talk to a lawyer, analysts and colleagues.
The fourth method costs more but gives you your own data: a trial run on one process. We have described separately how to run a short AI agent pilot, how to measure its result and why agent pilots fail.
I know the price of an untested decision from my own life. In How Not to Go Bankrupt!? I told how a story with fraudsters posing as investors cost me 2 months, a lot of nerves, about $1,000 and the illusion that business is easy [12]. I believe any of the four tests would have cost me less.
Data, AI and intuition: what the machine calculates and what stays with the leader
Data and models narrow the gaps in knowledge that Simon described [3], but the goal, the choice and the responsibility stay with a person. This is my position, and I showed how it works with numbers and pilots in our piece on data driven decisions with AI. The same logic of options, cost and reversibility applies to choosing between building and buying an AI solution.
In my book I wrote about the technology itself with caution. AI already helps us by analyzing data and suggesting decisions, but will it always be so? Technology changes, and we can’t be sure it won’t turn from a support into a trap tomorrow [12].
I don’t set intuition against calculation. In the chapter on leadership I wrote that intuition adds to logic by catching patterns hidden in the data [12]. Leaders who trust it wisely can spot new opportunities or threats earlier.
Science puts a limit on intuition. According to the Nobel committee’s 2002 summary, judgment in an uncertain situation slides into heuristics that systematically depart from the principles of probability theory [2]. So my rule is this: I trust my gut where I have experience and fast feedback, and in a new situation I check it with a calculation.
What Russian law says about a director’s failed decision
Russian law judges how a director made a decision, not only how it turned out. This follows from Resolution No. 62 of 30 July 2013, on losses caused by a company’s managers [11]. The Plenum of Russia’s Supreme Arbitration Court, then the top court for commercial disputes, adopted it; such resolutions tell commercial courts how to apply the law.
Under paragraph 1, negative consequences for the company “do not in themselves indicate bad faith and (or) unreasonableness of his actions (inaction)” (translated) [11]. The reason is given in the same paragraph: such consequences “accompany the risky nature of business activity” (translated).
Paragraph 3 lists cases in which a director’s unreasonableness counts as proven. One is a decision made “without taking into account information known to him” (translated) that matters in the situation. Another is when, before deciding, the director took no steps to obtain “the information necessary and sufficient for making it” (translated) [11].
A court looks at the same thing as the normative theory: was information gathered, and were the options weighed. That is why, for irreversible decisions, I recommend a written record: options, calculation, information sources with dates, risks and who objected. In the chapter of my book on investment illusions I wrote that critical thinking, legal advice, analytics and human awareness are your main allies [12]. This is an analysis of the document’s logic, not legal advice.
The decision process on one page, and the main objection
For a choice that is hard to undo, I use five steps, and each rests on a source from this analysis.
- Decide the type: reversible or not. Amazon’s 2015 letter advises against one slow process for every decision [8].
- Name the situation: risk or uncertainty in Knight’s sense (1921) [7]. Under uncertainty, run a pilot first.
- Calculate the expected value of each option with the table above, and find the chance at which the choice flips.
- Run a premortem and write out three or four scenarios with an answer to each.
- Record the decision, the information sources and who disagreed: this shows the information was gathered before the decision (paragraph 3 of Resolution No. 62 of 2013) [11]. For a reversible decision, about 70% of the information is enough (Amazon’s 2016 letter) [9]. Set a review date: compare the outcome with the calculation and the record, and decide whether the process failed or you were unlucky.
The main objection: theory is academic, and in business the gut decides. Simon’s work (prize of 1978) and Kahneman’s (prize of 2002) show that in an unfamiliar situation judgment departs from probability in systematic ways. So in a new and irreversible decision, a process matters most for those used to deciding by gut.
My conclusion matches what I wrote in my book [12]. Relying on verified information lowers the chance of falling into a trap; emotions inspire, but at the decisive moment you need to keep your balance and resist illusions. In practice, that means three habits: calculate expected value, separate reversible decisions from irreversible ones, and write down how you decided.


