A board asks you for "the risk" on a new product, and you give them a number. But press on it for a second: do you actually know the odds, the way a casino knows the house edge, or are you handing over a guess dressed up as a probability? Most of the time it's the second one. And that gap, between a known chance and a confident-sounding hunch, is where good decisions quietly go wrong.

The quick version

  • Risk = you don't know the outcome, but you can put trustworthy odds on it (a fair die, an insured event). The move: calculate and optimise.
  • Uncertainty = you can't put trustworthy odds on it; you don't even know the full range of what could happen. The move: stay flexible, run small probes, buy information.
  • Ambiguity = the odds themselves are unknown or contested, the data is missing, vague or means different things to different people. The move: reduce it before you decide; don't let it freeze you.
  • Naming which one you face is half the job. Treating an uncertainty as a risk is how confident spreadsheets produce expensive surprises.

The idea in depth: three problems, not three synonyms

The cleanest line here was drawn over a century ago. The economist Frank Knight, in Risk, Uncertainty and Profit (1921), separated two things that everyday language smears together. Risk, he said, is "a quantity susceptible of measurement", situations where you can assign reliable probabilities. Uncertainty is the rest: outcomes whose probabilities are "not capable of measurement" at all. As an MIT explainer later put it, risk is when "we do not know the outcome… but can accurately measure the odds," whereas uncertainty is when "we cannot know all the information we need in order to set accurate odds in the first place" (MIT News, 2010).

That distinction is still called Knightian uncertainty, and it matters because the two demand opposite habits. With genuine risk, the rational thing is to compute the expected value and optimise, this is the home turf of decision theory and expected value. With genuine uncertainty, optimising a number you made up is theatre. The habit that saves you is small: before you build the model, ask where these probabilities actually came from. If the honest answer is "we don't really have them," stop treating the output as a forecast. It's a scenario.

flowchart TD
  Q(["Facing a decision"]) --> A{"Can you put trustworthy odds on the outcomes?"}
  A -->|"Yes, reliable probabilities"| R("RISK
Calculate & optimise expected value")
  A -->|"No, but you know the possible outcomes"| U("UNCERTAINTY
Stay flexible, probe, preserve options")
  A -->|"The odds themselves are missing or contested"| M("AMBIGUITY
Reduce it: buy information, align definitions")
One question, "can you put trustworthy odds on it?", sorts most decisions into the right bucket. Leaders Loop

Where ambiguity enters, and why it freezes people

Knight gave us two buckets; a third idea sharpens the second. In 1961 the decision theorist Daniel Ellsberg published "Risk, Ambiguity, and the Savage Axioms" in the Quarterly Journal of Economics, and ran a now-famous thought experiment. Imagine an urn with 90 balls: 30 are red, the other 60 are some unknown mix of black and yellow. Bet on red and you know your odds exactly (30 in 90). Bet on black and you don't, it could be anywhere from 0 to 60. Most people strongly prefer the red bet, even when the maths says the two are equivalent. That preference for the known chance over the unknown one is ambiguity aversion, and it's so reliable it now carries Ellsberg's name (the Ellsberg paradox).

Ambiguity, then, isn't quite the same as Knightian uncertainty. Uncertainty is about the world being genuinely unpredictable. Ambiguity is about your information being thin, vague, or contested, you can't even pin down the odds because the underlying data is missing or two stakeholders read it differently. The practical danger is specific: people are so averse to ambiguity that they'll over-pay for false precision, or stall entirely rather than act on a fuzzy picture. So treat ambiguity as a thing you can shrink. A week of customer interviews, a competitor teardown, a pilot, each converts "we have no idea" into "we now have a rough range." You rarely buy your way to certainty. You can usually buy your way out of ambiguity.

You rarely buy your way to certainty. But you can almost always buy your way out of ambiguity.

The honest limitation: the boundaries are fuzzy

Here's where intellectual honesty earns its keep. The line between "measurable" and "unmeasurable" is itself a judgement, not a fact stamped on the situation. A Bayesian would argue you can always assign a probability, even to a one-off event, by stating your subjective prior and updating it as evidence arrives. On that view, Knight's hard wall between risk and uncertainty is really a spectrum of confidence in your numbers. That critique is fair, and worth holding. The value of the three labels isn't that nature comes pre-sorted into them; it's that they force the question Knight cared about, how much do I actually trust these odds?, instead of letting a precise-looking output hide a shaky input. Use the categories as a diagnostic, not as a law of physics.

One more framing is worth borrowing, because it's entered the language. In a 2002 briefing, US Defense Secretary Donald Rumsfeld distinguished "known knowns," "known unknowns," and "unknown unknowns" (DoD transcript, 12 Feb 2002). Mocked at the time, it maps neatly onto the three buckets: known knowns are settled facts, known unknowns are risks you can scope and price, and unknown unknowns are uncertainty proper, the outcomes not yet on anyone's list. The leadership skill is widening the circle of known unknowns so fewer things stay genuinely unknown.

flowchart LR
  KK(["Known knowns
settled facts"]) --> KU(["Known unknowns
scoped risks, price them"])
  KU --> UU(["Unknown unknowns
true uncertainty, probe for them"])
  UU -. "good leaders keep pushing this boundary left" .-> KU
The work is converting unknown unknowns into known unknowns, then pricing them. Leaders Loop

A worked example: the £400k market-entry call

You run a product unit. Your team proposes launching into a new country, and someone hands you a deck with a single headline: "Expected first-year return: £400k, with 15% risk of loss." The figures here are illustrative, but the shape of the problem is real, and it's a trap.

Pull the "15% risk" apart. Where did it come from? If it's the failure rate of forty comparable launches you've already run in similar markets, that's close to genuine risk: you have a reference class, the odds are earned, and optimising against them is sound. You can insure it, hedge it, or size the bet accordingly.

But suppose this is your first launch in a region with different regulation, an unfamiliar competitor, and no comparable history. Now that "15%" is not a probability, it's a feeling with a decimal point. The real situation is uncertainty (you don't know the true range of outcomes) wrapped in ambiguity (your local-market data is thin and your finance and sales teams read it differently). The single worst response is to argue about whether the number should be 15% or 25%. You're polishing a guess.

The better response follows the three buckets. Name it: "This is uncertainty, not risk, the number is a placeholder." Shrink the ambiguity: spend a small, fixed sum on the cheapest information that would most change your mind, a regulatory opinion, ten customer interviews, a four-week soft launch in one city. Stay flexible against the uncertainty: structure the entry as a staged commitment with a real exit, not an all-in bet, the logic of real options. You've spent a little to convert an unmeasurable hunch into a roughly measurable range, and kept the option to walk away cheaply if the probe comes back ugly. Name it, shrink it, stay flexible, that's most of the method, and it beats arguing about a decimal point.

Frequently asked questions

Isn't this just semantics? A risk is a risk.

The words are cheap; the consequences aren't. Call a true uncertainty a "risk" and you'll optimise a fabricated number, commit hard, and get blindsided by an outcome that was never in your model. Call a manageable risk an "uncertainty" and you'll dither over something you could have priced and decided in an afternoon. The label changes which tool you reach for, so getting it right is the opposite of semantics.

If I can't measure it, why not just guess a probability and move on?

You can, Bayesians do exactly that, and a stated prior beats a hidden one. The trap is forgetting it was a guess. The discipline isn't refusing to put a number down; it's labelling how much weight that number can bear, and committing more cautiously the thinner the basis. Cheap information first, big commitment second.

How do I tell genuine uncertainty from a team that just hasn't done the homework?

Ask one question: "What would we have to learn to put real odds on this, and can we afford to learn it?" If a fortnight of work would produce a defensible reference class, it's homework, not uncertainty, send them to do it. If no feasible amount of digging would pin the odds down, it's genuine uncertainty, and your job shifts from forecasting to staying flexible.

Doesn't ambiguity aversion mean I should just avoid ambiguous options?

No, that's the bias, not the lesson. Ellsberg's experiment shows people over-avoid the unknown bet even when it's a fair deal, and in business the ambiguous opportunities are often the under-competed ones precisely because rivals flinch. The point is to notice the flinch, then decide whether the ambiguity is reducible. Reduce it where you can, price the rest, and don't let discomfort alone make the call.

Where does plain old complexity fit in?

Snowden and Boone's Cynefin framework (HBR, 2007) is the natural companion: it sorts situations into clear, complicated, complex and chaotic, and prescribes a different decision mode for each. Roughly, "clear/complicated" is the land of risk (analyse, then act), while "complex/chaotic" is the land of uncertainty (probe and sense before you commit). Same instinct, different lens.

Related in the Toolkit

Where to go next