Most candidates who say "decision tree" draw an issue tree and call the branches decisions. The interviewer is listening for something narrower: controllable choices at square nodes, uncertain outcomes at chance nodes, and a recommendation that can refuse the highest expected value when risk, irreversibility, or the value of waiting for information says so.
That vocabulary is not case folklore. It comes from decision analysis as Howard Raiffa set it out in Decision Analysis: Introductory Lectures on Choices Under Uncertainty (1968): trees that make choices and chance explicit, with numerical probabilities, so you can fold the tree back to a strategy. Ronald Howard then showed that uncertainty has an economic price: the value of information is not Shannon entropy; it is how much better you can choose once you know.
After this piece you will be able to build a genuine decision tree under interview pressure, compute expected value without getting seduced by it, and say out loud when a pilot beats a high-EV launch.
What a decision tree is (and what it is not)
A decision tree answers one question: given alternatives I control and outcomes I do not, which strategy should I pick now?
| Node | What it holds | Interview signal |
|---|---|---|
| Decision | Mutually exclusive acts the client can choose (launch, pilot, wait, exit) | You name what management can still decide |
| Chance | Uncertain states with probabilities (demand high/low, competitor enters) | You refuse to pretend the future is known |
| Terminal | Payoff or consequence under that path | You can score the tree, not just narrate it |
An issue tree answers a different question: what could explain this problem, or what topics must we cover? It decomposes a diagnosis. A driver tree (see the sibling piece on driver trees) decomposes a metric into algebraic drivers so you can find where the number moves. Neither structure starts with a choice under uncertainty.
I stop listening when a candidate opens a go/no-go prompt with "I'd look at company, customers, and competitors." That is a topic list. The decision is already on the table. The room wants options, chance, and a way to pick.
The fold-back: expected value as a benchmark, not a verdict
You solve a tree from right to left. At each chance node, compute expected value: sum of probability times payoff. At each decision node, pick the branch with the best expected value (or the best expected utility if risk aversion matters). That rollback is the arithmetic engine Raiffa popularized for business and policy choices.
In a case, the interviewer rarely needs a perfect utility function. They need to see you:
- Keep choices mutually exclusive and exhaustive enough for the prompt.
- Put probabilities on chance, even if they are your judgmental estimates.
- Multiply and add without double-counting paths.
- State the recommendation as a strategy ("pilot, then scale only if X"), not as a hope.
Expected value is the long-run average if the same decision were repeated many times. Most client decisions are one-shot. That is why EV is a rational benchmark and a frequent trap in the same breath.
Where expected value misleads
Three places the high-EV branch is the wrong recommendation in the room.
Risk. Two paths can share the same EV while one has a ruinous downside. A cash-constrained client, a partner with a board that hates variance, or a firm that cannot survive a -$18M hit will not take the EV-max path. Say the risk profile out loud. If payoffs differ mostly in the left tail, EV alone is incomplete.
Irreversibility. Full launch commits capacity, brand, and channel relationships. A pilot or wait keeps options open. Real options matter when the cost of reversing is high and learning is cheap. High EV that burns the option to wait is not automatically best.
Information value. Sometimes the best "decision" is to buy information first. Howard's information value theory treats the worth of resolving uncertainty as an economic quantity: how much better you choose after you know. The ceiling is the expected value of perfect information (EVPI): value of deciding with the true state revealed, minus the best EV under current beliefs. Modern health-economics guidance on value of information uses the same logic. In a case, if a $1.2M pilot can change the launch decision and EVPI is $4M, the pilot is not dithering; it is the strategy.
Ned's rule. If your highest-EV branch is irreversible and the EVPI of a cheap test exceeds the cost of that test, recommend the test. The interviewer is scoring judgment under uncertainty, not arithmetic bravado.
Worked example: HarborLane's roast-plant choice
Invented client, consistent numbers. HarborLane Roasters, a mid-size specialty coffee supplier, must choose how to add roasting capacity for a new grocery channel. Three acts:
- Build a dedicated plant now ($12M capex embedded in payoffs below).
- Contract with a co-packer for three years (lower upside, lower downside).
- Pilot a six-month limited SKU run with the co-packer, then decide whether to build.
Uncertainty that matters: grocery demand over five years. Judgmental probabilities from management (treat them as given):
| Demand state | Probability |
|---|---|
| Strong | 0.40 |
| Base | 0.35 |
| Weak | 0.25 |
Five-year incremental NPV payoffs ($M), net of the path's costs:
| Path | Strong (0.40) | Base (0.35) | Weak (0.25) |
|---|---|---|---|
| Build now | +22 | +8 | -18 |
| Contract | +11 | +5 | -2 |
| Pilot then decide | see fold-back | see fold-back | see fold-back |
Step 1 — EV of the pure strategies
Build: (0.40(22) + 0.35(8) + 0.25(-18) = 8.8 + 2.8 - 4.5 = 7.1)
Contract: (0.40(11) + 0.35(5) + 0.25(-2) = 4.4 + 1.75 - 0.5 = 5.65)
On raw EV, Build wins at $7.1M vs Contract at $5.65M.
Step 2 — Pilot as a strategy (imperfect information, then act)
Pilot cost: $1.2M. After the pilot, HarborLane observes a signal that is imperfect but useful. For this interview math, simplify: with probability 0.55 the signal says "scale" and with 0.45 it says "stop." Conditional on "scale," management's updated view is Strong 0.55 / Base 0.30 / Weak 0.15. Conditional on "stop," it is Strong 0.15 / Base 0.35 / Weak 0.50. After "scale," they Build; after "stop," they Contract (they still need some capacity for the channel commitment).
EV if signal = scale, then Build:
(0.55(22) + 0.30(8) + 0.15(-18) = 12.1 + 2.4 - 2.7 = 11.8), minus pilot already paid separately.
EV if signal = stop, then Contract:
(0.15(11) + 0.35(5) + 0.50(-2) = 1.65 + 1.75 - 1.0 = 2.4)
EV of Pilot strategy before subtracting cost:
(0.55(11.8) + 0.45(2.4) = 6.49 + 1.08 = 7.57)
Net of $1.2M pilot: $6.37M.
So Pilot nets below Build on this imperfect signal. A weak candidate stops here and builds. A stronger one asks whether perfect (or better) information would change the answer.
Step 3 — EVPI: the ceiling on learning
With perfect foresight of demand, the optimal act by state:
| State | Best act | Payoff |
|---|---|---|
| Strong | Build | 22 |
| Base | Build | 8 |
| Weak | Contract | -2 |
EV with perfect information: (0.40(22) + 0.35(8) + 0.25(-2) = 8.8 + 2.8 - 0.5 = 11.1)
Best EV without information: Build at 7.1
EVPI = 11.1 - 7.1 = $4.0M
A pilot that costs $1.2M and can approach that learning is economically interesting even when this particular imperfect signal lands at $6.37M. In the room, I want you to say: "Build has the highest EV, but EVPI is $4M and a weak demand path loses $18M. If the board cannot absorb that left tail, Contract or a cleaner pilot design beats Build."
Step 4 — When risk overturns EV
Suppose HarborLane's covenant blows up below -$10M in any scenario. Build's weak path at -18 is unacceptable; Contract's worst case is -2. Expected value still ranks Build first. The recommendation does not. That is the point of the tree: make the downside visible, not averaged away.
How this sounds in a first-round case
A clean opening for HarborLane-style prompts:
"I'd treat this as a decision under uncertainty. The acts are build, contract, or pilot-then-decide. The chance node is demand. I'd fold expected values, then check whether risk limits or the value of information overturns the EV ranking."
Then ask for the three numbers that close the tree: payoffs by path and state, probabilities, and any hard constraint (covenant, capacity, brand risk). Do not invent precision you were not given; label judgmental probabilities as judgmental.
If the interviewer never gives numbers, still draw the nodes. Structure without arithmetic still separates you from the candidate who lists "market attractiveness" for ten minutes.
Mistakes that make the tree look like decoration
Overlapping acts ("grow" and "enter groceries") that are not mutually exclusive. Chance nodes without probabilities, or probabilities that do not sum to one. Treating an issue tree as a decision tree. Maximizing EV and ignoring a hard risk limit. Recommending a pilot with no decision gate ("we'll learn something"). Comparing a driver breakdown of roast cost per pound to a go/no-go plant choice: wrong tool for the prompt.
For how structure fails more broadly, see consulting framework mistakes. For cost–benefit logic that pairs with trees, see cost-benefit analysis.
Drill this under a clock
Take any "should we launch / acquire / expand / build" prompt. Spend eight minutes: list three mutually exclusive acts, one chance node with two or three states, rough payoffs, fold the EVs, then write one sentence on whether risk or information value changes the pick. Then run a timed structure rep.
Everything on CoachNed is open for seven days, no card; then $120 for a recruiting season or $49 a month. For a full voice case where you have to choose under uncertainty out loud, try /interview. For the five-minute typed first rep, use /start.
CoachNed is not affiliated with McKinsey, BCG, Bain, Deloitte, or any firm named in interview prep.
Frequently asked questions
What is a decision tree in a consulting case interview?
It is a structure that starts with the client's controllable choices, branches into uncertain outcomes with probabilities, and ends in payoffs so you can recommend a strategy. It is not a list of topics to investigate; that is an issue tree.
How is a decision tree different from an issue tree?
An issue tree decomposes a problem into causes or workstreams. A decision tree compares acts under chance. Use an issue tree when you still need diagnosis; use a decision tree when the decision is already on the table.
When should I not use expected value alone?
You solve a tree from right to left. At each chance node, compute expected value: sum of probability times payoff. At each decision node, pick the branch with the best expected value (or the best expected utility if risk aversion matters).
When downside risk can break the client, when the high-EV act is hard to reverse, or when a cheap test has information value that can change the choice. EV is the baseline; risk and EVPI are the overrides.?
You solve a tree from right to left. At each chance node, compute expected value: sum of probability times payoff. At each decision node, pick the branch with the best expected value (or the best expected utility if risk aversion matters).
Do I need to draw a literal tree diagram on paper?
No. A compact table of acts, states, probabilities, and payoffs is often clearer and faster. The interviewer cares that the logic is a tree, not that the ink looks like one.
How do I practice decision trees for cases?
Use go/no-go and investment prompts, force yourself to name acts and chance nodes first, then practice fold-back math. Structure drills and live cases on CoachNed are built for that pressure.
Sources
- Larry Summers on His Decision-Making Process for the Auto Bailout (HKS excerpt citing Raiffa 1968) — primary-context citation that Raiffa's Decision Analysis introduced decision trees and numerical probabilities to business and government decision-making. Checked 2026-09-24.
- Information Value Theory (Ronald A. Howard, IEEE Transactions on Systems Science and Cybernetics, 1966) — primary source that the economic value of reducing uncertainty depends on probabilities and payoffs together, not on information entropy alone. Checked 2026-09-24.
- Value of Information Analysis for Research Decisions (ISPOR Task Force Report 1) — defines EVPI as the expected gain from eliminating uncertainty relative to deciding under current evidence. Checked 2026-09-24.
- Foundations of Decision Analysis (Ronald A. Howard and Ali E. Abbas, Pearson) — textbook lineage of Howard's decision-analysis cycle, including the informational phase that values learning. Checked 2026-09-24.