Case mode · fictional firm, illustrative numbers
Maison Delaunay: a strategy case you solve inside a digital twin
Case mode turns an MBB-style case interview into an investigation inside a live model. You do not get a tidy prompt with all the data. You spend tokens to lift the fog on parts of the firm, interview its agents, bet on hypotheses before testing them, and compare your strategy with what an optimiser finds. Here is a full game on a fictional firm, with the trap most players fall into.
On this page
What is wrong at Maison Delaunay?
Brief
Maison Delaunay Search is an 18-person executive search firm in Lyon. Over 12 months, revenue fell 9% while signed mandates rose 12%. EBITDA margin went from 22% to 14%. Hélène, the founding partner, wants to sell in three years.
Question: how do you get back to 20% margin within 12 months without making the firm harder to sell?
Rules: 12 investigation tokens, 90 minutes, the whole map under fog at the start.
Maison Delaunay, Hélène and every number on this page are invented for the exercise.
How is a case played in the twin?
| Loop | Who plays | With what | What it brings |
|---|---|---|---|
| Game | The player: founder, partners, a consultant or a buyer | Flow map, investigation tokens, agent interviews, A/B universes | Exploration, intuition, ownership of the answer |
| Optimisation | The machine | A multi-objective search over the levers, with a margin, risk and dependence frontier | The measured gap between the best human idea and the best possible one |
| Strategy | The twin in crisis mode | Hundreds of random futures, a minimum-regret rule | A strategy that holds in many futures, not just the likeliest one |
How did the investigation unfold?
| Turn | Player action | Tokens | What the twin reveals | Effect on hypotheses |
|---|---|---|---|---|
| 1 | Draws the tree: revenue = mandates × fee × completion rate; costs = hours per placement | 0 | The partner agent points out that "price" and "discounts" overlap | Tree fixed, mutually exclusive |
| 2 | Bets on three hypotheses: discounts 40%, completion rate 30%, delays 30% | 0 | The bets are timestamped in the log | Starting point |
| 3 | Asks for the average discount by quarter | 1 | It went from 3% to 5% | Discounts drop to 20%: too small to explain −9% |
| 4 | Asks for the completion rate of mandates | 1 | 78% down to 61%, including 7 mandates cancelled mid-search | Completion rate up to 50% |
| 5 | Lifts the fog on the flow map | 1 | The "partner review" station blinks: 14 shortlists waiting; 7 active mandates per consultant instead of 4 | New hypothesis: founder bottleneck |
| 6 | Interviews the "senior consultant" agent | 2 | "Everything waits for Hélène's review, and we juggle seven searches." Confidence: green | Founder bottleneck at 60% |
| 7 | Interviews the "HR director, mid-cap" client agent | 2 | "We cancelled two searches: too slow, a competitor showed candidates in three weeks." | Mechanism confirmed: delay, then cancellations, then revenue |
| 8 | Applies Little's law on the map | 0 | Time to shortlist went from 5.5 to 9 weeks, consistent with the work in progress | Root cause: too much work in progress behind a founder bottleneck |
Seven tokens out of twelve. The opening intuition, price, was a false lead. The root cause is too many searches in progress, queued behind the founder. It is the classic trap of a case interview, except the player found it by playing instead of reading it in the prompt.
Which strategy wins when tested in the twin?
Each universe runs 1,000 times. Bars show EBITDA margin at 12 months from P10 to P90, the dark tick is the median, the dashed red line is the 20% target.
| Universe | Levers | Margin at 12 months (P10 / P50 / P90) | Shortlists reviewed by the founder | Verdict |
|---|---|---|---|---|
| A | Cap of 4 active mandates per consultant, waiting list | 15 / 18 / 20% | 60% | Good lever, not enough alone |
| B | Review delegated to two seniors; the founder sees mandates above €150k only | 14 / 17 / 19% | 15% | Gains margin and makes the firm easier to sell |
| C | No more discounts, fee up 5 points | 11 / 16 / 19% | 60% | Two client types leave: high risk |
| A + B + D | A and B, plus an AI sourcing agent (−40% search time) | 18 / 21 / 23% | 15% | Chosen; the bottleneck moves to client interviews |
| Optimiser | Best combination found on the frontier | 18 / 22 / 24% | 10% | Adds C, for new clients of one segment only |
- Player against machine. A + B + D reaches about 95% of the optimum at the median (21% against 22%), with comparable risk.
- The bottleneck moves. Once review is unblocked, the map makes the next queue blink: client interviews. The game carries on next turn, the way Factorio does.
- Stress test. In crisis mode (top client leaves, mandates −15%, a senior quits), A + B + D keeps at least 18% margin in 8 futures out of 10; C alone drops under 12% in one future out of three.
What does the final answer sound like?
Answer first: getting back to 20% goes through work in progress and shortlist review, not price. Three moves: cap work in progress at four mandates per consultant, delegate review to seniors, equip sourcing. Expected: 21% margin (18 to 23%) and a founder present on 15% of shortlists instead of 60%. Main risk: perceived shortlist quality. Guardrail: the founder keeps mandates above €150k for six months.
Then the partner agent asks the question that kills weak answers: "What would make you change your mind?" The reply becomes an alert criterion, and the decision enters the prediction log, checked at 3, 6 and 12 months.
| Score | What is measured |
|---|---|
| Value | Result of the strategy against the optimiser's frontier |
| Robustness | Share of simulated futures that reach the goal, shocks included |
| Investigation economy | Tokens spent to reach the root cause |
| Discipline | Hypotheses bet before tests, well-calibrated confidence (Brier score) |
| Synthesis | Answer first, mutually exclusive structure, numbers, risks, a change-of-mind criterion |
Who plays cases like this, and why?
- Founders and partners, on their own firm twin, in an evening or a one-day war game. On a real firm there is no planted answer: the case becomes a real investigation and the prediction log settles it at 3, 6 and 12 months.
- Buyers in due diligence, on a twin of the target built from the data room, before they sign.
- Teams in training, on fictional firms of increasing difficulty, as in a classroom business simulation, where the designer plants a root cause and false leads like a level designer.
One guardrail: cases are never used to score job candidates. An AI system that evaluates candidates is high-risk under the EU AI Act, so the exercise stays run and judged by a person.
Play this case on your own firm.
Your firm has its own Maison Delaunay question. A scoping call turns it into the first question your twin answers. Firm twin build from $55,555.
- Warm up with Twin Lite: the same mechanism on ten of your numbers.
- Read how a firm twin is built.
- Book 30 minutes.