Interview experience
McKinsey Interview Questions and Answers (2026)
Overview
Section titled “Overview”McKinsey’s process runs the gamified McKinsey Solve assessment into case-plus-PEI interviews across two rounds, with McKinsey Digital/QuantumBlack as the dedicated track for engineering and data-science candidates.
McKinsey interview process at a glance
Section titled “McKinsey interview process at a glance”| Round | Duration | What they test |
|---|---|---|
| McKinsey Solve (Imbellus game) | ~65 min (2 of 3 modules) | Problem-solving process and analytical thinking via scenario games |
| Recruiter screen | - | Resume, background, motivation |
| First round interviews | 2 x 45-60 min | Case + PEI each, with Associates/Engagement Managers |
| Final round interviews | 2-3 interviews | More open-ended cases, deeper PEI, partner-level interviewers |
McKinsey Solve (Imbellus game)
Section titled “McKinsey Solve (Imbellus game)”An online, gamified assessment built on the former Imbellus test. Candidates typically complete two of three scenario-based modules - such as building a self-sustaining ecosystem or running a geological “redrock” study - in about 65 minutes. It scores your problem-solving process (how you explore, test, and decide), not just the final answer.
Common questions
- Ecosystem-building scenario: balance interacting variables to reach a stable state
- Redrock-study scenario: gather and interpret data under a resource/time constraint
- Scenario decisions that reward methodical exploration over rushing to an answer
Recruiter screen
Section titled “Recruiter screen”A short conversation on background, resume, and motivation before you’re routed to first-round interviews. It’s a lighter filter than Solve or the case rounds, but a mismatched story here can still cost you a slot.
Common questions
- Walk me through your resume
- Why McKinsey, and why this office/practice?
- What other firms or roles are you considering?
Round-by-round breakdowns are on the McKinsey interview experience page.
First round interviews
Section titled “First round interviews”Two interviews of 45-60 minutes each with Associates or Engagement Managers. Each pairs an interviewer-led case (a business problem you structure, analyse with exhibits, and turn into a recommendation) with a PEI question on leadership or personal impact.
Common questions
- A market-entry or profitability case with interviewer-provided exhibits
- Tell me about a time you led a team through ambiguity (maps to the Leadership PEI trait)
- Describe a time you had to build trust with someone skeptical of your approach (Connection)
- Walk me through how you’d structure this unfamiliar business problem
Final round interviews
Section titled “Final round interviews”2-3 interviews with partner-level interviewers. Cases get more open-ended with less interviewer scaffolding, and PEI questions dig deeper into your specific personal contribution versus the team’s outcome.
Common questions
- An open-ended case that shifts direction partway through
- Tell me about your biggest failure and what you changed afterward (Growth)
- Describe a time you showed resilience under real pressure (Drive)
- Why consulting, and why McKinsey over other offers?
Sample answer frameworks for each of these are on the McKinsey HR interview questions page.
McKinsey Digital / QuantumBlack: the track for engineering/CS-background candidates
Section titled “McKinsey Digital / QuantumBlack: the track for engineering/CS-background candidates”McKinsey hires generalist Business Analysts alongside McKinsey Digital and QuantumBlack (McKinsey’s AI and analytics arm), which is where most engineering, data-science, and CS-background campus candidates land. The core case-and-PEI format carries over, but QuantumBlack loops add coding assessments and technical interviews weighted toward machine learning, algorithms, Python, statistics, and probability, with generative-AI concepts increasingly tested in 2025-2026 interviews. Confirm which track your application or recruiter screen is routing you into before you prep - a pure case-interview prep plan will leave you underprepared for the technical rounds.
Common technical interview questions and answers
Section titled “Common technical interview questions and answers”Q: A client’s profits have fallen 15% year on year. How do you structure the case?
Start from the identity Profit = Revenue - Cost, and confirm with the interviewer whether the fall is a revenue problem, a cost problem, or both, before going deeper. Decompose revenue into price times volume, and volume into market size times market share, so you can separate a shrinking market from a share loss. Decompose cost into fixed and variable, then by major line item, and check whether cost per unit rose or whether volume fell so fixed cost is spread thinner. Then segment - by product, geography, channel and customer type - because an aggregate 15% decline is usually one segment collapsing rather than uniform decay. Close with the two or three drivers that explain most of the gap plus what you would do about each.
Q: Estimate the annual market for washing machines in India.
State assumptions out loud and work top-down. India has roughly 1.4 billion people at about 4.5 people per household, so around 300 million households. Take urban plus higher-income rural households as the addressable base, roughly 35 percent, so about 105 million households. Assume penetration of about 50 percent, giving roughly 50 million machines installed. With an average life of 8 years, replacement demand is about 6 million units a year; add first-time adoption of roughly 2 million a year for around 8 million units. At an average price of about 20,000 rupees that is roughly 16,000 crore rupees a year, or under 2 billion US dollars. The number matters far less than a clean structure, sane assumptions and a sanity check at the end.
Q: Explain the bias-variance trade-off.
Expected prediction error decomposes into bias squared, variance, and irreducible noise. Bias is the error from a model too simple to represent the true relationship - a linear fit on a curved signal underfits, giving high training and test error. Variance is sensitivity to the particular training sample - a deep unpruned decision tree fits noise, giving near-zero training error and much worse test error. Increasing model complexity lowers bias and raises variance, so the goal is the complexity that minimises total error, found with cross-validation. Practical levers are regularisation and pruning to cut variance, more features or a richer model to cut bias, and more data, which reduces variance without adding bias.
Q: When would you optimise for precision instead of recall?
Precision is true positives divided by all predicted positives - of the cases you flagged, how many were real. Recall is true positives divided by all actual positives - of the real cases, how many you caught. Optimise precision when acting on a false positive is expensive: flagging a legitimate high-value client for fraud review costs relationship damage and analyst hours. Optimise recall when missing a positive is expensive: screening for a disease, or detecting a safety defect, where a missed case is far costlier than an extra check. Because the two trade off through the decision threshold, report both, use F1 or an explicit cost-weighted metric, and read the precision-recall curve rather than accuracy when classes are imbalanced.
Q: What is overfitting, and how do L1 and L2 regularisation differ?
Overfitting is when a model learns noise specific to the training sample, so training error keeps falling while validation error rises. Regularisation adds a penalty on coefficient size to the loss. L2 (ridge) penalises the sum of squared coefficients, shrinking all of them smoothly towards zero without eliminating any, which handles correlated predictors well. L1 (lasso) penalises the sum of absolute values, and because its penalty has a corner at zero it drives some coefficients exactly to zero, performing feature selection - useful when you want a sparse, explainable model for a client. Elastic net combines both. Other defences are cross-validation, early stopping, dropout in neural networks, and simply collecting more data.
Q: A test is 99% sensitive and 95% specific for a condition that affects 1% of the population. Someone tests positive - what is the probability they have it?
Apply Bayes’ theorem. Take 10,000 people: 100 have the condition and 9,900 do not. Of the 100, sensitivity of 99 percent gives 99 true positives. Of the 9,900, specificity of 95 percent means a 5 percent false-positive rate, giving 495 false positives. So 594 people test positive and only 99 of them have the condition: 99 divided by 594, about 16.7 percent. The point is that with a low base rate most positives are false positives even with an accurate test - which is why McKinsey and QuantumBlack interviewers use this to check whether you reason from base rates rather than from the headline accuracy number.
Q: In Python and pandas, how would you compute each customer’s share of total revenue by region?
Group and then normalise within group. Aggregate first with df.groupby(['region','customer'])['revenue'].sum().reset_index(), then add a share column using a group transform: df['share'] = df['revenue'] / df.groupby('region')['revenue'].transform('sum'). transform is the key idea - it returns a series aligned to the original index, unlike agg which collapses rows, so you avoid a merge back. For the top customers per region use .sort_values('revenue', ascending=False).groupby('region').head(3). Groupby-transform is vectorised and runs in roughly linear time, whereas looping over regions with iterrows is orders of magnitude slower on real client data.
Q: What is retrieval-augmented generation, and why would a client use it instead of fine-tuning?
Retrieval-augmented generation puts a search step in front of a large language model: client documents are chunked and embedded into a vector store, the user’s question is embedded and used to retrieve the most similar chunks, and those chunks are injected into the prompt so the model answers from supplied evidence rather than from memory alone. Compared with fine-tuning it is far cheaper, updates the moment the underlying documents change, and lets you cite sources, which matters for auditability. Fine-tuning is the better choice when you need a consistent output format, a specialised tone, or lower latency per call, and the knowledge is stable. In practice the failure mode is usually retrieval quality, not the model - poor chunking or a weak embedding gives the model the wrong evidence and the answer is confidently wrong.
Frequently asked questions about McKinsey interviews
Section titled “Frequently asked questions about McKinsey interviews”What is the McKinsey interview process for freshers/campus hires?
McKinsey’s process typically runs application - McKinsey Solve (a gamified problem-solving assessment) - recruiter screen - first-round interviews - final-round interviews - offer. The full cycle usually takes several weeks, and well under 1% of applicants receive an offer, so treat every stage as a real filter.
What is the McKinsey Solve game?
McKinsey Solve (built on the former Imbellus assessment) is an online, gamified test with scenario-based modules - candidates typically complete two of three, such as an ecosystem-building or redrock-study scenario - in about 65 minutes. It scores your problem-solving process, not just the final answer, so how you explore and decide matters as much as the outcome.
What questions are asked in McKinsey interviews?
Each live interview pairs a case - an interviewer-presented business problem you structure and analyse out loud - with a Personal Experience Interview (PEI) question on leadership, entrepreneurial drive, or personal impact. Final-round cases tend to be more open-ended, with interviewers pushing harder on business judgment.
How should I prepare for McKinsey interviews?
Practise timed case structuring and mental math, try sample McKinsey Solve scenarios so the game format doesn’t cost you time, and prepare several PEI stories - each with a clear personal contribution and what you’d do differently - covering leadership, resilience, and impact.
What is McKinsey Digital / QuantumBlack, and is its interview different from a generalist Business Analyst hire?
McKinsey Digital and QuantumBlack (McKinsey’s AI/analytics arm) are the tracks most engineering, data-science, and CS-background campus candidates join rather than a generalist Business Analyst seat. The process still includes case and PEI rounds, but QuantumBlack loops add coding assessments and technical interviews weighted toward machine learning, algorithms, Python, statistics, and probability - generative-AI concepts have increasingly come up in 2025-2026 rounds - on top of the standard McKinsey case and PEI format.
What does McKinsey’s PEI (Personal Experience Interview) actually assess?
McKinsey updated its PEI dimensions in mid-2025 to four traits: Connection (influencing and building trust), Drive (resilience and initiative), Leadership (leading diverse teams), and Growth (learning and adapting). Each PEI story should map to one of these rather than being a generic ‘tell me about a challenge’ answer - interviewers probe for your specific contribution, not just the team outcome.