In this article
Before you start
For the wider method, see behavioral interview questions; this page builds on the general method for answering interview questions.
Why interviewers ask this
A weak answer to this question describes an engineer who never pushes back, or one who pushes back and never lets go. Interviewers use the question to see which of those two behaviors you show when someone who depends on your model wants something you doubt.
The format is behavioral. Northwestern Career Advancement (accessed 2026-10-26) says "Most employers use behavioral interviewing, which is based on the idea that past behavior predicts future performance." The best-known public wording of the underlying value is Amazon's Leadership Principles (accessed 2026-10-26): "Leaders are obligated to respectfully challenge decisions when they disagree, even when doing so is uncomfortable or exhausting." The same page adds "Once a decision is determined, they commit wholly." The Amazon Leadership Principles page covers the full list.
Candidates report the question in loops at large tech companies; no public source says how often. A strong answer shows both halves, challenge and commitment. A Staff-level answer adds how the evidence changed the decision, or how you protected the project after you lost.
Disagreeing with a stakeholder from first principles
- Evidence-first pushback
- Evidence-first pushback is a way of disagreeing in which you restate the other person's goal, replace your opinion with a measurement, baseline or small experiment that could settle the question, and propose it as the next step. The argument becomes about a number both sides can read.
In plain words
Think of two people planning a drive. The navigator believes the left road is faster and says so, with the map open. The driver picks the right road. The navigator then helps drive the right road and does not sulk or quietly slow the car.
The analogy breaks in one place. In ML work the map is a metric you can measure, so you can often settle the argument before anyone chooses a road.
In one precise sentence, disagree and commit is a rule that you state your objection openly and with evidence while the decision is open, and then carry out the decision made as if it were your own.
Step 1: Choose a disagreement the evidence could settle
The answer starts by naming the decision and one measurable claim. Pick a story where the disagreement was about something checkable: a launch date, a model against a rule, a metric, a data source.
Input: a real moment where you doubted a stakeholder's plan. Output: one sentence such as "She wanted a model shipped; I doubted it beat a simple rule." Be careful with metrics. A metric is a number that you care about, and an objective is a metric that your algorithm is trying to optimize, as Google's Rules of Machine Learning (accessed 2026-10-26) defines them. A stakeholder's key performance indicator (KPI) may not be the objective the model trains on, and many disagreements hide in that gap. You now have the claim; the next step makes it testable.
Step 2: Turn the opinion into a measurement
Evidence replaces opinion before anyone is overruled. Google's guide gives the general rule as "measure first, optimize second". For a stakeholder who wants ML, Rule 1 supplies a fair challenge: "If machine learning is not absolutely required for your product, don't use it until you have data."
Propose a heuristic baseline and a short test on the same metric, framed as a shared goal: "I want the feature to work; can we check what the model adds?" A design doc is a short written plan of what a team will build and why. Google says it "functions as a way to receive feedback and address questions and concerns" from stakeholders (Google, Stakeholders, accessed 2026-10-26).
In an illustrative test, the current ordering gets a 4.0% click-through rate, a recency rule gets 5.0% and the model gets 5.6%.1 The rule delivers 1.0 of the 1.6 points of lift, or 62.5%, and the model adds 0.6 points, a 12% relative gain over the rule.1 The decision is now about whether 0.6 points is worth the cost.
Step 3: Commit wholly, and say what you did
After the decision is made the answer shows full commitment. Bezos describes the phrase in his 2016 shareholder letter (April 2017): "Third, use the phrase “disagree and commit.” This phrase will save a lot of time." He frames the commitment as sincere, a candid statement of your view followed by a quick commitment to go the team's way, not private doubt while you go through the motions.
Commit does not mean give in, stay quiet or split the difference. Amazon's page says leaders "do not compromise for the sake of social cohesion." In the story, say the concrete things you did afterward: you recorded your concern in the design doc, built the monitoring, set a rollback threshold, reported results. For reversible decisions, the same letter says to decide with "somewhere around 70% of the information you wish you had", and it pairs that with recognizing and correcting a bad decision quickly. Waiting for 90% is, in his words, probably being slow.
Step 4: State the result and what changed
Close with the result and one change in how you work. In STAR, the result is the final part, and it must be stated. There are three honest endings: you were right and the decision changed; you were wrong and say what you learned; or the outcome was mixed and you say what the evidence showed.
Do not invent a number. If you have a real one, use it with its unit. If you have none, say what you observed, such as "the model held its metric and the rollback never triggered." End on a habit, such as asking for a baseline at kickoff. You now have all four parts: the decision, the evidence, the commitment and the result.
Step-by-step description
- Name the decision and one claim.
- Replace opinion with a measurement.
- Execute the decision fully.
- State the outcome and the change.
- The answer starts by naming the decision and a measurable claim.
- Evidence replaces opinion before anyone is overruled.
- After the decision is made the answer shows full commitment.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
The story below is an illustrative composite, not a real case, and it names no company or person.
Tell me about a time you disagreed with a stakeholder on a machine learning project.
A product manager wanted to ship a learned ranking model for notifications this quarter. I disagreed because we had no baseline, so nobody could say what the model added. I told her that directly, restated her goal of more engagement, and proposed a two-week test of a simple recency rule against the model on the same metric. The rule captured most of the lift, and I showed her the numbers. She still chose the model because it opened up personalization later. I recorded my concern in the design doc, then committed fully: I built the monitoring, set a rollback threshold and reported weekly. The model worked, and I now ask for a baseline at kickoff.
What makes it Staff-level: it replaces opinion with a test, records dissent in writing, commits through concrete actions and ends on a lasting habit. The spoken structure: name the decision, show the evidence, say what you did after losing, state the result.
On timing, Northwestern suggests 15%, 10%, 50% and 25% for situation, task, action and result, and says action "should always be the longest part of your response". For a 90-second clip that is 13.5, 9, 45 and 22.5 seconds, and for 60 seconds it is 9, 6, 30 and 15.1 Action takes 45 of 90 seconds, the longest part of the answer. Situation and task together take 22.5 seconds, about a quarter of the answer.
Text description of the diagram
- Action takes 45 of 90 seconds, the longest part of the answer.
- Situation and task together take 22.5 seconds, about a quarter of the answer.
The rubric
This rubric is advice on what a strong answer shows, not a published company scorecard.
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Evidence | "I just felt the model was wrong." | Cites a metric or a comparison. | Designs a quick test that could prove you wrong. |
| Respect | Blames the stakeholder. | Restates their goal first. | Frames the test as a shared goal. |
| Commitment | Complies silently or keeps arguing. | States the objection once, then executes. | Records dissent in writing, then builds the safeguards. |
| Result | No outcome given. | States what happened. | Gives the outcome, including a loss, honestly. |
| Learning | None. | Names one lesson. | Shows a changed habit, such as asking for a baseline. |
Three variants of this question
What if you disagreed with your manager about launching a model?
The structure is the same. Raise the objection once, in private or in the design doc, with the evidence. If the manager still decides to launch, say you committed and name the safeguards you added.
What if the stakeholder turned out to be right?
Say so plainly and make it the lesson. A strong answer shows you could be persuaded by the evidence and changed your approach, which is the same skill as pushing back.
What if you could not run an experiment?
Use the best evidence you had: historical data, a small held-out slice of past traffic, or published guidance. For reversible decisions, Bezos's letter argues for deciding with about 70% of the information, so name what you did not know and how you would catch a wrong call early.
Common mistakes
- "I told them they were wrong, and I was." (wrong) It makes the story about winning. (right) "I raised the concern with evidence, and then I executed the decision."
- "I disagreed in the meeting, then went along." (wrong) It leaves out the evidence and the actions. (right) "I proposed a test, recorded my view in the design doc, and built the monitoring."
- "We met in the middle." (wrong) Splitting the difference to keep the peace is what the principle warns against. (right) "We chose one option and I committed to it."
- "I agreed on the surface but kept my doubts to myself." (wrong) Bezos describes commitment as sincere. (right) "I voiced my doubts openly, then gave the decision full effort."
Follow-up questions
How do you disagree with a stakeholder without sounding negative?
Restate their goal first, then offer a measurement as the next step instead of a verdict. The interpersonal communication page covers the wording in more depth.
What is the difference between disagree and commit and compromise?
Compromise splits the difference to keep the peace. Disagree and commit picks one option after a candid argument, and everyone then executes it fully.
How should you split the time in a STAR answer?
Northwestern suggests 15%, 10%, 50% and 25% for situation, task, action and result, so most of your time goes to what you did. The common interview questions guide shows the same structure on other prompts.
What if you have no real disagreement story?
Use a smaller real one, such as a metric choice or a data-quality concern, and do not invent a dramatic case. Interviewers ask for past behavior, so a modest true story beats a larger made-up one.
Is this the same as the Amazon Have Backbone question?
It tests the same two halves, challenge and commitment, which the Amazon Leadership Principles page explains. Pair it with a strong answer to why we should hire you so your stories show the same strengths.
- Challenge with evidence and respect while the decision is open; commit wholly once it is made.
- In ML, turn the disagreement into a metric, a heuristic baseline or a small test.
- Commitment is shown by concrete actions, such as monitoring, a rollback threshold and written dissent.
- Spend about half of the answer on action, and end with a result and one changed habit.
Sources
- Amazon Leadership Principles — Amazon, 26 Oct 2026
- Jeff Bezos' 2016 Letter to Shareholders — Amazon, 12 Apr 2017
- Rules of Machine Learning: Best Practices for ML Engineering — Google for Developers (Zinkevich), 26 Oct 2026
- Stakeholders, Managing ML Projects — Google for Developers, 26 Oct 2026
- Behavioral Interviewing and The STAR Approach — Northwestern Career Advancement, 26 Oct 2026
