behavioral · standard
Tell me about a time you used data to influence a decision
Tell me about a time you used data to influence a decision.
Interviewers score this question on two axes at once: data fluency (did you understand what the data actually said?) and stakeholder influence (did you change someone’s mind, and how?). Most candidates pass one axis and fail the other. A strong answer shows both.
What the question is really asking
The word “influence” is load-bearing. The question is not “tell me about a time you looked at data.” It is asking whether you changed a decision that was already in motion, with someone else holding it, using data as the argument. That stakeholder navigation layer is what separates a reporter from an analyst.
The interviewers are listening for five things: the decision that was already being made (and why it was wrong), the specific data gap you identified, who needed convincing and why they were resistant, what the data actually showed, and the counterfactual (what would have shipped without you).
At Amazon, this question maps directly to two Leadership Principles: “Are Right, A Lot” and “Dive Deep.” Candidates who treat it as one LP signal leave points on the table.
What kills an answer
The single most common failure is conflating correlation with causation, or presenting a data output without describing the insight that changed the room. Saying “our metrics showed improvement, so we shipped” makes you a reporter. The interviewer learns nothing about how you think.
Weak answers also treat data as decoration. They describe a result (20% retention lift) with no mechanism connecting the data to the outcome. At companies running genuinely data-informed cultures like Google, Meta, Stripe, and Databricks, interviewers know the difference between statistically significant and practically significant results, and they notice when a candidate does not.
The STAR method is table stakes. What differentiates is the “Complication” layer: the moment where the obvious interpretation of the data was wrong or incomplete, and you saw through it.
weak
"We were seeing a drop in engagement and I pulled some analytics that showed users were churning after the onboarding step. I put together a presentation with the data and brought it to leadership. They agreed we should fix onboarding, so we redesigned it and retention went up 20% in three months."
This fails because the PM is a reporter, not an analyst. There is no tension (who disagreed? why?), no insight (what specifically in the data changed the interpretation?), no influence mechanics (why did leadership believe them?), and no counterfactual. The 20% result sounds fabricated because it lacks any mechanism. The interviewer learns nothing about how this person thinks.
strong
"Six months into running our search feature, the team was planning to invest in indexing speed. Leadership believed slow results were the retention problem. I wasn't convinced. I built a cohort analysis segmenting users by first-search-result quality score (did they click result 1, or did they scroll and bounce?) rather than by query latency. What it showed: users with sub-200ms results but low result relevance churned at the same rate as users with 600ms results. Users with slower but more relevant results retained at 15 points higher. The data reframed the problem from infrastructure to ranking.
That was a significant budget reallocation. The infrastructure team had a strong sponsor. I brought in two supporting signals: session replay clips showing the click-then-bounce pattern, and a competitor teardown showing their result quality scores. Stakeholders could see the behavior, not just the chart. We shifted 60% of the planned infrastructure budget to ranking improvements. Within two quarters, 7-day retention for new users went from 34% to 41%. The latency investment went on the roadmap for the following quarter, informed by actual bottleneck data, not assumption."
This works because it names the existing decision and who held it, shows a PM who defined the measurement frame (result quality score, not just clicks), describes the specific insight that changed the argument, names the resistance and the multi-signal influence strategy, gives a credible counterfactual, and reports a result with a timeline that makes it checkable.
Picking the right story
Senior and staff PM candidates are expected to show that they defined the measurement framework, not just consumed existing dashboards. “I showed a dashboard” is weak. “I reframed what the metric meant and changed the decision criterion” is strong.
Calibrate story scope to the role. An IC PM should show cross-functional influence (convincing engineering or design). A senior PM should show influencing a director-level stakeholder or reallocating budget. A staff PM should show changing a strategic direction, with the data architecture they built to make the argument legible.
The 2026 AI PM layer
At AI-native companies (Anthropic, OpenAI, Cursor, Perplexity, Glean), “data” now includes evals, quality scores, hallucination rates, cost-per-query at scale, and latency-quality tradeoff curves. Candidates who anchor exclusively on A/B test examples are answering the 2022 version of this question.
A strong 2026 answer might sound like: “I ran a structured eval on 200 adversarial prompts and showed that our 90th-percentile failure mode was a billing context gap, not a model capability gap. I used that to redirect a fine-tuning budget toward RAG retrieval.” That demonstrates data fluency at the frontier of what the role now requires: knowing which data source is authoritative for a given decision type, what a valid eval looks like, and how to distinguish a model problem from a data problem from a product problem.
Data fluency in 2026 is not just product analytics literacy. It includes knowing what your evals measure and what they miss, and being honest about both when you’re in the room.