A practice prompt we wrote. No company or candidate report names it, so it carries no company tag.
How to answer
OpenAI’s San Francisco posting, as of September 2026, writes measurement into the job: it says FDEs measure success through production adoption, measurable workflow impact, and eval-driven feedback that changes product and model roadmaps. Source 1Forward Deployed Engineer (FDE) - SFPublisherOpenAI CareersSource typecompany job posting This question checks whether you have done the measuring yourself, not quoted someone else’s dashboard.
- The goal, in the customer’s words, and who owned it. “The head of customer service was losing accounts because orders shipped a day late.” The metric follows from this.
- The metric, and why that one. Fixed before you built anything, with the goal’s owner. Name what you rejected: an activity count you could raise without helping anyone, or a model score that matters but isn’t the goal. Add the guardrail you watched so the win didn’t hide a cost, an error rate or a complaint count.
- The baseline. How you got it before anything changed: the logs you pulled, the sample you timed by hand, over what period. Say you pulled it.
- The comparison. A staged rollout with a group left on the old way, or at least last year’s same weeks. Name what else changed in that period.
- What it showed, read plainly. The number you claim, the part that didn’t move, the guardrail’s result, who on their side checked your number, and what they did with it. If your metric stands in for the real goal, say what you could not yet show about the goal itself.
Step five is where senior answers pull ahead. Say what you would not claim: “Our comparison sites improved too, so I credit us with the difference, not the whole gain.” That sentence makes the rest believable.
The traps are a metric chosen after the fact because it went up, and an activity count presented as impact.
Follow-ups
What the interviewer may ask next, once your first answer is on the table.
- What else changed during that period that could explain the result?
- Which metric did you consider and reject, and why?
- Did the customer accept your number? Who checked it?
Where answers go wrong
- Reports an activity count, such as documents processed or users onboarded, as the impact, with no baseline and no tie to what the customer wanted.
- Presents a clean before-and-after win and has no answer when asked what else changed at the same time.
Answer this in two minutes
Write the answer you would say out loud. The clock starts with your first word.
Illustrative answer about a fictional project
I built order entry from emailed purchase orders for an industrial parts distributor with four branches. Their customer service reps were retyping PDF orders into the ERP by hand.