Tell me about a time you embedded with a customer team and shipped something to production inside their systems.
Answer it out loud first, then compare. That's what makes it stick.
Situation: a logistics customer needed an AI tool to classify inbound shipping emails, running inside their locked-down network. Task: get it to production in six weeks. Action: I spent week one with their dispatchers, got read-only access to a sample of 2,000 emails, shipped a narrow version for one depot, and logged every miss. Result: it handled 71 percent of emails unaided, their own engineers now run it, and I left a runbook and alerting behind.
- Situation: their stack, constraints and who controlled access
- Task: the deadline and the definition of done
- Action: how you earned access and trust, what you built
- What broke on the way to production and how you fixed it
- Result with a number, plus the handover so they could run it without you
This is the closest thing to the core FDE job description. Anthropic's public postings ask FDEs to work within customer systems to build production applications, and prep guides built from those postings say interviewers probe whether you can deliver in an environment you do not control. The two most common failures are telling a demo as if it were production, and telling a story where the customer's engineers did the building.
Structure the story so each beat proves something
- Constraints: name them plainly. Network allow-lists, no internet from the data zone, a security review that takes three weeks, a legacy system with no API. These make the story credible.
- Trust: say how you earned it. Sitting with end users on day one, shipping a tiny win in the first week, writing down every decision and risk in a shared page.
- Build: describe a narrow first release (one depot, one email type), not the whole vision. Narrow scope is what makes production in six weeks possible.
- What broke: pick one real failure. For example the model mis-routed forwarded emails with long quoted threads, found through a miss log; you fixed it by stripping quoted text and adding a 60-case regression set.
- Production reality: auth via the customer's identity provider, secrets in their vault, logging and alerts in their tooling, rollback plan, an owner named on their side.
- Handover: runbook, dashboards, a training session, and a clear list of what they should and should not touch.
Numbers worth stating
- Time to first production user (weeks)
- Share of cases handled without a human (71 percent in the example)
- Hours saved per week, or cost per item
- How many incidents in the first month and how long each took to resolve
Expect this follow-up: what did you hand over so their team could operate it without you? Have a concrete list: runbook, eval set, alert thresholds, a named owner and a 30-day check-in.
Be careful not to leak customer names or confidential figures; anonymise to 'a mid-size logistics company'. Say 'I' for what you did and 'we' only for what was truly shared.
- What did you hand over so their team could run it without you?
- How did you get access to their data and systems, and who said no?
- What was the first thing that broke in production?
- How did you earn the customer team's trust in the first week?
- Anthropic Forward Deployed Engineer Interview (2026) (prep, 2026-10-25) Lists an embed-and-ship-to-production behavioral question tied to the posting's responsibility to build within customer systems.
- Scale AI Forward Deployed Engineer (FDE) Interview Guide - Aced (prep, 2026-08-18) Reports that the hiring-manager round at that company checks fit for embedded customer work.