career · career
Build your AI graveyard: how killed projects strengthen a PM portfolio
Interviewers at AI-first companies are not impressed by a portfolio that only ships. They want the graveyard: AI features killed on purpose, calls made before sunk cost took over, redirected budget that produced something real. In 2026, that is the rarer proof of judgment.
Why kills signal more than ships
Feasibility is nearly free. Any PM with a cloud API key and a Cursor subscription can ship an AI feature in days. That collapses “I shipped X” from differentiator to baseline. What you cannot script is the judgment to stop.
KORE1’s 2026 AI PM hiring guide says it directly: “They have never killed an AI feature. Nobody who has shipped real AI has a perfect record. The graveyard is part of the resume.” Interviewers probing for difficulty are not hunting for recovery narratives. They want evidence you can read signal fast, act on it before cost compounds, and redirect the team.
The viable/lovable lens makes this concrete. A killed AI project is usually a viability failure (not worth solving at this cost, for this market) or a lovability failure (technically works, users don’t want it this way). Neither is a story about incompetence. Both are stories about knowing what matters, which is the judgment AI-first companies pay $305K-plus nationally to find.
Three types of kill: which belong in your portfolio
Principled kill. You had evidence, acted on it, and stopped something with momentum. DAY-7 retention for the AI-suggested workflow was 4% against 31% for the manual path. Compute cost per session was 14x what the premium tier could bear. You named the signal, got alignment against pushback, and stopped. This belongs in your portfolio.
Pivot kill. The team redirected before full build-out because discovery surfaced a narrower problem. Worth including if you can show your specific role in surfacing the signal.
Org kill. Leadership pulled the plug, layoffs ended the team, a reorg changed priorities. Least useful unless you were actively advocating for the stop before it happened. Including a kill you did not influence reads as borrowed credit.
The portfolio artifact: a graveyard entry
A graveyard entry is a brief structured artifact, not an apology. Four parts:
Signal. What evidence told you this was wrong. Name the metric, the eval result, the cost number. “Users weren’t engaging” fails. “DAY-7 retention was 4% against 31% for the manual path; the AI added steps without reducing uncertainty” passes.
Cost of stopping. What was sunk, who resisted, what alignment required. A kill that nobody cared about is not a judgment story.
What the kill unlocked. Where the engineering capacity went next and what it produced. GetProductPeople’s 2026 recruiter guide frames the three kill-decision questions as ego impact, financial impact, and opportunity cost. The opportunity cost side is the portfolio proof: a kill that redirected a team to something that drove a real metric is worth more than a shipped feature with modest impact.
What you would catch earlier. One sentence on what the discovery process missed. Signals you treat failure as data, not something to minimize.
Answering “tell me about a project you killed”
Anthropic, OpenAI, and Google DeepMind include this as a standard behavioral screen. The failure-to-reflect answer (“nothing major, it all went well”) is a red flag.
strong
"Six weeks into an AI-suggested reordering feature, evals looked clean in staging but production told a different story: DAY-14 adoption was 9%, and shadowed sessions with operators showed the model had no visibility into supplier lead times. Its suggestions were technically plausible and contextually useless. I wrote the kill case with the fix cost ($240K minimum), the opportunity cost, and what operators actually needed: a simple anomalous-gap alert with no model required. The VP pushed back. I showed her the retention number and asked what her threshold was. We killed it that week. The alert shipped in three weeks and drove a 19-point lift in weekly active use. What I'd catch earlier: lead time data availability should have been a discovery gate, not a post-launch finding."
weak
"We tried an AI reordering feature but it wasn't working the way users expected, so we pivoted to something simpler that ended up being more successful. We learned a lot about what users actually needed." This centers the ship, not the kill. "Wasn't working the way users expected" is not a signal. 9% DAY-14 adoption is. Interviewers at AI-first companies have heard hundreds of pivot narratives. They're looking for the PM who can name the metric, name the stakeholder who resisted, and name what the stop cost them personally to push for.
A portfolio with only ships shows execution. A portfolio with principled kills shows product judgment. For the specific behavioral question structure, see killed a project you loved. For the upstream call about when to stop an AI idea before it becomes a project, see kill the AI idea.