🌏 中文版
Today's Topic
Since mid-2026, Meta has added a new round to its PM loop: "product sense with AI," where candidates use an AI tool live to generate a solution to a product sense prompt. The thing being graded isn't how well-crafted your prompt was — it's whether you can spot what's wrong with the AI's first draft and fix it. That skill has a name: Human Delta, the specific insight you add on top of what the model produced. Today works through a real Meta interview prompt leaked from this round, then closes with a case from a Wharton study published last month — not a hypothetical, a real company that paid weeks of cleanup cost for skipping exactly this check.
Core Framework Cheat Sheet
The Human Delta stack (what Meta's product sense with AI round actually grades)
The interviewer isn't reading your prompt — they're watching what you do with the AI's first output. These four layers go from table-stakes to the ones that actually separate strong candidates, in that order:
| Layer | What the interviewer is watching for | Common way to lose points |
|---|---|---|
| Human Delta | Can you catch hallucinations, generic answers, or outputs that drift from the product vision, and explain exactly how to fix them | Reading the AI's draft out loud as if it were your own answer, with no judgment added |
| Strategic Leverage | Did you offload commodity work — segmentation, feature drafts, metric lists — to the AI, so you have time left for the real trade-off | Doing all the basic groundwork by hand, leaving no time to discuss the actual hard decision |
| Product Direction | Can you turn a vague prompt into a specific instruction for the AI, and push back with added constraints when the output isn't deep enough | Feeding the AI a prompt as vague as the question itself, then getting a vague answer back |
| Audit and Redirect | Can you spot feasibility, policy, privacy, or cross-functional friction the AI missed, and redirect the solution | Only checking whether the plan sounds reasonable, not whether it's actually buildable |
These four layers aren't a one-time checklist — they're a loop that keeps running: the AI drafts, a human audits for gaps, a failed audit sends it back with tighter instructions, and only a passed audit gets signed off:
- AI generates a first draft from the PM's instructions
- PM audits for hallucinations, generic output, drift from product direction
- Decide whether this draft passes the audit
- Fails — add constraints and send back to the AI to redraft
- Passes — sign off, with final judgment call staying with the PM
Simplified view mapping to the Human Delta (catch the gap) and Audit and Redirect (catch feasibility/policy risk) layers.
Today's Practice Question
The Prompt
"You're the sole PM for Zoom. Imagine that AI is sending transcripts of meeting notes to all invitees, and suddenly, meeting attendance is down. What do you do?" (Source: a real Meta product sense with AI interview prompt, compiled by IGotAnOffer from 2026 candidate reports)
How to Break It Down
- Clarify the problem: Ask whether "attendance down" means invitees declining outright, or accepting and then leaving early or going camera-off. Check whether the drop started when the transcript feature launched, or was already trending that way. You need the real timeline before you can even tell whether the transcript feature is the actual cause.
- Don't accept the frame that this is automatically bad: This is today's Human Delta move. The AI handed you a correlation — attendance is down — not a verdict. Split meetings into two buckets: low-value meetings where a transcript genuinely substitutes for showing up (people skipping these is a good outcome), and high-value meetings that need real-time interaction but are getting skipped anyway because the transcript feels "good enough" (this is the actual problem).
- Structure the analysis: Use Audit and Redirect on the feature itself — is the transcript going to every invitee or only confirmed attendees? Is it sent after the meeting, or does its mere existence get advertised beforehand in a way that discourages showing up? Also check whether transcripts carry sensitive content that optional or external invitees shouldn't be seeing — a privacy risk the AI won't flag on its own.
- Propose a solution: Short term, tier the transcript delivery — full transcripts go only to confirmed attendees or people the host designates, everyone else gets a summary, so showing up is once again the reason you get the full picture. Long term, flip the signal and use it on purpose: flag meeting types with low attendance but high transcript-download rates, and push hosts toward async formats, turning the time saved into an actual team win.
- Define success: Don't use overall attendance as the North Star — use attendance for high-engagement meetings, since filtering out low-value ones is exactly the direction the signal should move. Track tiered-delivery attendance as the input metric, and track "sensitive transcript sent to a non-attendee" incidents as the guardrail — that number needs to hit zero.
Sample Answer (how you'd actually say this in an interview)
Question the AI's conclusion before accepting it: "I'd start by pinning down what 'attendance down' actually means — declines versus no-shows — and whether the timing lines up with when the transcript feature shipped. But even if the timing matches, I wouldn't treat that as bad news right away, because a drop in attendance can mean two very different things: people realizing some meetings didn't need to be live because the transcript covers it, which is actually the product saving the company time, or people skipping meetings that genuinely needed real discussion. That second case is the one I'd actually need to fix."
Problem and solution: "Right now every invitee gets the full transcript regardless of attendance. I'd tier that — confirmed attendees or host-designated people get the full transcript, everyone else gets a summary, so showing up is again the reason you get the complete picture. I'd also check whether transcripts are going to external or optional invitees who shouldn't see sensitive content — that's a risk the AI generating this feature wouldn't flag on its own, so it's on me to audit it."
Close on metrics: "My North Star would be attendance for high-engagement meetings, not overall attendance, because filtering out low-value meetings is exactly what this feature should be doing. I'd add a guardrail metric tracking whether transcripts are leaking to people who shouldn't see them. If high-engagement attendance is also dropping, that's when I'd treat it as a real problem and go back and check whether the tiered delivery is actually working."
Self-Check List
Use this table to check whether your answer covered the key points:
| Checklist item | Covered? |
|---|---|
| Did you question whether "attendance down" is correlation, not a verdict handed to you by the AI | |
| Did you split meetings into "being filtered out is good" versus "genuinely needed but skipped anyway" | |
| Did you catch the transcript's own policy/privacy risk (something the AI won't flag on its own) | |
| Did your solution split into short-term (tiered delivery) and long-term (use the signal on purpose) | |
| Did you name both a North Star and a guardrail metric | |
| Bonus: did you argue that the attendance drop could be proof the product is working, not a bug |
Today's Case
The cost of skipping Human Delta: a consumer electronics company's flagship device got auto-listed at 85% off for three hours
A report published in mid-October 2026 by the Wharton School and consulting firm GBK Collective found almost no consensus among executives on who should be blamed when an AI agent "makes a consequential error" — roughly a quarter pointed to the employee operating it, a quarter to their manager, a quarter to the company's risk team, and the rest split between the vendor or simply not knowing. Jeremy Korst, a partner at GBK Collective and co-author of the study, said the most common answer he gets from senior executives when he asks this question is that they hadn't thought about it yet. The report cites a real case from Leeron Walter, VP at employee-monitoring software company Teramind: at a US consumer electronics company, a marketing employee used AI to auto-generate and publish product prices. A single prompt error caused a flagship device to be listed at an 85% discount for three hours before anyone caught it — more than 700 orders went through in the meantime, and the customer-service cleanup took weeks. A second case in the same report: a US retailer's AI-automated customer email system sent incorrect refund and cancellation notices to more than 18,000 customers in under six hours because of one prompt error.
How to use this in an interview: These two cases are the real-world version of what happens when Human Delta gets skipped. Auto-generating prices and auto-replying to customers both look like the right call on Strategic Leverage — offloading commodity work to AI — but without a human audit before anything goes live, the risk turns straight into real financial loss. If you get asked something like "would you let AI publish automatically," these are the counter-examples to reach for: any AI output that touches pricing, customer notifications, or other irreversible outward-facing actions needs a human checkpoint in the loop, not full automation.
Further Reading
- Meta Product Sense Interview (questions, framework, prep) — the full explanation of the Human Delta stack and the real question bank from Meta's product sense with AI round.
- Product Manager Interview Questions 2026: 60 Questions That Reveal Real Skill — how AI product literacy gets folded into 2026 interview scoring, and the specific follow-ups around eval harnesses and fallback paths.
- Product Manager Interview Practice & Mock — more product design, metrics, and strategy prompts with answer outlines to keep practicing with.
References
- Meta Product Sense Interview (questions, framework, prep) — source for the Human Delta, Strategic Leverage, Product Direction, and Audit and Redirect framework, and the Zoom AI transcript practice question.
- Who's to blame if AI fails at work? Wharton study shows a split — original source for the consumer electronics 85%-discount incident and the retailer's 18,000 incorrect notices (Bloomberg, citing the Wharton/GBK Collective October 2026 report).
- Product Manager Interview Questions 2026: 60 Questions That Reveal Real Skill — source for the AI product literacy interview-scoring notes.