How to hire an AI-native product manager

Adam Faik

Jul 20, 2026

The headcount finally came through. You need a product manager, your recruiting partner needs a job description by Friday, and the process sitting in your ATS (the software your whole recruiting pipeline runs on) is the one you’ve run since 2023. Résumé screen, recruiter call, case-study interview, take-home, panel, offer. Every stage works the same way: the candidate produces something, and you grade it. A résumé. A structured answer. A polished deck. Look at it honestly and your hiring process is a machine for grading documents.

Here’s the problem: AI just made polished documents free. The tight résumé, the crisp case structure, the take-home that used to signal “this person is serious”: any candidate with Claude and an evening now produces all three. You’ve probably felt it already, reading applications that all sound suspiciously excellent. And it happened at the worst possible moment, right when you were asked to make your team AI-first and every open seat became a chance to hire the mindset. The signal collapsed at the exact moment the stakes went up.

So I went and read what the companies furthest into this shift actually did about it. I pulled the live PM job posts at Anthropic, Ramp, Notion, and Stripe, straight from their own job boards this week. I read the interview loops Sierra and Canva published, and the one candidates describe at Shopify. I read the fluency rubric Zapier now applies to every single hire. Different companies, different stages, one converging move. Stop grading the deliverable. Watch the candidate work. The bar isn’t “knows AI.” It’s “directs AI and catches it when it’s wrong.”

Here’s the whole arc in one picture.

The diagnosis. I’ll show you why every stage of your current process stopped measuring anything real.

The rebuild. Five moves, one per stage, with full requirements quoted from live postings and the interview questions from companies that already made them.

The calibration. You’ll see the data on how much AI to actually demand, and the walk-back that warns against overshooting.

The walkaway. One forwardable paragraph for leadership and your recruiter, plus a checklist to run on the req you have open.

By the end you’ll have a full replacement for the old process: job-description language you can borrow whole, two screening questions, a live interview format with a published AI policy, a work-sample design, and a decision rubric, all compressed into a closing checklist you can run this week. Your next PM hire stops being a coin flip on polish and becomes the clearest AI-first signal you send your team this year.

Let’s rebuild it, stage by stage.

Why your hiring loop went blind

Walk through the classic stages and name what each one actually measures. The résumé screen grades a document. The case interview grades a rehearsed structure, the kind a decade of candidates drilled from Cracking the PM Interview, McDowell and Bavaro’s 2013 prep book (”How many pizzas are delivered in Manhattan?”, “How do you design an alarm clock for the blind?”). The take-home grades another document, produced somewhere you can’t see. The panel grades a performance, and every stage grades some kind of output.

For years we told ourselves polish correlated with skill, that producing a crisp PRD required a crisp mind. The honest version is less flattering. Selection studies have long ranked résumé screens and unstructured interviews among the weakest predictors of actual job performance. The old process wasn’t a good instrument that AI broke. It was a weak instrument whose weakness AI made undeniable.

Picture the machine you’ve been running.

There’s a famous precedent for this kind of audit. Back in 2013, Google ran the numbers on its own famously clever interviews. Laszlo Bock, who ran People Operations there, went public with the verdict: brainteasers were “a complete waste of time” that “serve primarily to make the interviewer feel smart.” Google moved to structured interviews as a result. The lesson stings a little: a signal-free process can feel rigorous for years, because nobody checks. AI didn’t create the weakness. It removed the excuse for ignoring it.

You can watch the collapse happen in real interviews. In Nikhyl Singhal’s The Skip (Singhal ran product at Meta and Google), hiring manager Sam Stone observes that candidates who used AI to prep their case structure “will abandon the structure the moment Q&A starts, because they never internalized it.” In the same piece, product leader Mckenzie Lock shares the probe she used when screening at Netflix: when new information invalidates an assumption, “can you say ‘oh, that changes things’ and walk through updated logic?” And Meta gives the same diagnosis from the company side. Its new interview format puts an AI assistant inside the session, partly because, in Meta’s own words, it “makes LLM-based cheating less effective.” The companies that moved first aren’t fighting AI-polished candidates with detection tools. They moved the test to where polish can’t fake it.

Meanwhile the requirement itself went up, not down. In Microsoft and LinkedIn’s Work Trend Index, a survey of 31,000 people, 66% of leaders said they wouldn’t hire someone without AI skills. That survey is from May 2024, and nothing published since suggests the number softened. So you’re expected to hire for a skill your instruments can’t see, using stages that stopped measuring anything. The fix isn’t a better lie detector. It’s a process that watches work instead of grading output. Five stages, five moves.

Rebuild the loop stage by stage

Everything below comes from choices real companies have published, plus my own read on how to adapt them to a PM seat. Treat the five moves as parts to adapt, not a recipe to obey. Your org, your seat, your constraints. Run the ones that fit.

Here’s the full map before we go stage by stage.

Write AI into the job description

Stop writing “familiarity with AI tools is a plus.” The most advanced companies on this front put AI usage in the requirements section and name tools, the way postings used to name SQL. The strongest reqs treat AI usage as a requirement with teeth, not a vibe. All of these lines come from live postings I pulled off the companies’ own job boards this week, quoted in full so you can borrow whole requirements, not fragments:

Notion (PM, Workspaces, an everyday workspace-product seat): “AI-pilled: You have comfort and an ongoing curiosity with using AI tools for your product development process.” And on every one of its 141 live postings, Notion adds: “You don’t need deep AI expertise for every role, but we do expect every Notino to be intellectually curious, drawn to tinkering and discovery, and excited to use AI as a real collaborator in their work.”

Supabase (PM, Interfaces; the line recurs across their PM postings): “You use AI to move faster. You use AI tools to compress the slow parts of the job (research, drafting, synthesizing feedback) so you spend more time on judgment calls, and you’ve built enough with agents to have real opinions on what makes an interface agent-friendly.”

Airbnb (PM, Incubations, a consumer seat asking for 10+ years of experience): “Demonstrated curiosity and point of view on incorporating AI tools into the product workflow.”

Canva (PM for Pexels, their stock-photo product): “We’re looking for an AI-embracing entrepreneurial type product manager.” And in the experience list: “You embrace AI for automation, prototyping, report generation, data analysis, etc.”

Stripe (growth PM, in Minimum Requirements): “Experience building AI-powered, conversational, or agentic products. You understand what it takes to ship AI that works reliably for real users, not just in demos.” And under preferred qualifications: “You have shipped products on top of these technologies, not just evaluated them.”

Ramp (PM, Agentic CX): “We’re looking for builders who can vibe-code a prototype before lunch and write the rollout plan after.” The requirements spell it out: “Deep hands-on experience building with AI: you’ve prototyped, shipped, and iterated on AI tools, agents, or workflows, not just managed roadmaps about them. Technical fluency with modern AI coding harnesses (Cursor, Claude Code, Codex).” And: “Fluency in data and AI evals: you know when to trust numbers and when to lean on principles.”

Anthropic (research PM, in Minimum Qualifications, in full): “Have a deep passion and curiosity for AI and LLMs. Use AI regularly.”

One distinction before you borrow. At Anthropic or on Ramp’s agentic products, asking for AI depth is as natural as a payments seat asking for payments experience: the product is AI, so the requirement writes itself. The lines worth copying are the other kind: a workspace product, a home-rental team, and a stock-photo library writing AI usage into ordinary PM seats. That’s the expectation spreading beyond AI products, and it’s the part that applies to your posting whatever you build.

Now decide which ask you’re actually making. The postings above split into three distinct tiers, and mixing them up is how job descriptions turn into wish lists.

Tier one asks for a PM who works faster with AI: the Notion, Supabase, and Airbnb language. Tier two asks for a PM who prototypes with AI coding tools by name: the Ramp language, echoed in Anthropic’s Labs PM posting (”Prototype with AI tools like Claude Code”). Tier three asks for a PM who owns evals, the tests that measure whether an AI feature actually works: Anthropic’s Claude Code PM posting wants someone who has “personally built agentic evals,” and Figma and Databricks write eval ownership straight into PM responsibilities. Most seats need tier one or two, and writing all three into one req doesn’t get you a unicorn, it gets you an empty pipeline. Pick the tier, write two or three specific bullets, and move on. The req is now making a promise the rest of your process has to keep.

Screen for their last real AI workflow

The old screening question (”are you familiar with AI tools?”) is answerable by anyone with a pulse and a ChatGPT subscription. Replace it with two questions that can’t be prepped.

First: “Walk me through the last thing you used AI for, end to end.”

Second: “When has AI been confidently wrong for you, and how did you catch it?”

Both were published by Metaview, a recruiting-software company, building on the fluency rubric Zapier’s co-founder Wade Foster made public. Vendor source, but the questions stand on their own. Ten minutes with these two questions tells you more than an hour of tool-name bingo.

The second question is the real detector. Anyone who genuinely works with AI has been burned by it: the hallucinated citation, the confidently wrong data pull, the beautiful answer built on a misread premise. They’ll tell you the story with texture and a little embarrassment, and then, crucially, they’ll tell you what they changed about how they verify. A candidate who has only read about AI gives you abstractions about “hallucination risks.” Real usage leaves scar tissue, and scar tissue can’t be improvised in a screening call. The question is kind, too: you’re inviting a story they’ll enjoy telling, not running a quiz.

Zapier, the most codified case I found, is worth mining beyond the two questions. It checks AI fluency at four fixed points for every hire in every role (”The application, Initial (human or AI) screen, Skills test, Executive interview”) and probes four components: “AI mindset, strategy, building, and accountability.” The detail worth stealing came in the rubric’s second version: they now assess “AI fluency slope, not a snapshot,” the candidate’s trajectory rather than today’s toolkit. Listen for the slope in the workflow answer: someone whose story ends “and last month I changed how I do it” is climbing; someone reciting a setup from a year ago has plateaued. Metaview’s read of the same rubric describes

[truncated for AI cost control]