The Hollowed-Out PM: What AI Took From Product Management — and What It Can't

A founder asked me last month whether she still needed to hire a product manager. Her engineers were shipping features they'd scoped themselves with Claude Code, her competitive research came back from Perplexity in minutes, and her last three "PRDs" were drafted by an LLM from a Slack thread. "It feels like the PM job is evaporating," she said. "Am I about to hire someone to do work the tools already do?"

She's half right. The product manager role is being hollowed out — but only at the edges. The tasks that used to fill a PM's calendar are exactly the tasks AI is now good at. What's left when you scoop those out isn't a smaller job. It's a denser, harder one. And if you're a founder deciding how to staff product, knowing the difference is the whole game.

The outer ring: what AI is actually taking

Walk through a traditional PM's week and most of it is connective tissue — moving information from one place to another, summarizing, formatting, triaging. That's the layer coming off.

Backlog grooming. Deduping tickets, clustering related requests, suggesting priority based on stated criteria. Linear's AI triage alone reportedly cuts backlog work by around 70%. The machine is genuinely better at this than a tired human at 4pm — it doesn't get bored, and it doesn't have favorites.

Research synthesis. This was the big one. Taking forty customer interviews and clustering them into themes used to take a PM a week and consumed real reading time. Tools like Dovetail now tag transcripts, surface themes, and pull supporting quotes automatically; Claude or ChatGPT will do a one-off synthesis from raw notes in minutes. The thing that made discovery slow — reading everything — is no longer the constraint.

Status reports. Standup summaries, stakeholder updates, release notes, the weekly "where are we" deck. All of it is text-transformation over information the system already has. Granola captures the meeting; the agent writes the update. Nobody misses writing status reports.

Competitive research. Mining public RFPs, summarizing competitor docs and changelogs, building a first-pass feature comparison. Perplexity produces cited market research on demand. The work that used to eat a Friday afternoon is a prompt.

Add it up and you're reclaiming a serious fraction of the week — the part of the job that was administrative all along. Over 70% of PMs now use AI tools daily, and the consistent report is the same: less time on documentation and synthesis, more time on everything else.

The mistake my founder friend was making is assuming that because the visible work is shrinking, the job is shrinking. It's the opposite. When you remove the busywork that used to fill the calendar, what remains is the part that was always the actual job — and it's now the whole job.

The core: what's left when the busywork is gone

Here's the uncomfortable part for anyone hoping AI makes product management easier. The residual core is the hard stuff, and there's nowhere left to hide.

Deciding what's worth building. AI can tell you what customers said. It cannot tell you which customer to believe, which market to serve, or which feature to kill. Those are judgment calls that require organizational context, a real model of the customer, and accountability for the outcome — none of which fall out of a prompt. As Productboard's team put it, when the cost of building drops, the responsibility to choose well goes up. Velocity stops being the differentiator. Judgment is.

Strategy and sequencing. How you position against a competitor, when to cut a feature, how to order a roadmap so each release sets up the next — this is the connective reasoning that AI is worst at, because it depends on context the model doesn't have and tradeoffs the model isn't accountable for. (I've written before about why the traditional roadmap fails at seed stage; AI makes the lightweight alternative more important, not less.)

Taste. The judgment of whether a thing is actually good — whether the flow feels right, whether the copy lands, whether you'd be proud to ship it. AI can generate ten variants. Choosing the right one, and knowing why, is taste, and taste is a human's job.

Reviewing AI output. This is the genuinely new skill, and it's the one founders underrate. When an LLM drafts your PRD, synthesizes your research, or proposes a priority order, it does so with total confidence — including when it's wrong. Models hallucinate fabricated facts, invented sources, and capabilities that don't exist, and they present them as authoritatively as the true ones. The research on this is blunt: people over-trust confident-sounding AI, and extended reliance leads to deskilling — you stop being able to catch the errors because you've stopped practicing. The PM's new core competency is critical review: knowing the domain well enough to spot the plausible-but-false answer before it becomes a roadmap decision. This is the same discipline I push on engineering teams reviewing AI-generated code and why your eval suite matters more than your prompt — the output is cheap, so the judgment about the output is where the value moves.

What this means if you're a founder

So back to the original question: do you still need a PM?

The honest answer is that you need product thinking far more than you need a product manager — and AI changes where that thinking has to live. At the earliest stage, founder-led product plus AI tooling genuinely covers a lot: you can run discovery, draft specs, and validate prototypes without a dedicated hire. The handoffs that used to require a PM to coordinate are collapsing into shared context, and your engineers are absorbing more product decisions than they used to.

What you can't outsource to tooling is the judgment core. The moment you feel decisions getting made by default — features shipping because they were easy, not because they were right; research getting synthesized but never interrogated; a roadmap that's really just the loudest customer's wishlist — that's the signal you need someone whose entire job is the dense center, not the disappearing edges.

And when you do hire, the bar has moved. The PM who lists "wrote PRDs, ran competitive analysis, managed the backlog" is describing the part of the job a tool now does. The one worth hiring talks about decisions they got right, calls they reversed, and how they caught the model being confidently wrong. Strategy and business acumen now top every ranking of the skills that matter for the role — because they're what's left.

The counterpoint: where this thesis breaks

I don't want to oversell the clean story. A few places it gets messier.

Someone still has to do the edges — they just do them faster. "AI handles synthesis" doesn't mean synthesis happens unsupervised. A PM still drives the tool, checks the clustering, and decides what the themes mean. The work compresses; it doesn't vanish. If you fire your PM and assume the tools cover it, you've just moved the reviewing-AI-output burden onto a founder who has less time to do it well.

Deskilling is a real trap, especially for juniors. If the administrative work was how early-career PMs learned the craft — reading every transcript is how you develop a gut for customers — then automating it away risks producing PMs who can operate the tools but never built the judgment underneath. The leverage in this new world flows to people who already have taste. Building that taste from scratch, when the practice reps are automated, is an unsolved problem.

The core is hard to evaluate. Busywork was at least legible — you could see the deck, the backlog, the report. Judgment is invisible until it's tested by an outcome months later. That makes the hollowed-out PM harder to manage and harder to hire for, because the thing you're paying for doesn't show up in a status update. Which is, of course, the point.

What to do this quarter

If you're a founder or eng leader reasoning about product staffing in the AI era:

1. Audit where your product decisions actually get made. For one sprint, note who decided what to build and why. If the answer is "it was easy" or "the loudest customer," you have a judgment gap no tool fills.
2. Give your team the AI edge-tooling now. Discovery synthesis, competitive research, draft specs — get these into the workflow so nobody's calendar is full of busywork. This is true whether or not you have a PM.
3. Make "review the AI" an explicit job, not an assumption. Decide who is accountable for catching hallucinated facts and lazy priorities before they ship. Name the person.
4. Reframe the PM scorecard around judgment. Evaluate decisions and reversals, not artifacts produced. Artifacts are now cheap.
5. Hire for the core when defaults start deciding for you. Not before — but the day product decisions are being made by inertia is the day you need someone who owns the dense center.

The product manager role isn't dying. It's being concentrated. The tools took the ring of busywork that made the job look full, and left behind the part that was always the real work: deciding what to build, judging whether it's any good, and catching the machine when it's confidently wrong. That core doesn't shrink as AI gets better. It gets more valuable.

If you're weighing how product should work on your team — whether to hire, what to expect, how to structure the decisions that matter — that's exactly the work I do as a fractional CTO. Book a 30-minute call and we'll map it to your stage.