The PM role isn't disappearing.
It's splitting into three.
Everyone's writing think-pieces about how AI is killing product management. They're looking at the wrong layer. The role isn't dying — it's separating into three new roles, and the operators who own them will own the next decade.
Every week another headline says product management is finished. The model is going to write the spec. The agent is going to run the standup. The PMs are going to be the first office workers replaced. The math, the timeline, the confidence — all of it is wrong, and all of it is wrong in the same way.
The people writing those pieces are watching the old PM role. They're looking at the version of the job that was already breaking before any of this shipped — Jira hygiene, status decks, ceremony management, calendar Tetris. That role is dying. It deserves to die. But the function it sits inside of — translating a messy reality into shippable software — isn't going anywhere. It's just splitting into three roles that used to live inside one person.
The job didn't disappear. It got unbundled.— What I've been telling every PM who DMs me about this
I run delivery programmes by day and the AI Delivery Lab by night. I have seen this shape now in maybe sixty cohorts of teams — startups, agencies, public-sector transformation orgs, banks. The teams that are actually shipping with AI have all quietly reorganised around the same three roles. They don't always call them this. But this is what I see.
/ 01 — THE FIRST SPLITThe Discovery Operator
The first role is the one most PMs are worst at, which is also the one AI makes the cheapest. Discovery. Interviews, synthesis, problem framing, market reads, opportunity sizing. The work upstream of any ticket.
For ten years the orthodoxy has been “PMs do discovery.” In practice, almost nobody did. There was always a deadline, always a quarter, always a roadmap to defend. Discovery got squeezed into a half-day off-site and a Miro board full of stickies that nobody read.
Now the Discovery Operator runs ten interview transcripts through a custom agent and gets back a structured synthesis in twenty minutes. They run the same loop against support tickets, sales call recordings, the competitor's release notes. The artefact still has to be good — somebody has to design the synthesis prompt, calibrate the rubric, sanity-check the output, and turn it into an opportunity statement a team can argue about. That's the role. The leverage is enormous. The skill is taste.
What good looks like: a Discovery Operator can compress a month of qualitative research into a week, and the synthesis is betterthan what a PM would have done by hand — because the model surfaces patterns across more sources than a human can hold in working memory.
What bad looks like: a generic summary of ten interviews that reads like LinkedIn slop. No distinct point of view. No falsifiable claim. The model is doing the work and the operator is laundering it.
/ 02 — THE SECOND SPLITThe Delivery Conductor
The second role is the one closest to the old PM job — but stripped of the ceremonies and re-armed with agents. Call it the Delivery Conductor. They own the path from “we believe this is worth building” to “this is in production and the metric moved.”
The Delivery Conductor doesn't run standup. The standup runs itself — agents write the status doc from the commits, the ticket activity, the CI signals. The Conductor reads it, asks one good question, and gets out of the way. What they do own is the queue — the order things get built in, the dependencies, the trade-offs nobody else is paid to hold in their head.
They also own one thing the model can't do: the politics of the org. Which exec needs to be told first. Which team will block the launch if they're not in the room. Which compromise is fine and which one will rot the product. That's still a human job. It's a smaller human job than it used to be — and that's the point.
- Read the agent-written status. Don't re-write it. Trust the artefact.
- Identify the one decision that's stuck. Make it, or escalate it. Don't host a meeting about it.
- Tell three people what shipped this week. Not in Slack. In sentences.
/ 03 — THE THIRD SPLITThe AI-Native Builder
The third role is the new one. It didn't exist three years ago. Call it theAI-Native Builder. Half-PM, half-engineer, all-operator. They don't write the production code. But they write enough code — usually with Claude Code or an equivalent — to ship the v1 of a feature themselves, see what's wrong with it, and hand a working prototype to the engineering team.
This is the role that breaks the old PM-engineer contract. The old contract was I describe, you build. The new contract is I prototype, you harden. The prototype is the new spec. The engineering team gets to spend their time on the parts that actually need engineering — performance, reliability, security, the hard data model — instead of arguing about whether the dropdown should be a radio.
The prototype is the new spec. The engineering team gets the parts that actually need engineering.
/ SO WHATWhat this means for you
If you're a PM today, you don't have to pick one of the three. You probably already do parts of all of them. But you should know which one you're best at — and which one you should be deliberately practising.
- Stand up Claude Code locally this week. Use it for one real thing.
- Replace one status meeting with an agent-written digest. See what breaks.
- Pick one ten-interview dataset you've been sitting on. Run a real synthesis.
The PM role isn't disappearing. It's splitting into three — and the operators who own the new roles will own the next decade. That's the whole thesis. The rest is just getting your hands on the tools.