Thursday afternoon, product review. The PM walks in with the document she has worked on for two weeks: fifteen pages, competitor benchmarks, three solution options, a priority matrix in the right colors. She sits down, opens her laptop, and before she can share her screen an engineer on the team, one who six months ago would have asked "where's the ticket?", projects a working prototype of the third option. He built it Tuesday night with a coding agent, had two customers try it on a call Wednesday, and already has the session logs. The fifteen-page document isn't wrong. It just arrived after reality.
Anyone working in product has seen this scene, or will see it within the year. And the common reading, "AI will save PMs time, freeing them to focus on strategy", is the most polite and most wrong way to describe it. What happened in that room is not a time saving. It's a change of currency.
My thesis is this: AI agents don't automate the product manager's work, they demonetize his artifacts. The PM's craft has always stood on two legs we've grown used to confusing: producing coordination artifacts (PRDs, tickets, research summaries, roadmaps, decks) and making decisions you answer for (what to build, what not to, when, for whom). Agents are driving the cost of the first leg close to zero. Anyone who built a career on the first leg will find the market no longer pays for it. Anyone who can tell the two legs apart will find the second one suddenly worth much more than before. And for anyone running a company, the founder reading this, it changes how you hire PMs, how you count them, and how you grow them.
The mechanism: why the PM's artifacts specifically
To see where the wave hits, you have to look at how an agent works, not what the vendor's slide promises.
An agent is a language model placed inside a loop: it receives a goal, produces a plan, uses tools (reads documents, queries data, writes code, browses), observes the result, corrects, repeats. Nothing magical: the new thing of the last two years isn't the model's intelligence, it's the ability to sustain this loop for hours without losing the thread, on tasks that used to take a person days.
Now look at a PM's daily work through the eyes of this loop. Most of the artifacts a PM produces have three properties that make them the perfect target:
-
They're text-to-text transformations. A PRD is the transformation of conversations, data, and constraints into a structured document. A user research summary is the compression of twenty transcripts into five themes. A refined backlog is the translation of a strategy into tickets. These are exactly the operations language models are built for.
-
The context needed is largely already written. Call transcripts, support tickets, analytics, prior documents, Slack threads. An agent with access to these sources starts with 80% of the context the PM would have gathered by hand in a week.
-
The cost of the second copy is zero. Here's the real economic point. A PM produces a PRD in two weeks; an agent produces three alternative versions of it in an hour, and three more tomorrow if a constraint changes. When the marginal cost of an artifact collapses, so does its signal value: a polished document no longer proves two weeks of thinking, it proves a good prompt.
This last point deserves a pause, because it's the hinge for everything else. In organizations, the PM's artifacts were never just containers of information: they were currency. A well-made PRD bought credibility with engineering. The research summary bought a seat at the table with management. The detailed roadmap bought trust from the board. The PM traded artifacts for influence, and the organization accepted the trade because producing those artifacts was expensive, so the document was proof of work. Agents break the trade on the supply side: when anyone can produce a flawless document in an hour, the flawless document no longer proves anything.
Where the loop stops
If the mechanism stopped here, the conclusion would be brutal: the PM is finished. But the agent's loop has three precise stopping points, and they're exactly the points where the craft retreats and concentrates.
First: the unwritten context. The agent works on what's documented. But the most important product decisions depend on what isn't: the reason that enterprise customer is nervous and won't say so on recorded calls, the history of the feature we already tried to build in 2023 and why it failed, the political balance between sales and product that makes a certain choice impractical regardless of its merits. This context lives in heads and hallways. A PM who holds it has something no agent can retrieve with a query, not because the model is stupid, but because the data exists nowhere you can query.
Second: the asymmetry between generating and verifying. Producing an artifact has become cheap; deciding whether it's right hasn't at all. The agent's research summary is plausible, but did it overweight the three most talkative users? The PRD is complete, but is the missing edge case exactly the one that will blow up the biggest customer's renewal? Verification requires precisely the judgment and context the agent lacks, and the volume to verify grows with the agent's productivity. The bottleneck hasn't disappeared: it has moved downstream, from writing to verification. Anyone who saw what happened in software development with coding agents recognizes the pattern: the constraint is no longer writing the code, it's trusting the code that was written.
Third, and decisive: accountability. A product decision is a bet with irreversible consequences: months of a team's work, a positioning, sometimes a startup's cash. Organizations don't delegate bets like that to anyone who can't answer for them, and an agent can't answer for them by construction: it has no reputation to lose, no equity, no career.
An agent can't be fired. That's why it will never be asked to decide.
This limit isn't technical, it's institutional, and it won't move with the next model. Even an agent with superhuman judgment would stay a consultant: brilliant, tireless, and structurally unaccountable.
The implications: what changes for those who decide
So much for the mechanism. Now the chain of "and therefore", because that's where the game is played for a PM and for a founder.
And therefore: the PM/engineer relationship changes, but not the way people think. The lazy prediction is "fewer PMs needed". The real dynamic is more interesting. Engineers armed with agents no longer wait for the PRD: they prototype directly, like the engineer in the opening scene. This means the PM stops being the sole producer of proposals and becomes the arbiter among proposals, no longer a writer of options but an editor of options built by others (humans and agents). Fewer PMs per surface of product, yes; but the ones you need do work of much higher decision density. For a founder, the hiring question changes accordingly: not "how good is he at writing a PRD", the agent writes it, but "how good is he at taking apart a plausible PRD and finding the rotten assumption". These are related but not identical skills, and the second is far rarer.
And therefore: the bottleneck becomes decision bandwidth, and that's the founder's problem. If every team produces five times more prototypes, analyses, and proposals, but the company makes decisions at the same speed as before, the result isn't five times more progress: it's a queue of orphan prototypes aging while they wait for a verdict. I've seen companies celebrate their "AI-first transformation" by measuring the output generated, documents, experiments, demos, while the throughput of actual decisions stayed identical. It's the corporate version of a traffic jam: add lanes to the highway and keep the same single-bar tollbooth. For the CEO, the move isn't to buy more agents: it's to redesign who decides what at which level, pushing reversible decisions as close as possible to whoever has the context, and reserving the management table for the irreversible ones.
And therefore: proprietary context becomes the PM's asset. If artifacts are no longer currency, what is? Two things the agent can't get on its own: direct, continuous access to customers, and the institutional memory of past bets. The PM who spends the hours freed by agents talking to customers is accumulating exactly the asset that makes his verifications and his decisions better than those of anyone, agent included, who works only on the documented. The PM who spends those hours producing more artifacts is running faster on a treadmill that's dropping him off the back.
And therefore: the career ladder is broken, and almost no one is talking about it. This is the second effect the public discussion ignores. How has a senior PM always been formed? By doing for years the work agents do today: writing tickets, synthesizing research, preparing documents, and getting it wrong, and being corrected. The artifact grind was the training program for judgment. If agents make the artifacts, where does the senior judgment come from five years out? A founder who today stops hiring juniors "because we have the agent anyway" is running a cash maneuver with tomorrow's money: he saves a salary now and deprives himself of the senior he'll need in 2031. The alternative isn't to go back, it's to redesign the apprenticeship: juniors who verify agent output under supervision, who own the customer calls, who are judged on the quality of their objections and not the quantity of their documents. It costs managerial attention. Not doing it costs a generation of judgment.
And therefore, for those building in Italy. Italian product teams are chronically understaffed: the digital SME with a PM who is also the project manager, customer success, and sometimes sales is the rule, not the exception. This, paradoxically, is an advantage in this transition. Large American product organizations have to dismantle layers of roles built on artifact production, and that's painful. An Italian team of eight has nothing to dismantle: the agent doesn't replace a colleague, it fills a chair that was always empty. For once, the leverage favors the small and the short-handed. On the condition that you don't use it to produce ten times more documents no one will read.
The serious objections
The strongest version of the dissent isn't "agents don't work". It's this: we've heard this before. No-code was going to eliminate developers, offshoring was going to empty Europe's technical offices, and every wave of "revolutionary" PM tools mostly produced more meetings to discuss them. Why would this time be different?
The answer is in the mechanism, not the enthusiasm. No-code and offshoring moved the work, toward another tool, another country, leaving its coordination cost intact, which in fact often grew. Agents lower the marginal cost of production inside the daily work loop, with no handoff, no time zone, no contract. It's the difference between delegating a task and halving its price: the first reorganizes the work, the second changes what's worth doing. The earlier waves were the first thing disguised as the second.
A second objection: agent judgment will improve, and verification will be automated too. That's partly true, and it's honest to separate it by horizon. Today, verifying a product artifact requires a human with context. Eighteen months out, it's plausible that specialized agents will verify other agents on everything that can be formalized: consistency with the data, completeness against a checklist, internal contradictions. But the part of verification that depends on unwritten context and the part of the decision that requires accountability stay out, for the institutional reasons above: whoever signs answers for it, and to answer you need to be able to lose something. Anything beyond that, "agents will develop reliable autonomous judgment on irreversible bets", is speculation today, and should be treated as such.
There's finally a case where my thesis really gives way, and it should be said: in very small teams, the technical founder with three engineers, agents can make the dedicated PM redundant, because the founder already holds both legs of the craft, and the machines make the artifacts. But this isn't something the agents created: in those teams the dedicated PM was already a questionable choice. The agents just made it obvious.
The landing
The test I'd propose to every PM, and to every founder evaluating his PMs, is unpleasantly simple. Take the last week of calendar and split the hours into two columns: those spent producing artifacts a good agent would do at 80% of the quality, and those spent doing what the agent can't, talking to customers, verifying with real knowledge, deciding and answering for the decision. In most cases I've seen, the first column wins by a wide margin.
That first column is already lost. Not this year, not for everyone at the same time, but the direction isn't in question: it's value the market will stop paying for, because its cost of production is going to zero. The question that matters isn't how to defend it. It's what to put in its place.
Three moves, in order of priority. Deliberately shift the hours from production to verification and decision, and make it measured: a PM should be able to say how many decisions he advanced this week, not how many documents he delivered. Invest the freed hours in the one asset that appreciates in this scenario: proprietary context, meaning customers seen in person and the memory of past bets, written down while there's still someone who remembers it. And redesign how juniors grow before the old ladder finishes crumbling, because the senior judgment of 2031 is being formed, or not formed, now.
The product manager who produces documents is a craft in liquidation. The product manager who answers for the bets has never been worth so much.