What does product management look like now?

Marty Cagan thinks the typical product team could shrink from eight people to three. If you’re a PM, the uncomfortable part is figuring out whether that smaller team still needs you.

In February 2025, Cagan published a prediction on the SVPG blog: within a few years, AI could allow much smaller product teams to do work that currently requires eight people. In his scenario, a fifteen-team scale-up with roughly $24 million in fully loaded headcount costs could get comparable output for closer to $1.8 million.

He was direct about which roles were most exposed. Product owners and feature-team PMs, he wrote, are “very vulnerable.”

That’s a lot to hear from the person whose book many of us were handed when we started in product.

The usual response is reassuring: AI will take care of the busywork, leaving PMs more time for strategy and stakeholder alignment. I’d like to believe that’s how it plays out. For some people, it probably will.

But it assumes companies will keep the same number of PMs once a substantial part of their workload disappears. I’m not sure that assumption holds.

What happens when the busywork is most of your job?

Productboard and UserEvidence surveyed 379 product managers in October 2025. Respondents reported saving about 33 hours a month with AI, largely on familiar tasks: writing PRDs, researching competitors, summarizing customer calls, preparing presentations, and drafting roadmaps.

Almost a working week each month. That’s a meaningful amount of time.

Whether it’s good news for your career depends on what you were hired to do. A senior PM with real authority over product direction might happily hand off those tasks. They already have a backlog of decisions that need their attention.

A PM on a feature team may be in a different position. If the company mainly expects you to turn requests into specs, keep the roadmap current, and explain progress to stakeholders, AI is eating into the work your manager sees you doing.

You might use the time to take on more valuable responsibilities. Your company might also decide it needs fewer people in the role.

Adam Judelson, who runs product at Palantir, has argued that generative systems already perform the less technical parts of PM work better than humans. You don’t have to agree with the full claim to take it seriously. It tells you something about what at least some product leaders will expect from their next hire.

Being good at the existing job may no longer be enough.

Faster delivery puts more pressure on product decisions

In Lenny Rachitsky’s research on the changing PM skill mix, data literacy, customer insight, systems thinking, and strategic prioritization ranked highly. The process-heavy parts of the job, including writing specs and maintaining roadmaps, lost ground.

That makes sense alongside what’s happening in engineering. Studies at Microsoft, GitHub, and large banks have reported productivity gains from AI coding tools, although the size of those gains varies by task and context.

As implementation gets faster, teams have less time to resolve uncertainty before someone starts building. A vague product decision can become working software surprisingly quickly. It can also become a month of work on something customers never needed.

That puts pressure on the PM to have a useful point of view earlier.

After a customer call, can you explain what should change because of what you heard? Can you tell whether a request reflects a broader problem or one customer’s unusual setup? When engineering challenges your proposed scope, do you understand the tradeoff well enough to make a decision?

AI can help you work through those questions. It doesn’t take responsibility for the answer.

This is where I think the role gets harder. Producing a plausible document is becoming easy. Knowing whether the document points the team in the right direction still takes experience, context, and a willingness to be wrong.

Some PMs are already working differently

The useful examples aren’t necessarily people who have automated the most tasks. They’re people who have found something more valuable to do with the tools.

Shreyas Doshi has written about using LLMs to challenge his thinking. Pre-mortems, strategy reviews, and product critiques give him ways to test an argument before he acts on it. A model that helps you notice a bad assumption can be worth more than one that finishes your memo.

Lenny Rachitsky offers a different example. His business spans a newsletter, podcast, community, job board, and courses. AI supports work across research, editing, transcription, and analytics. His operation is an example of how an experienced person can extend their reach well beyond what they could do alone.

Claire Vo went further by building ChatPRD. Her experience as a product leader gave her a close view of a recurring problem, and she built a tool to address it. There’s something worth paying attention to there: PMs can now test and build more of their own ideas instead of waiting for a team to become available.

For PMs moving into AI products, Aman Khan’s work at Arize AI is more directly relevant. He writes about evaluations, agent observability, and the difficulty of making LLM features reliable in production. That work demands enough technical understanding to investigate failures and make informed product decisions. Knowing how to get a convincing demo out of a model won’t get you very far.

Ravi Mehta has applied his product leadership experience to advising teams on AI and how it changes their work. The value is in being specific about what an organization should do differently, rather than telling it to “embrace AI.”

These are prominent people with advantages most PMs don’t have. I wouldn’t treat their careers as a template. Still, they offer useful examples of where the work can go: deeper into technical problems, into building, or into helping teams make decisions they’re struggling to make themselves.

What I’d do as a working PM

I expect PM job descriptions to change faster than many companies’ career frameworks. You may still have the same title while the expectations underneath it shift considerably.

I’d start by looking honestly at the last month of work. Which parts required your experience? Which parts needed your access to customers or your understanding of the business? How much of the rest could someone produce with a model and a decent brief?

That’s a more useful exercise than asking whether AI can “replace product managers” in the abstract.

From there, I’d go deeper in a domain. Learn why customers work around the product, what makes a sale fall apart, and which constraints keep defeating otherwise sensible ideas. A model can retrieve plenty of information about an industry. It’s much harder to reconstruct the context you gain by staying close to the work.

I’d also get comfortable enough with AI tools to recognize when their output is weak. Use them on real problems. Build something, test it, and investigate where it fails. You’ll learn more about what to trust than you will from collecting prompts.

And I’d look for decisions I could take responsibility for. That might mean resolving a disagreement over priorities or recommending a smaller release when everyone wants the bigger one. These are uncomfortable parts of the job. They’re also where a team most needs someone who understands the situation and will stand behind a call.

Cagan’s three-person team may or may not become the norm. I wouldn’t build a career plan around the exact number.

But I would pay attention to the question behind it: if your team were hiring from scratch today, what would they need you to do?

Sources

• Marty Cagan, “Product Teams in the Age of AI,” SVPG, February 2025
• Productboard and UserEvidence, “The State of AI in Product Management,” October 2025 (n=379)
• Lenny Rachitsky, “What product managers actually need to be great,” Lenny’s Newsletter, August 2024
• Adam Judelson, Palantir, public remarks 2025
• Shreyas Doshi essays at shreyasdoshi.com
• Lenny’s Newsletter at lennysnewsletter.com
• Claire Vo at chatprd.ai
• Aman Khan at arize.com
• Ravi Mehta at ravi-mehta.com