The framing I see most often: AI makes you 20% faster. Autocomplete for knowledge work. Useful, not transformative.
I think this undersells it. The more interesting effect is what it does to scope.
What used to require a team
A year ago, taking a product from problem statement to customer-ready demo involved: discovery interviews, synthesis, requirements document, wireframes (designer), copy (writer), slide deck (someone), data pull (analyst). Weeks of calendar coordination before you had anything to show anyone.
Now a meaningful chunk of that is compressible into days. Not because each step is 20% faster. Because several steps that required specialist time can be done with AI assistance at a quality level that’s good enough for the purpose.
The wireframes I make aren’t as good as a senior designer’s. But for a hypothesis-validation session with a customer? They’re fine. They do the job.
What this changes about prioritisation
If doing a thing well takes 3 weeks, you pick your bets carefully. You say no a lot. You queue things.
If doing a thing well enough takes 3 days, you can afford to be wrong. You can build the prototype, show it, learn it’s the wrong direction, and still have time to pivot. The cost of a wrong hypothesis drops.
That’s not just an efficiency gain. It changes the optimal strategy. The team that can validate in days and the team that needs weeks should be running different playbooks.
The PM’s job shifts
If speed of output is no longer the bottleneck, the bottleneck moves. It moves to: what’s worth building? What’s the right problem? What would validate this?
Those are judgment calls. AI doesn’t make them. You do.
The PMs who adapt well to this aren’t the ones who use AI to do more of the same work faster. They’re the ones who use the reclaimed time to do more of the thinking that only humans can do: talking to users, challenging assumptions, deciding what not to build.
The honest version
AI also makes it easier to produce confident-looking output that’s wrong. A well-formatted document with coherent structure is not the same as a correct document. The polish can obscure the gaps.
The skill that matters more, not less, in an AI-accelerated workflow: knowing when the output is good enough, and when it’s plausible but hollow.
That judgment doesn’t get automated. It has to be developed.