On World Models
How should product managers take advantage of the leverage they get from AI?
A mentor and I were talking recently about how product leaders are actually changing the way they work with agentic AI tools. Not whether to use them, but how to get the most out of them without burning tokens, losing proficiency, or creating new alignment problems across the team.
The conversation led somewhere I keep coming back to: world models.
But before getting there, it helps to be clear about where the actual leverage is for product managers.
Where the leverage is
AI tools have compressed the time between an idea or insight and a tangible prototype you can validate. That’s the leverage for product managers. Everything else in the product development cycle still needs to happen.
Product still needs to know what to build, for whom, and why it matters to the business now or later. Design still needs to figure out the right interaction patterns for the solution, and how to handle the edge cases that come with every use case. Engineering still needs to assess how the current architecture supports what needs to get built, and what the delta is between what exists and what needs to exist.
Those functions haven’t changed. What’s changed on the product side is the process of getting from insight to validated idea, can now move faster.
That speed is real. But it creates a problem.
The alignment gap
When you’re moving faster from idea to validation, you’re also generating more frequent and more rapid updates to your understanding of the user’s problem and the right solution. Each of those updates needs to travel across your EPD pod and across the organization. To your cross-functional teammates. To senior leadership.
The faster you move, the more surface area there is for misalignment. Not just within your immediate team, but with everyone who needs a clear and actionable understanding of what you’re building and why.
This is the part the AI tools don’t solve. Getting to a validated idea quickly is only useful if the people responsible for designing and building it are working from the same understanding you are.
So the question becomes: how do you take advantage of the new speed without constantly rewriting the PRD, without running alignment sessions every time your understanding shifts, and without the organization losing the thread?
World models as a possible answer
The idea of a world model is essentially a persistent, continuously updated representation of three things: the user’s problem, the current solutions in production, and where the product needs to go next.
It’s flexible enough to absorb the continuous learning that comes with faster iteration. And it gives product, design, and engineering a shared surface to work from rather than a document that’s already out of date by the time it’s shared.
What that actually looks like in practice is still an open question. A few directions worth exploring:
One version is an always-present prototype that mirrors your design system and interaction patterns, continuously updated with what your team is learning about the user and the problem. Not a static spec. Something living.
Another version sits somewhere between prototype and production. A staging environment that’s continuously updated with new learnings, where product can interact with and refine features before deciding what actually goes to production. Almost like an interactive backlog you can feel rather than just read.
There may be other versions entirely.
Why this matters
The risk of not having something like a world model is that the speed AI tools give you becomes a liability. You’re moving fast, but the organization can’t keep up. You’re validating ideas quickly, but the team is building from different mental models of what the product is supposed to be.
The goal isn’t to slow down to stay aligned. It’s to build the infrastructure that lets alignment travel at the same speed as the work.
I don’t have a definitive answer on what that infrastructure looks like. But I think a version of a world model is where product leadership probably needs to go if we’re serious about using these tools in a way that’s both tightly aligned and cost effective.
The leverage is real. The question is whether we’re building the systems to use it well.