Product managers have always been in the business of solving for ambiguity — they insert themselves to resolve uncertainty around execution, product craft, analytics, and customer needs. As AI advances, so does the argument that the PM’s role is being made obsolete by increasingly capable models. However, debate around the viability of the function misses an important point: there isn’t any less ambiguity involved in bringing products to market and scaling them today, but the tools and opportunities are completely new. Product managers that ignore this dynamic risk irrelevance.
The current class of product leaders came of age during the mobile/web transition and were trained in mobile app-specific methodologies like growth accounting, mobile-first product craft, and owning an app icon on the homescreen. There was a lot of discussion around how to avoid injecting desktop-era product holdovers like hover state cues and click-first design metaphors into mobile products.
What are the new skeuomorphic metaphors that will constrain mobile product management when building in AI? And what are some of the best practices to consider? A few come to mind:
1. “Interview” your models alongside your customers.
Large models are inherently probabilistic, meaning that given the same input, the outputs are varied rather than deterministic. Because AI models have these stochastic outputs and exhibit emergent behaviors — unexpected actions that aren’t explicitly programmed — PMs now have to spend as much time “interviewing” their models as they do their customers, probing to understand the models’ capabilities and constraints. PMs should be asking: What types of “noise” or unpredictable outputs do these models generate that I can leverage in my product? How does the model respond to edge cases, and when do I need it to be stable?
Honing this intuition can lead to products like Websim, the AI-powered simulator that generates strange, unexpected websites that make you feel like you’re peering into the model’s mind. Instead of reining in the model to produce polished, conventional outputs, Websim’s builders leaned into the weird.
To find the value in the unexpected, PMs also need to become skilled at writing evals: structured tests that help you see where your model performs well and where it’s falling short. Evals aren’t just about measuring accuracy, they’re identifying and assessing emergent capabilities to inform your product design.
2. Don’t shy away from extreme products at extreme prices.
We’re now seeing a class of software products that can do things that were unimaginable just a few years ago: an AI nurse that calls patients with information and reminders before surgery; a tool that generates complicated web applications from a single prompt; a product that performs sophisticated research and analysis that previously would have required a team. In this world, there really isn’t a ceiling on how much you can charge.
When ChatGPT launched its $200/month subscription last year as a mass consumer product, it seemed like wishful thinking. Now it’s a go-to, daily tool for power users. Similarly, AI products like Krea, Cursor, Midjourney and many others have been successful in aggressively exploring price ceilings, rather than optimizing for price floors.
Our belief is that software will be the #1 category of discretionary spending for consumers in the near future. Against that backdrop, the PM’s product design prompt should be: “What does the $1,000/month version of our product look like?” Then work backwards to deliver it.
3. The elusive AI moat: first and fast.
Consumer AI companies should be intentional about exploring new moats, whether through LoRAs, proprietary workflows, integrations with other software, or new channels like voice and phone. I believe one of the most overlooked moats is emotional: simply being willing to build products that carry emotional valence. Apple, Google, and the like have a thousand committees designed to ensure that the messy aspects of the human experience (disagreement, persuasion, sexuality) are never surfaced in their products. Language models function as mass “averaging machines” optimized for consensus, leading to products that can be bland, uninspiring, or outright bad. Startups, on the other hand, can build around those edges — emotion, friction, intensity — to create products that feel unique in the marketplace.
Network effects remain the gold standard of software moats. However in the competitive landscape of AI, where the volume of products being spun up is so enormous, many of the traditional frameworks for establishing moats may not apply. For example, systems of record can now be indexed via vision models + RPA, potentially rendering the strength of this moat less powerful. As builders gain access to the same models and infrastructure, “soft” moats — like mindshare and momentum — that once seemed too weak to sustain a competitive advantage are becoming increasingly important in consumer AI. A playbook we’re seeing more of: first and fast. Leading founders are the first to build a product, then stay at the front of the pack by continually shipping new features and capabilities.
4. Models are platforms, not products.
The first generation of AI products were really webpages in front of models, with the foundation models doing the heavy lifting of generating images, writing poems, and delighting users with new capabilities.
As these foundation models increase in number and sophistication, users will increasingly require opinionated workflows around the model to make the most of them. For example, text-to-app products like Replit, Lovable, and Bolt are a miraculous experience for prototyping new ideas. But moving from prototype to production will likely require more advanced interfaces that support fine-tuning and customization. Thus, we believe the next generation of large-scale AI products will be opinionated and sophisticated products built around foundation models.
5. Reflexive AI use is tipping from differentiator to default.
You can’t productize a system you don’t understand. That means it’s not enough to dabble in ChatGPT, you need to understand the difference between a language model and a reasoning model. Have you tried Deep Research, Operator, Gemini Flash, custom GPTs, and GPT-4o in multimodal mode? Have you read up on chain of thought, or observed it when using DeepSeek or any of the other reasoning models that expose it? The single most important intuition-builder for PMs is reflexively using AI products every day, in every part of their job. This view is quickly tipping into consensus, as the CEOs of Shopify, Duolingo, Box and many more declare their companies pivoting to AI-first in all efforts.
Over time, PMs should naturally incorporate AI products into their daily habits and become beacons for best practices at their company.
