The Three Flavors of AI Product Management — and What Separates the Top 5%

Aman Khan starts by defining the base job: a product manager brings together design, engineering, business development, operations and sales to ship impact, representing the customer and driving toward a solution. On top of that, he breaks the AI component into three distinct flavors. The first is the AI platform PM — the bucket he sits in at Arize, an observability and evaluation platform for AI that began with machine learning ranking, regression and classification models and has since pivoted toward any kind of AI, now mostly large language models.

Platform PMs build tools for AI engineers, and Aman argues this space is unusually raw. Interfaces for building on top of OpenAI and Anthropic APIs are clunky, and people have only been building on top of LLMs for a couple of years. The platform PM's job is to let AI engineers answer whether their app does the thing they expect, and to let PMs at customer companies show leadership that the app is actually moving the business.

The second flavor is the AI product PM, where the core product is centered on AI: ChatGPT, or NotebookLM as a recent example. Here researchers and engineers push the boundary of the technology, and the PM's job is to package that technology into something a consumer or enterprise can actually use. The core competency being packaged and distributed is the model itself.

The third flavor is the AI-powered PM, who uses AI across the entire stack of a role they may already be doing. This PM probably does not have access to a research team that can build a model from the ground up, but can still use LLMs to build the best experience for their customers. Aman expects AI to become as common as the database in SaaS applications, which means most PMs will soon be one flavor of AI product manager.

Asked which tools make him more productive, Aman says he is obsessed with Cursor and Replit, which he uses constantly to build prototypes and to understand for himself what is possible. Showing up with a functional prototype is super enabling for a PM: it will not be what ships to production, but the buttons work and it is wired up well enough to tell the story.

For interface work he points to v0, Vercel's prompt-to-UI starting point. It is best known for producing a polished landing page, but with the right prompt and iteration it can produce a pretty good working UI — a strong starting point for the first mock-up conversation with a design team.

For visual thinking he reaches for image generation: Midjourney or DALL-E for a logo, or for turning a user story into something visual and immediately legible to the team. Which tools matter depends on the company — an experience-led company will lean on image generation, and if you are building physical products, emerging 3D modeling tools have made that work much easier too.

Aman describes himself as an outsider: he studied mechanical engineering, took no computer science courses, learned much of it on his own, and has no machine learning coursework, master's or PhD. What he believes matters for both platform PMs and AI product PMs is caring about the end customer — being in love with the problem, and then pushing the boundaries of what technology can do to solve it. He compares it to his own surprise at how hard it is to build a dashboard that shows whether a model is working, and how that deceptively simple question forces you to push the stack forward.

On skills, he recommends two dimensions running in parallel: understand the fundamentals of what machine learning and AI actually are, and then let curiosity drive you to apply that knowledge to problems or industries you care about. He points to Andrej Karpathy's roughly hour-long video on the foundations of LLMs as a great starting resource — and notes you can feed it into NotebookLM to get a more succinct version. There is no college degree in product management, and certainly none in AI product management, so you stand out by showing genuine interest and building a skill stack rather than credentialing.

Aman has sat on the hiring side, and he says the process is designed to answer three questions: can this person do the job, are they excited about the work we do here, and do I like working with them. A portfolio of prototypes you have actually built — even imperfect ones — answers the first two before the interview starts, because it shows how you think, how you spend your time, and what you are curious about. That leaves you to nail the phone interview and demonstrate culture fit, which is what most multi-round interview panels are really probing for.

Counterintuitively, Aman thinks it is easier to break into AI product management now than it was before. Earlier, candidates needed real depth in machine learning, model types, training data splits and launch infrastructure to get a shot at an AI company. Today the role is largely about building experiences around or for other AI builders, and the tooling has collapsed the cost of demonstrating that you can do it.

Lenny offers the hot take that AI tooling makes PMs more valuable, not obsolete: these tools are excellent at building whatever you tell them to, which makes knowing what to build, finding the problem and articulating it clearly the scarce skill. Aman agrees, adding that many executives feel pressure to get into AI but default to doing what has worked so far, and the AI PM's role is to be the representative of how to use these tools at their highest leverage. Because these tools also let a PM communicate an idea to design, to engineering and to leadership in their native language, they max out the core PM skill of influence.

Aman marks the ChatGPT launch roughly two years ago as the iPhone moment for AI products: the interface was intuitive, the model felt human-like, and everyone was impressed. Looking back is like looking at the original iPhone — by today's standards it barely does anything, because users get conditioned to the technology continuing to improve.

From Arize's vantage point across customers like Spotify and Instacart, Aman saw the same first move almost everywhere: another ChatGPT, but on top of the company's own data. Every AI PM he knew was building an internal chatbot over the knowledge base, and how much usage those products actually got is not well understood. The top 5%, in his view, resist that gravitational pull and ask whether this is the problem the business needs right now — at Arize the real AI interface was automating data analysis, with a small chatbot shipped mostly to learn how the technology works.

He sees the same dynamic repeating with AI agents: a flood of pitches that describe a solution — an agent for this, an agent for that — rather than a problem. His advice is to ask whether you should build that agentic layer at all, or whether it is better left to the foundation model companies while you make the experience so seamless in your own product that it does not even feel like AI. You probably will not invent the next agent framework, but you can apply the technology where your organization needs it, which is where the AI PM stands out.

Every PM is tasked with moving a business metric, but Aman points out there is no metric for building AI prototypes — companies routinely say AI is not expected to move revenue, which leaves no way to tell whether the work is succeeding. His answer is to measure shots on goal: how many experiments are you taking in the first place.

Hackathons are his second lever, because getting people hands-on removes the feeling that the technology is unapproachable. Bring a list of real problems and expect most attempts to fail: out of ten problems you apply AI to, nine may not work, and finding out where the ideas stick is the point.

Arize's own hackathon produced a useful failure case. Engineers proposed a Slackbot that would read a question in the support channel and ping the right on-call person — an obvious AI use case for classifying problems and routing them. It turned out to be genuinely hard, because so much context is missing: someone worked on that problem two weeks ago, but now they are on another project and a different team owns it.

His third lever is studying the details of AI products that feel magical. The cautionary tale is Betty Crocker's cake mix that only required adding water, which tanked sales; changing the instructions so customers had to add an egg made sales skyrocket, because people want some control over the experience — the Ikea effect applied to AI. The same logic explains why a Waymo still lets you adjust the air conditioning and the music, and why Aman argues Blue Apron is not competing with DoorDash on feeding people but on the feeling of cooking.

Aman borrows Kevin Yien's phrase that an AI PM has to walk and chew gum: the job is not to ship AI products, it is to solve customer problems while being pushed by the company to move a business metric. The goal is to make space for the things that do not move a KPI — prototyping yourself, letting the team run hackathons, and tearing down great experiences. Arize livestreams a teardown of a cutting-edge product such as NotebookLM every week or two, which builds the team's understanding of the space without a direct metric attached.

That also means accepting that some experiments will not land. Aman borrows Lenny's line that even a great chef occasionally serves a bad meal: you will have bad experiences in your AI product diet, but the point is to keep iterating, which is how both products and organizations get better at deploying AI.

The first thing he names for surviving as an IC PM is energy, a lesson from watching Arize's CEO. When you are unsure which direction to go, energy counts: show up to a meeting with more of it and much of the friction disappears, while showing up flat is something people pick up on. He describes the stuck feeling as product block rather than writer's block — you are trying to reach the next idea or milestone and cannot see it yet.

During one existential decision — how much to invest in LLMs versus serving the existing customer base — he did not wait for clarity. He became the company's outbound salesperson, writing LinkedIn copy, asking sales how to do it, and maxing out his five LinkedIn messages a day to reach people with AI in their title. The point was partly to learn how to message the product, and partly to show the team he would get into the details with them — the player-coach, LeBron-style mentality of drawing up the play and then playing it.

Aman frames the IC PM's core challenge as waiting versus wandering, using the image of a camp sending out scouts to decide where to move next. Companies, especially large ones, lean toward doing what has worked so far: let's wait and see. In AI that shows up as waiting for the next OpenAI model before building, and while there are moments where waiting makes sense, he believes the PM should be the one who wanders out and comes back with a direction.

He illustrates the asymmetry with a beer with one of his engineering managers, who said the team was doing great — roadmap set, next three weeks of sprints planned, morale high — while the product side was confused and unsure where to go. Being the wanderer in the unknown, hunting for any signal to bring back, felt tough, but that is precisely the space where an AI PM earns their keep by pushing what the company thinks is possible.

When he admitted to leadership that he was not sure what to do next, Arize's chief product officer Aparna told him that is what building zero to one actually feels like: you do not know the right decision, but you know when you get there. The advice was to keep wandering and iterating until the drag of the product starts pulling you in a direction.

To shorten that wandering, Aman scales himself with AI. At Arize the team pulled Gong call transcripts and fed them into long-context models to surface the most common problems customers raise — signal from hundreds of conversations he could never attend in person, and now something that can be done with voice input as well. His advice is to find ways to be in many places at once so you can come back to the team with real signal through the noise.

Aman's closing advice for IC PMs came from a serial entrepreneur on Arize's board who, asked what a first-time PM should know, said simply: have fun. Curiosity and enjoyment make you iterate faster, and caring about what you are building is simply more fun than the alternative. Lenny adds that he keeps a battered Post-it on his desk that just says "have fun," and Aman points out the same reframe has been the core advice in public-speaking coaching.

His most-recommended book is Bill Bryson's A Short History of Nearly Everything, which he loves because it tells the history of science through the people who made the discoveries — Newton being unkind to colleagues, Darwin not being seen as a great scientist at the time — and because the scale of the universe puts our moment, including the AI era, into perspective as a cosmic blip.

For career questions he recommends Designing Your Life by one of the founders of IDEO and Stanford design professors, because it is practical rather than abstract. The exercise that helped him most was writing one page on what work means to him and one page on what life means to him, then finding the overlap — clarifying when he felt stuck and giving him a direction even without a route.

His life motto is Steve Jobs's line that your time is limited, so don't waste it living someone else's life. Aman says career pressure pushes people toward the next title, a startup, or something impressive, and the more useful practice has been reassessing his own goals rather than other people's projections of what his life should look like. In the same spirit, he answers every cold message he receives, as long as the sender shares their top three books, movies, podcasts or videos — a ritual he learned from John Dory at his college and now carries forward.