# Transcript: How to Build AI-First Products: Models, Memory, Mastery **Date:** March 17, 2026 · 10:30 PM **Session:** [How to Build AI-First Products: Models, Memory, Mastery](/sessions/2026-03-17/pp1162317-how-to-build-ai-first-products-models-memory-mastery) ## Summary A Microsoft product executive shares lessons from building AI-first products, emphasizing the shift from traditional engineering to managing AI agents. The speaker argues that products must now be built for both agents and humans, highlights the rapid commoditization of AI capabilities, and explores what matters in teams when intelligence becomes abundant: agency, judgment, and focus. ## Topics `ai-first products` · `agent management` · `product design` · `ai commoditization` · `team building` · `conversational ux` · `model evolution` · `human-agent collaboration` ## Key Takeaways 1. Building with AI agents feels more like management than traditional engineering—you're steering, nudging, giving instructions and feedback rather than writing code directly. 2. Products must now be designed for agents AND users, not just humans. This affects every decision from programming language choice to system architecture to API design. 3. In AI, 'today's magic is tomorrow's commodity'—capabilities that were groundbreaking three years ago (like DALL-E) are now routine, requiring constant attention to where models are evolving. 4. When intelligence is abundant and on-tap, the scarce resources become agency (initiative to pursue problems), judgment (which problems to pick), and density (maintaining focus amid infinite possibilities). 5. The 80-90% solution with humans filling gaps often delivers more value than waiting for full automation. ## Full Transcript Think about places where AI can tackle the big, hard 80-90% of the job, and then figure out the software engineering and the human in the loop folding around it. That gets you pretty much more of the value than trying to wait in eternity for the full automation. This is something I've been experiencing, and pretty much everybody in our teams have been experiencing. How many of you have been engineers in the past? How many of you use AI coding tools to make code, produce code? This is a new phenomenon. In the past, you had very clear divisions—there were parts of the team that were managing code, and then there were managers and schedules, and the dynamic was very different. But the biggest insight I see in our work is that with 80-90% code being generated by AI agents, it feels more like management. You're managing the agents, whether you use GitHub Copilot or Claude or whatever is your product of choice. You're running these instances, firing off these tasks like a boss. But a lot of the skills you now need are to steer them, nudge them, give correct instructions, give feedback and training. What does all of that sound like? Management. So I think there's more connection between management theory and business and the mindset that even builders will have to have right now. This has been really striking for us to realize: for the last 30 years, many of the productivity products have been built assuming there's a human at the other end of it. Whether you think of Illustrator or Word or Excel or any of these things, there's an assumption that at the other end of the PC, you have a human looking at it. And that assumption, of course, will continue in places, but the big opening is that you have to think about building your products for agents and users. That percolates through all sorts of choices and decisions you make. What programming language you pick—whichever is the one that agents want. What system architecture do you pick—whichever interfaces well with agents. What MCPs and skills you build—something that the agents can compose and recompose and build on top of. This has been one of the recent revelations for us: how do we pass every decision through the lens of which is the more agent-friendly option? And as you build products, assume you're building not just for humans, but for agents, which means a lot more scale too. Previously, when you had 10 humans looking at your product, you may have 10,000 or 100,000 agents wanting the same data. So how do you build for that performance and scale? I want to leave you with a few lessons and insights. The biggest lesson of all is that in the AI space, having lived through the internet and mobile and cloud, today's magic is tomorrow's commodity. I remember three years ago waking up one day and seeing DALL-E create a picture of an astronaut on a horse on the moon. How wonderful is that? And then now it's like, oh, it's just Tuesday, right? So to me, the meta lesson as a product builder has been: how do you have the stance to be very tuned into the first derivative? Where are the models changing and eating the product? Where are the model capabilities lacking some of the product capabilities—that's an advantage for you as a product builder. And what are these loops that create that compounding advantage? I leave you with one question. The other vantage point I have is running an organization that designs the teams that build AI products. One of the questions on top of our minds is: in a world where intelligence is on tap, that's no longer the scarce commodity. The coding products are better at pretty much everything than the most expert engineers in teams. So in a world where intelligence is no longer scarce, what do you select for in the teams that you build? These are a few things I think about. One, as cliche as it sounds and as ironic as it sounds because we talk about agents, it's really agency. Do you actually have the initiative to pursue a problem doggedly? Because now you have a team of agents that can go do anything. The other one I find in our team and across the board in the industry, especially if you spend any time on Twitter, is that there's an infinite range of things you can do, and every day there's a new model. In fact, as we speak, I'm sure there's a new model that drops somewhere. So how do you preserve your density—that's the main thing to maintain in your product. The question of which problem to pick and how important it is matters more than ever. What we're seeing is that it's like we are all kids in a candy store in terms of product building. All these ideas that would have been bottlenecked by engineering or product development are now all possible—of course, still compute constrained. So the question becomes: what is the resulting judgment you're going to impose on which problems to pick and how you're going to solve them? The rest of it, of course, the models and the AI products can easily get done for you. But this is true: this entire deck has been generated by a new copilot core agent. So I did exactly what I said—shared the Substack post, internal talks and insights from our lessons, and this is what it came up with. We call it the inception mode of product building, where we build this AI agent to build itself. So it generates the status reports for its project, it generates the text and updates. And of course, there's agentic code. I did ask it to put some style in it, and it generated this title. So I think we have five or six minutes for questions if there are any. --- *Source: stt · Language: en · Model: anthropic/claude-sonnet-4-5* [← Back to session](/sessions/2026-03-17/pp1162317-how-to-build-ai-first-products-models-memory-mastery)