How 80,000 companies build with AI: products as organisms, the death of org charts, and why agents will outnumber employees by 2026 | Asha Sharma (CVP of AI Platform at Microsoft)

The episode opens with a cold-open montage of four ideas that the rest of the conversation develops: an "agentic society" that has barely begun to take shape, a shift from product as artifact to product as organism, the difficulty of roadmapping when a new model such as GPT-5 can land at any time, and a planning habit of asking which "season" the industry is in. These are teasers, and the concrete reasoning behind each comes later.

When all of that happens, the org chart starts to become the work chart.

Asha Sharma's compressed argument in the teaser runs as follows. The marginal cost of good output is heading toward zero, so demand for productivity and output will grow exponentially, and agents are how organizations scale to meet it. A consequence she draws is that organizations will not need as many management layers.

In the long introductory block, host Lenny Rachitsky introduces Sharma and previews the topics: product as organism, graphical interfaces giving way to code-native interfaces, post-training as the new pre-training, an agentic society, what it takes to be a successful builder, and a leadership lesson from Satya (Nadella, as context implies, though the transcript only says "Satya"). The transcript describes her as "Chief Vice President of Product" for Microsoft's AI platform (the title in the episode name is CVP, so this is likely a transcription slip) and as a former Meta "VPR product" (probably a VP of product; the transcript says VPR). The same block contains Lenny's newsletter promotion and two sponsor reads, for Enterpret (a customer-intelligence platform) and DX (a developer-intelligence platform). These are advertisements and carry no analytical content.

Lenny asks Sharma to explain the idea that products are moving from artifact to organism. Her account starts from what changed at the model layer since she joined Microsoft. After OpenAI and the large foundation models arrived, there was an explosion of proprietary and open frontier models. They became more efficient, some began to show domain-level expertise, and more recently they can call tools and functions and take action. In her view this opens the way for a new type of product that is beginning to see some success.

The contrast she draws is with the static artifact: come up with an insight, solve a problem, ship it, improve it a little and watch a dashboard. In the organism model the key measure becomes the "metabolism" of the product team, meaning its ability to ingest data, digest the rewards model and produce an outcome. Because the models are effective, teams can tune them toward outcomes such as price, performance or quality. She calls such products living organisms that improve with more interactions, and says she sees this as the new IP of every company. The claim is framed as her view of where things are going rather than a measured result, and she offers no data in this passage to show how widespread it already is.

All of a sudden, these are these living organisms that just get better with the more interactions that happen.

Lenny connects the idea to Michael Truell's point (from an earlier episode) that the Cursor moat is the data captured when users accept or reject suggestions, and asks whether Sharma means proprietary usage data or something broader. Sharma answers by pointing to the rise of post-training. She cites a study by Nathan Lambert, as she describes it, across the top leaderboards: once a model reaches 30 billion parameters, the capital expense of pre-training with billions of tokens stops making economic sense and one can instead optimize on the loop. This is her summary of a study she did not author, and the passage does not give the study's details or limits.

On the data question she says that using your own data is, in many ways, the best route, but that data can also be generated synthetically. The harder work lies elsewhere: designing the rewards, rolling the system out, running rigorous A/B tests, and finding the job to be done or use case where it fits best. Only then does the process yield data to learn from. She adds an empirical observation, that she has never seen a single loop serve a product. What she sees instead is multiple tracks running in parallel, like assembly lines.

Asked whether the thesis applies only to model companies or also to SaaS products and tools, Sharma says software as a primitive is changing: the artifact inside a product is now a model alongside the ordinary software components, so she expects software products to become model-forward. Her answer is short and offers no examples or counter-cases.

Lenny reports a conversation with Nick Turley, head of ChatGPT. According to Lenny, Turley said ChatGPT is the same product when GPT-5 arrives, because the model tells the team what to do in the product. Lenny extends this: if the model can drive the product, why couldn't GPT-5 build its own interface as you use it, with the UX evolving without product teams doing anything? He notes that Canvas is a partial step in that direction. Sharma says she believes 100% that this is where the world is going, with each person's experience looking and feeling different from anyone else's, and says it would be possible on the fly in the future. She adds that the experience would also differ for agents, power users and new users.