LLenny's Podcast
← All episodes
Asha Sharma (CVP of AI Platform at Microsoft)28 August 2025

How 80,000 companies build with AI: products as organisms, the death of org charts, and why agents will outnumber employees by 2026

5Frameworks
14Insights

Frameworks in this episode

Innovation4 steps

Adapt the Model, Don't Build One

Post-training is the new pre-training — steer an off-the-shelf model to your outcome instead of pre-training your own

Read →
Strategy4 steps

Planning in Seasons

Replace rigid roadmaps with secular 'seasons', loose quarterly OKRs, and deliberate slack

Read →
Innovation5 steps

Product as Organism: the metabolic loop

Treat an AI product as a living system that ingests signals, tunes on rewards, and improves with every interaction

Read →
Strategy3 steps

The Enterprise AI Adoption Ladder

Three sequential stages that separate companies winning with AI from those spinning their wheels

Read →
Strategy3 steps

Win on the Platform, Not the Features

Durable products win on invisible infrastructure — reliability, reach, privacy — not on the feature list

Read →

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Hot Take· 4

Hot Take11:00

Build for the slope, not the snapshot

With something like 70,000 enterprise AI tools launched in a single year, Sharma argues it's impossible to know which one to bet on. Instead of getting locked into any single technology, companies should bet on a platform or app-server layer that lets them swap tools in and out. The mindset should optimize for where things are heading, not the current state.

  • Roughly 70,000 enterprise AI tools launched in one year
  • Don't get beholden to any single tool — the whole stack will change
  • Bet on a platform layer that lets you swap technologies in and out
  • Build for the slope (trajectory) rather than the snapshot (current state)

“something like 70,000 enterprise tools like in the AI space launched last year.”

Asha Sharma · 11:30

“Feel like you have to actually build for the slope instead of the snapshot of where you are.”

Asha Sharma · 11:30
#platform#tooling#enterprise#strategy
Hot Take20:30

The org chart becomes the work chart

Sharma predicts that as the marginal cost of good output approaches zero, demand for productivity explodes and you scale it with agents — both embedded (tools/software) and embodied (assignable workers). When that happens, the org chart becomes the 'work chart': tasks and throughput matter more than hierarchy, and organizations need far fewer layers. Humans still decide how AI is used, but teams organize outward and task-based rather than upward.

  • The marginal cost of a good output is approaching zero, driving exponential demand for output
  • You scale that with agents — embedded (tools) and embodied (assignable workers)
  • The org chart becomes the work chart; tasks and throughput beat hierarchy
  • Organizations need fewer layers; humans still decide how AI is applied

“we're approaching this world in which the marginal cost of um a good output is approaching zero.”

Asha Sharma · 21:00

“the work chart uh the org chart starts to become the work chart.”

Asha Sharma · 21:30
#agents#future-of-work#org-design#productivity
Hot Take39:30

Satya's lesson: optimism is a renewable resource

Sharma's single biggest leadership lesson from Satya Nadella is that optimism is a renewable resource. Despite 50 years of reasons not to succeed, his ability to generate energy and clarity and renew everyone's commitment to the mission — every day, in a fiercely competitive talent market — is what she finds most remarkable. She frames it as 'vibes': you have to follow a mission bigger than yourself.

  • Satya Nadella's standout trait is treating optimism as a renewable resource
  • He renews everyone's dedication to the mission daily through energy and clarity
  • The growth mindset is real, but generating energy in a competitive talent market is rarer
  • Sharma frames it as 'vibes' — a sense of duty toward a mission bigger than yourself

“optimism is a renewable resource.”

Asha Sharma · 39:30

“the only people that'll remember you working late are your kids.”

Lenny Rachitsky · 41:30
#leadership#satya-nadella#culture#motivation
Hot Take48:00

Bet on a model system, not one model to rule them all

Sharma is firmly in the 'model system' camp and believes in model diversity rather than one model to rule them all. Different models fit different use cases — Sonnet 4 is great for some, GPT-5 for others — and some tasks care about latency while others tolerate thinking time. She distinguishes this from an 'ensemble of models,' which she defines as multiple models you fine-tune and deploy independently.

  • Sharma believes in model diversity, not one model to rule them all
  • Different models suit different use cases (e.g. Sonnet 4 vs GPT-5) and latency needs
  • An 'ensemble' is multiple models fine-tuned and deployed independently
  • Terminology is still being invented as the field moves fast

“I'm much more in the model system camp. Like I believe in uh model diversity.”

Asha Sharma · 48:00

“I'm much more in the like model system uh rather than one model to rule them all.”

Asha Sharma · 48:30
#model-diversity#ensemble#architecture#llms

Explainer· 4

Explainer04:30

From product as artifact to product as organism

Asha Sharma argues products are no longer static artifacts you ship and dashboard, but living systems that improve with every interaction. Because models can now tool-call and act, the winning skill is a product team's 'metabolism' — ingesting data, digesting a rewards model, and tuning to outcomes like price, performance, or quality. She calls these self-improving products the new IP of every company.

  • Products shift from static artifacts to systems that get better with more interactions
  • The new KPI is a team's 'metabolism': ingest data, digest a rewards model, produce an outcome
  • You tune models toward specific outcomes — price, performance, or quality
  • Self-improving products become the core IP of a company

“all of a sudden these are these living organisms that just get better with the more interactions that happen.”

Asha Sharma · 06:00

“products that think and live and learn, which is kind of exciting.”

Asha Sharma · 06:00
#ai-products#product-strategy#post-training#ip
Explainer06:30

Why post-training is the new pre-training

Sharma explains that once base models are strong enough, the economic leverage moves from expensive pre-training to fine-tuning and reinforcement learning on your own data. She cites a Nathan Lambert study finding that past ~30 billion parameters the capex to train a model from scratch stops making sense, so you optimize on the loop instead. She predicts as much money will eventually be spent on post-training as pre-training.

  • Past ~30B parameters, training a model from scratch stops making economic sense
  • Leverage shifts to fine-tuning and reinforcement learning on proprietary or synthetic data
  • She predicts post-training spend will eventually match or exceed pre-training spend
  • Adapting an off-the-shelf model beats building your own for most companies

“you know once a model hits 30 billion parameters the capex to actually train a model”

Asha Sharma · 06:30

“I believe we will see you know just as much money spent on post-training uh as we will on pre-training and in the future more…”

Asha Sharma · 45:00
#post-training#fine-tuning#reinforcement-learning#economics
Explainer12:00

The rise of the full-stack builder: the loop, not the lane

Sharma notes that a normal org needs ~10 steps and 5-7 functions across 6-7 layers to ship a product — around 500 touch points, which is untenable when hundreds of new models and technologies arrive weekly. The answer is the full-stack 'polymath' builder who owns the whole loop. Whatever your function, you must obsess over cost, rewards design, and UX for both people and agents.

  • A normal org needs ~10 steps, 5-7 functions, and 6-7 layers — around 500 touch points to ship
  • That process is insufficient when 500 new models/technologies appear weekly
  • Full-stack 'polymath' builders own the whole loop and gain velocity and throughput
  • It's about the loop, not the lane — every function must obsess over cost, rewards, and UX

“when there are 500 models available a week or 500 new technologies, that is just insufficient.”

Asha Sharma · 12:30

“it's all about the loop not not the lane here.”

Asha Sharma · 13:30
#full-stack#roles#polymath#product-teams
Explainer16:30

From GUIs to code-native interfaces: composability over canvas

Sharma sees products moving from graphical UIs to code-native interfaces, the same historical pattern as databases going to SQL and cloud consoles going to Terraform. Because a stream of text connects better with LLMs, the future favors composability over the canvas. Product makers should rewire their mindset away from obsessing over UI and toward how something composes and how an agent can read it.

  • Interfaces follow history: desktop-to-SQL, cloud consoles-to-Terraform, now GUIs-to-code-native
  • A stream of text connects better with LLMs, favoring composability over the canvas
  • Builders overspend on UI and underspend on how things compose and scale
  • Design for how agents read and collaborate, not just how humans click

“databases kind of went from the desktop kind of down into SQL. I think cloud was all about consoles and now it's about Terraform.”

Asha Sharma · 16:30

“there's a bunch of trends that are kind of working in the favor for like the future of products being about composability and not the…”

Asha Sharma · 17:00
#interfaces#agents#composability#product-design

Story· 2

Story15:00

How Microsoft's Dragon jumped to 83% acceptance with expert labeling

Sharma gives a concrete example of the fine-tuning loop with Dragon, Microsoft's AI product for physicians. Moving from synthetic fine-tuning to annotating 600,000 physician-patient interactions by experts, then continuously optimizing, drove the character acceptance rate from 30-60% up to about 83%. Crucially, this was done by a small cross-functional group, not a large organization.

  • Dragon is an AI product for physicians
  • Switching from synthetic data to 600,000 expert-annotated interactions transformed results
  • Character acceptance rate rose from roughly 30-60% to about 83%
  • A small cross-functional team, not a big org, drove the improvement by iterating in the loop

“we annotated 600,000 patient um physician interactions by experts and actually fed that into the model”

Asha Sharma · 15:30

“I think we're sitting between 30 and 60 character acceptance rate depending on the run to something like 83 3%.”

Asha Sharma · 15:30
#healthcare-ai#fine-tuning#data-labeling#case-study
Story37:00

Platforms win on invisible infrastructure, not features

Sharma's biggest career lesson is that platforms win on the invisible work, not the pixels. WhatsApp didn't win on stickers or dark mode — it won on the phone book, reliability, speed, and end-to-end privacy. Instacart is really a billion items updating 3,000 times a minute. She wishes she'd known sooner that data residency, availability, reliability, and the right tool selection are what actually matter.

  • WhatsApp won on the phone book, reliability, speed, and end-to-end privacy — not features
  • Instacart is a billion items updating 3,000 times every minute
  • Porch's breakthrough was a matching engine, not the home-report features
  • For AI platforms, data residency, availability, and reliability are the real moat

“WhatsApp didn't win because it had stickers or stories or dark mode.”

Asha Sharma · 37:30

“It's the data residency so the hospital in Germany that's fine-tuning a model can do so in confidence and the data isn't going to leave…”

Asha Sharma · 38:30
#platforms#infrastructure#whatsapp#reliability

Tool· 1

Tool49:30

Asha Sharma's most-recommended books

In the lightning round, Sharma names the books she recommends most. At work it's Thinking Machine, which she frames as treating the cause not the symptoms — like solving traffic through walkability and mobility rather than speed bumps. Personally, her Instacart CMO recommended Tomorrow and Tomorrow and Tomorrow, which she has reread multiple years running.

  • Work recommendation: Thinking Machine — treat the cause, not the symptoms
  • Systems example: solve traffic via walkability and mobility, not speed bumps
  • Personal favorite: Tomorrow and Tomorrow and Tomorrow, recommended by Instacart's CMO
  • She has reread the latter several years in a row

“It's probably Thinking Machine. Uh so it's all about treating the the cause not the symptoms.”

Asha Sharma · 49:30

“recommended to me Tomorrow and Tomorrow and Tomorrow, and I read it like last month and last year and the year before because I love…”

Asha Sharma · 50:00
#books#recommendations#systems-thinking

Takeaway· 3

Takeaway09:30

What separates companies that win with AI from those that don't

Sharma describes the pattern behind successful AI adopters: everyone becomes AI-fluent, they apply AI to an existing process to feel real P&L impact, then use it to inflect growth. Companies fail when they do 'AI for AI's sake' — many scattered projects with no blueprint, no measurement, observability, or evals, and no treatment of AI as a real investment.

  • Winners make everyone AI-fluent first, so no one fears the tools
  • They apply AI to an existing process to feel measurable P&L impact before scaling
  • Then they use AI to inflect growth via LTV, retention, or new categories
  • Losers run scattered projects with no blueprint, measurement, observability, or evals

“one is they are embracing AI and everybody becomes AI fluent.”

Asha Sharma · 09:30

“where companies fail is that they're doing AI for AI sake.”

Asha Sharma · 11:00
#ai-adoption#enterprise#measurement#strategy
Takeaway30:30

How to plan a roadmap in a world that changes weekly: seasons

Asked how anyone sets a roadmap when GPT-5 can drop overnight, Sharma describes planning by 'seasons' rather than fixed six-month semesters. A season is defined by secular changes in the industry or from customers and can last three months or a year; it grounds everyone on a shared north-star. Beneath that sit loose quarterly OKRs, four-to-six-week squad goals, and deliberate slack in the system for both the unplanned and the slope.

  • Fixed six-month semesters are too rigid when the frontier shifts weekly
  • A 'season' is defined by secular industry or customer changes and can last 3 months to a year
  • Seasons align everyone on a shared ethos and north-star metric
  • Loose quarterly OKRs, 4-6 week squad goals, and slack for the slope sit underneath

“think about it as what season are we in?”

Asha Sharma · 31:00

“we try to leave slack in the system not just for the unplanned but for the the slope.”

Asha Sharma · 32:30
#planning#roadmap#okrs#strategy
Takeaway53:00

Maximizing option value beats minimizing regret

Sharma shares how her life motto evolved from a minimize-regret framework to maximizing option value once she had a family. The shift gave new weight to family, health, trust, and relationships because rest and health compound into the future rather than trading off against extra work hours. Her aim at 70 is not to count regrets looking back, but to have accumulated skills, trust, and relationships that create adventures ahead.

  • Her motto evolved from minimize-regret to maximizing option value
  • The shift gave more weight to family, health, trust, and relationships
  • Rest and health compound into the future rather than trading off against work
  • The goal is looking forward to adventures, not counting regrets looking back

“it was all about maximizing kind of option value and it just gave the things that I naturally cared about like family and health and…”

Asha Sharma · 53:30

“when I'm 70 it's not about what do I look back on in my life and count the number of regrets. It's really about like…”

Asha Sharma · 54:00
#life-motto#option-value#decision-making#wellbeing