How we restructured Airtable's entire org for AI | Howie Liu (co-founder and CEO)

YouTube URL: https://www.youtube.com/watch?v=GT0jtVjRy2E&t=17s Host(s): Lenny Rachitsky (Lenny’s Podcast) Guest(s): Howie Liu (Co-founder and CEO, Airtable)

Podcast/Video Overview

Overall Summary

This conversation with Airtable’s CEO, Howie Liu, is a playbook for refounding a mature software company in the AI era. The core question: if you started your company today with the same mission, what would an AI‑native approach look like? If you can’t answer that credibly, you may need to rethink your strategy—or even your company. Howie shares how Airtable reorganized into “fast thinking” and “slow thinking” groups, rebuilt the product experience around an AI agent, and changed the CEO role into an ICEO—an individual contributor CEO who ships, prototypes, and tastes the product daily.

You’ll learn how to set a culture of speed, when to use evals versus vibes, and why roles across PM, engineering, and design must merge. You’ll also see practical steps for adopting AI: cancel meetings to play with tools, ship prototypes over decks, and obsess over experiential value. This is a candid, actionable masterclass on AI-era product leadership.

References

• PLG (Product-led Growth): A go-to-market motion where the product itself drives acquisition, activation, and expansion.
• EPD: Engineering, Product, and Design organization—the core product development triad.
• ICEO: “Individual Contributor CEO.” A CEO who actively prototypes, builds, and leads by doing.
• Agentic/Agent: AI systems that plan and take multi-step actions, often using tools and context, not just answering single prompts.
• Tool Calling: Letting an AI call external tools or functions (APIs) to complete tasks.
• DSL (Domain-Specific Language): A specialized set of commands or building blocks that make a class of tasks easier than raw code.
• LLM (Large Language Model): Foundation AI models like GPT-4/5 or Claude that generate and reason over text and code.
• Context Window: The maximum amount of text an LLM can “hold in mind” during a request.
• Map-Reduce (LLM context): Breaking large inputs into chunks (map), processing each with the model, then aggregating results (reduce).
• Inference Cost: The compute cost paid to run model requests (tokens and model time).
• Deep Research (OpenAI): An automated, long‑running research mode that finds and synthesizes information from the web.
• Fast vs Slow Thinking: From Daniel Kahneman. Fast = rapid, intuitive execution. Slow = deliberate, large, high‑stakes bets.
• Codegen: AI-generated code to implement features or entire apps.
• NUX: New user experience—the first-time user journey.
• Eval: A repeatable test or benchmark to measure AI output quality.
• “Vibes”: Early, qualitative sense-making before formal metrics.

Key Topics Covered

CEOs as ICs again: The rise of the ICEO

Summary: Howie argues CEOs must get back into the details. In AI, you cannot taste the soup from 10,000 feet. He prototypes, plays with models daily, and is Airtable’s “chief tastemaker.” He even tracks his own inference costs and is okay “overspending” on compute to learn faster. This is about speed, texture, and firsthand knowledge. Delegation alone won’t cut it in a rapidly shifting landscape.

• Main Point: In the AI era, CEOs must be builders again.
• Core Argument: AI’s pace and ambiguity require leaders to feel the product, not just review slides.
• Quotes: “I am the number one most expensive inference cost user of Airtable AI.” “You have to be in the details.”

Fast thinking vs slow thinking org design

Summary: Airtable split EPD into two groups. Fast thinking ships jaw-dropping AI capabilities almost weekly. Slow thinking invests in deep infrastructure like HyperDB to power enterprise scale. The two sides feed each other: fast creates excitement and adoption; slow enables durability, security, and expansion. This dual-track model balances invention and stability.

• Main Point: Separate rapid AI invention from deliberate platform building.
• Core Argument: Weekly AI shipping needs a different operating cadence than multi-quarter infra bets.
• Quotes: “We want to ship a bunch of new capabilities on a near weekly basis.” “You need fast and slow thinking and the common sense to operate.”

Refound your product for AI: The clean-slate test

Summary: Howie’s test: if you were founding the company today with the same mission, what would an AI-native strategy be? If your legacy assets don’t help, you may be better off selling. Airtable still has an edge: robust no-code primitives the agent can assemble, with codegen for bespoke needs. That mix keeps things reliable and editable by non‑developers.

• Main Point: Rebuild strategy from first principles and be honest about your assets.
• Core Argument: AI changes form factors and business models; legacy must prove it’s an advantage.
• Quotes: “You have to take a clean slate approach.” “If you can’t, you should find a buyer.”

Prototypes over decks: Ship, learn, iterate

Summary: Replace plans and decks with real prototypes. AI value is interactive and experiential. You learn more by testing real prompts than reading PRDs. Airtable runs weekly AI sprints and expects demos that handle open-ended inputs. This reveals latency, UX friction, and “wow” moments that slides never capture.

• Main Point: Experience beats documentation in AI product building.
• Core Argument: Only working prototypes expose real usability, performance, and model behavior.
• Quotes: “The proof is in the pudding.” “I want to see actual interactive demos.”

Role convergence: PM, design, and engineering become hybrid

Summary: The best AI builders are polymaths. Designers who understand tool calling. PMs who prototype. Engineers who lead product decisions. Each role should be “minimally good” at the other two. This breaks dependencies, speeds cycles, and raises product taste across the team.

• Main Point: The minimum bar is hybrid skill across PM, design, and engineering.
• Core Argument: AI reduces friction to build; people who can do more, ship more.
• Quotes: “You need to be like decently good at all three.” “There’s a strong advantage to hybrid unicorn types.”

PLG returns: AI value must be experienced

Summary: PLG is essential in AI because value is felt, not told. Sales-led motions work (Palantir, Harvey), but mass adoption comes from frictionless trials. Make onboarding fast. Let people try the thing now. Airtable’s homepage now starts with, “What do you want to build?” Then Omni, their agent, builds with you.

• Main Point: Let users feel AI value immediately.
• Core Argument: The fastest-growing AI products are experiential and self-serve.
• Quotes: “ChatGPT is arguably the most successful PLG product of all time.” “We made the entire product experience AI‑centric.”

Evals vs vibes: When to measure and when to explore

Summary: Don’t start with evals. Start with vibes to find the form factor and promising use cases. Once the shape is clear, use evals to iterate and harden. This avoids over-optimizing the wrong thing early and gives structure once you know what “good” looks like.

• Main Point: Explore first; measure second.
• Core Argument: Evals are powerful after you’ve converged on a viable product direction.
• Quotes: “You should actually not start with evals and start with vibes.” “Evals are useful once you’ve converged.”

Culture of play and curiosity

Summary: Howie encourages teams to cancel meetings for a day—or a week—and just play with AI tools. Try Runway, Cursor, HeyGen, Deep Research, NotebookLM. Make a weird video. Build a small app. The goal is hands-on fluency and a personal library of patterns. This is the fastest path to taste and innovation.

• Main Point: Play is a productive use of time in AI.
• Core Argument: Curiosity produces pattern recognition, instincts, and better product ideas.
• Quotes: “Go do it, period.” “AI is something you have to play with.”

Agents + no-code + codegen: Airtable’s new product shape

Summary: Airtable’s agent, Omni, is now the default way you build. The agent assembles reliable no‑code primitives (views, layouts, automations) and uses codegen for the final mile. That keeps apps stable, secure, and editable, while allowing bespoke features. It avoids context collapse and black-box codebases.

• Main Point: Combine agentic assembly with reliable primitives and targeted codegen.
• Core Argument: This hybrid beats pure codegen for business apps where reliability matters.
• Quotes: “Our agent can use our Lego kit as a more expressive DSL.” “We made the entire product experience AI‑centric.”

Key Themes and Insights

Operate with urgent curiosity

Play with AI tools weekly. Build tiny projects that scratch itches. Taste the soup yourself. This builds instincts faster than reading.

• “If you want to cancel all your meetings for a day or a week…go do it.”
• “AI is something you have to play with.”

Dual-track orgs win in AI

Create fast and slow thinking lanes. Ship weekly magic while building long‑term platform bets that scale.

• “We now have these two separate parts of the company.”
• “They actually complement each other very well.”

Roles must converge

Raise the baseline for PM, design, and engineering. Everyone should be competent at the other two.

• “You need to be like decently good at all three.”
• “There’s a strong advantage to hybrid unicorn types.”

Prototypes beat documents

Prioritize working demos, not decks. Learn from latency, failures, and real prompts.

• “The proof is in the pudding.”
• “It’s hard to get that feel with anything but a functional prototype.”

Explore with vibes, refine with evals

Start broad to find the right form factor. Add evals after convergence to harden and scale.

• “You should actually not start with evals and start with vibes.”
• “Evals are useful once you’ve converged.”

Refound your company for AI

Do the clean‑slate test. If your legacy isn’t an advantage, rethink or divest.

• “If you can’t, then you should find a buyer.”
• “Take a clean slate approach.”

Experiential PLG is back

Let users feel value immediately. Make onboarding agentic and fast.

• “ChatGPT…700 million weekly active users.”
• “We made the entire product experience AI-centric.”

Leaders must build

The ICEO sets taste, speed, and standards by doing the work.

• “I am the number one most expensive inference cost user of Airtable AI.”
• “You have to be in the details.”

Actionable Advice and Takeaways

Immediate Actions You Can Take:

• Block one day to try 5–7 AI tools end‑to‑end. Build a tiny project that is personally useful.
• Replace one deck this week with a clickable prototype. Gather feedback from real usage.
• Instrument your AI usage. Track inference spend to learn faster (and spot ROI).
• Add an “agent first” entry point in your product or internal tools for a narrow task.

Long-term Strategies:

• Reorg into fast and slow thinking lanes. Give each clear charters and cadences.
• Raise the hybrid bar in hiring and growth. PMs prototype. Designers understand tool calling. Engineers lead product tradeoffs.
• Build a hybrid product stack: agent + reliable primitives + targeted codegen for final mile.
• Treat PLG as an experiential funnel. Make the first minute magical and useful.

Questions for Reflection:

• If I refounded my product today with AI, what would change immediately?
• Which bets belong in fast track versus slow track in my org?
• Where am I over‑planning and under‑prototyping?
• Do my PMs, designers, and engineers meet the new hybrid baseline?

Noteworthy Observations and Unique Perspective

The cost of learning is cheap compared to the value of insight

• Spending hundreds on inference can replace weeks of human analysis.
• “Hundreds of dollars…is trivial compared to the potential strategic value.”

PLG for AI is about feeling, not telling

• You must let users try the thing right away.
• “It’s so much more powerful when you just open up the doors.”

Legacy advantage must be proven, not assumed

• Honest audit of whether your assets help or hurt in AI.
• “If you can’t…you should find a buyer.”

Collapse role silos beyond EPD

• Sales should demo like SEs. Marketers should run their own creative and performance loops.
• “Everybody needs to become more full stack to do the thing.”

CEO joy returns by getting back in the build

• Founder mode is energizing and effective in AI.
• “It’s just much more invigorating to get to play that role.”

Companies, Tool and Entities Mentioned

• Airtable (Omni, HyperDB, Field Agents)
• OpenAI (ChatGPT, GPT-5, Deep Research)
• Anthropic (Claude 4, MCP)
• Cursor (Composer)
• Runway
• HeyGen
• NotebookLM
• GitHub Copilot
• Figma
• Palantir
• Harvey
• Airbnb
• Stripe
• Salesforce
• ServiceNow
• CB Insights
• Flexport
• All-In Podcast
• DX (Developer Intelligence Platform)
• LucidLink

Final Thoughts

AI is forcing every software company to refound itself. Airtable’s playbook is simple but hard: lead as an ICEO, split into fast and slow tracks, ship prototypes, and rebuild the product as agent‑first. Collapse silos. Raise the hybrid bar. Explore with vibes, then measure with evals. If you act with urgency and curiosity, you can find new product-market fit again. Your next step: cancel a few meetings this week and go build something real with AI—then ship it.

Linkedin Ideas

The ICEO: Why CEOs must build again

• Core Argument: In AI, leaders need firsthand product taste to make good bets.
• Key Point: Replace reviews and decks with shipping prototypes and daily AI use.
• Quotes: “You have to be in the details.” “I’m the number one inference cost user of Airtable AI.”

Fast vs Slow Thinking Orgs: How to ship weekly and scale safely

• Core Argument: Separate rapid AI feature velocity from deliberate platform work.
• Key Point: Fast generates demand; slow ensures durability and enterprise readiness.
• Quotes: “We ship on a near weekly basis.” “You need fast and slow thinking.”

Prototypes over Decks: The new rule of AI product management

• Core Argument: Only interactive demos reveal real value, latency, and UX.
• Key Point: Replace one deck this week with a working prototype.
• Quotes: “The proof is in the pudding.” “AI is something you have to play with.”

Evals vs Vibes: When to measure and when to explore

• Core Argument: Start with vibes to find form; add evals to scale and harden.
• Key Point: Don’t over‑optimize the wrong shape early.
• Quotes: “Not start with evals—start with vibes.” “Evals are useful once you’ve converged.”

Refounding for AI: The clean-slate test every founder must run

• Core Argument: If your legacy assets aren’t an advantage, rethink or divest.
• Key Point: Build agent‑first experiences and hybrid stacks (primitives + codegen).
• Quotes: “Take a clean slate approach.” “If you can’t, you should find a buyer.”

Blog Ideas

How Airtable Refounded Itself for AI: A Practical Org and Product Playbook

• Core Argument(s): AI demands new org structure, role convergence, and agent‑first UX.
• Key Themes to explore: ICEO, fast/slow orgs, prototypes, PLG, evals vs vibes.
• Key Point (s): Lead by building; ship weekly; make the product experiential.
• Quotes: “We made the entire product experience AI-centric.” “Go cancel your meetings and play.”

The Hybrid Baseline: Why PMs, Designers, and Engineers Must Overlap in AI

• Core Argument(s): Polymaths ship faster and make better product tradeoffs.
• Key Themes to explore: Role convergence, tool calling, prototyping fluency.
• Key Point (s): Set a new baseline—be “minimally good” at all three roles.
• Quotes: “You need to be like decently good at all three.”

Evals vs Vibes: A Founder’s Guide to Finding AI Product-Market Fit

• Core Argument(s): Explore first to find the shape; measure second to scale.
• Key Themes to explore: Early exploration, convergence, defining “good.”
• Key Point (s): Don’t let metrics constrain discovery too soon.
• Quotes: “Not start with evals—start with vibes.”

Agent + Primitives + Codegen: A Better Way to Build Business Apps

• Core Argument(s): Hybrid stacks beat pure codegen for reliability and editability.
• Key Themes to explore: DSL-like primitives, context windows, final-mile codegen.
• Key Point (s): Keep users in control with transparent, editable building blocks.
• Quotes: “Use our Lego kit as a more expressive DSL.”

Fast Thinking, Slow Thinking: Designing AI Teams That Actually Ship

• Core Argument(s): Dual-track execution balances innovation and stability.
• Key Themes to explore: Cadence, governance, handoffs, metrics.
• Key Point (s): Fast creates demand; slow scales it.
• Quotes: “They actually complement each other very well.”