How Airtable restructured for AI: lessons from Howie Liu

Executive overview

Most software companies bolt AI onto existing structures and fall behind. Airtable split its engineering org into a fast-thinking team (shipping new AI capabilities weekly) and a slow-thinking team (handling infrastructure and enterprise scale), then rebuilt the entire product experience around an AI agent as the default interface.

The bigger lesson: every software company needs to refound itself for AI, not just add features. The question is whether your existing building blocks give you an unfair advantage — or whether you should sell and start fresh.

CEOs who use AI hourly, stay close to product details, and refuse to delegate taste are the ones finding new product-market fit.

Refounding Airtable for AI

• Original mission was always to democratise software creation — AI app-building is a different means to the same end
• Pure code-gen tools produce unreliable business apps as complexity grows; Airtable's no-code primitives act as a reliable DSL the agent assembles instead
• The agent (Omni) is now the default interface; the app itself has become an artifact the agent manipulates
• New homepage: just describe what you want to build — fully PLG and experiential, not sales-led
• Airtable also added custom code-gen to fill final-mile gaps where no-code primitives aren't enough
• Before committing to the AI pivot: ask honestly whether your legacy assets help or hurt versus starting from scratch; if they hurt, find a buyer

The fast/slow-thinking reorg

• Previous structure: feature-area teams (search, mobile, etc.) optimised incrementally, never made step-change bets
• Intermediate reorg into business units was better but still not fast enough for AI pace
• Current structure: AI platform group ("fast thinking") ships jaw-dropping capabilities on a near-weekly cadence; a slow-thinking group handles data infrastructure, scale, and enterprise durability
• Fast and slow complement each other: fast creates top-of-funnel excitement; slow turns adoption seeds into large deployments
• Fast-thinking team profile: entrepreneurial, full-stack thinkers comfortable with design ambiguity
• Slow-thinking prevents the "wide top of funnel, no retention" failure pattern seen in many AI-native startups

The ICCO: CEOs as individual contributors

• AI demands continuous product-market refinding, not a one-time form-factor change like mobile or cloud
• To be chief taste maker you must participate in making the soup, not just taste the final dish
• Howie cut standing one-on-ones; replaced them with urgency-driven meetings seeded by real alpha
• Barbell social model: brief topical check-ins plus higher-quality in-person catch-ups every month or two
• He tracks his own inference costs as a proxy for engagement; aims to be the highest-cost AI user at Airtable
• Key signal: does the CEO use Claude or ChatGPT daily? Howie's answer: multiple times an hour

Helping teams operate at AI pace

• Stress the value of play — curiosity-driven exploration, not task completion
• Lead by example: share Replit prototypes instead of docs; share deep-research links instead of memos
• Give anyone permission to block a full day or week just to try AI products with no other agenda
• Prototypes over decks: functional demos expose what words in a PRD cannot — latency, edge cases, feel
• Shift from deterministic roadmaps to an experimentation-and-iteration model; shipping teaches you what the product should be
• Invent small personal side projects to force genuine engagement with tools — an hour produces real insight

Skills every PM, engineer, and designer now needs

• The winner in each function is the hybrid: designer who understands tool-calling, PM who can prototype, engineer who thinks about product and business
• Minimum viable fluency: every role needs a baseline in the other two disciplines — not depth, just "dangerous enough"
• PM role evolves into hybrid PM-prototyper with design sensibility; spec-writing-only PMs are at risk
• AI design is primarily interaction design — what happens after you fire a prompt is the real UX
• The gap between idea and working software has collapsed; everyone can learn to build, especially with AI tutors available 24/7
• Google's original APM spec required technical and design literacy; Stripe engineers often hold the product DRI — this is now table stakes

On evals

• Don't start with evals for genuinely novel capabilities — start with vibes and open-ended exploration
• Evals are most useful once you've converged on a use-case cluster and want to measure iterative improvements
• Early discovery requires divergence first; evals lock you into convergence too soon
• At scale you can A/B test everything; at earlier stages, qualitative tasting is faster and more informative

Founder mode and the counterintuitive lesson

• Standard advice — hire experienced operators, create functional swim lanes, step away from details — optimises each lane but destroys holistic, step-change product thinking
• The real failure mode: everyone locally optimises, nobody reaches the global breakthrough
• Founder mode is not micromanagement; it is refusing to let integrative decisions happen by default without the CEO's involvement
• Meta-lesson: don't just follow advice — inspect the chain of thought behind it, then apply the reasoning to your own context
• What Howie would tell his 2013 self: never step away from the product details you love; that passion is the engine