Building AI-Driven Workflows at Airtable: Play, Prove, Productionise

 “Everything we’ve known is changing right now. Do not rely on what you know. Step back and figure out the next set of best practices.”  Jimmy Hillis, VP and Head of Engineering, Airtable

Three Waves, One Inflection

Jimmy doesn’t open with hype. He opens with perspective. For him, AI isn’t an add-on or an accessory; it’s a fundamental shift in how products are conceived, built, and refined. At Airtable, where the mission is to empower non-technical teams to build tools that fit their unique workflows, AI has to simplify, not complicate. It has to help people express intent in their own words and see tangible value fast.

He places this shift alongside two earlier revolutions. The first was the cloud, which turned static software into living, continuously improving systems. The second was mobile, which forced everyone to rethink how, where, and when people interact with products. Large language models, Jimmy argues, mark a third transformation, one that challenges every assumption about interface, logic, and process.

Five years ago, most engineers and designers would have laughed at the idea that a chatbox could become the dominant interface. Today, that’s the norm. The takeaway is humility. When new paradigms arrive, the only productive response is to experiment and learn fast.

What Airtable Built and Why It Matters

Jimmy has spent nearly six years watching Airtable scale from a 100-person company to over 750 employees. Around a third of the product organisation now builds AI-driven features not as gimmicks, but as ways to deepen accessibility and speed.

AI Fields allow users to write prompts enriched with their existing Airtable data, generating insights, content, and responses directly inside their workflows. Co-builder lets anyone describe what they want to build in natural language, “I’m in marketing, running global campaigns across time zones,” and generates the right schema, relationships, and UI in minutes. Assistant takes it further, enabling users to ask questions, analyse results, or generate summaries conversationally.

Each of these innovations strips away technical friction and expands who can build. For Airtable’s customers, AI isn’t replacing creativity; it’s amplifying it.

Play Before You Plan

Jimmy’s most emphatic lesson is that play must come before planning. Traditional product processes start with documents, research, and design, then move to building. With AI, that order no longer works. The models behave in unpredictable ways, and preconceptions often collapse on contact.

He encourages teams to start by playing with the models directly. Sit engineers, designers, and PMs together. Feed in real data. Write messy prompts. See what’s possible before defining what’s desirable. That spirit of play reshaped Co-builder’s origin story. Initially, the team proposed a cautious plan: generate recommended templates for different teams. Safe, sensible, and completely uninspired. Only when they experimented hands-on did they realise that LLMs already understood database relationships well enough to build full applications from scratch. That insight, Jimmy says, didn’t come from a whiteboard; it came from curiosity.

Play isn’t just a process shift; it’s a mindset shift. Jimmy urges every team member to find their “aha moment” with AI. That moment when a tool saves hours of work or unblocks a persistent problem turns abstract technology into personal conviction. Once people feel that spark, momentum becomes unstoppable.

From Prototype to Production

Prototypes today come easily. Production still takes rigour. Jimmy calls this stage productionisation, the hard, unglamorous bridge between “it works once” and “it works for everyone, every time.”

The first part of that bridge is evaluation. Because LLMs are non-deterministic, you can’t count on the same output twice. Airtable addresses this with evals, structured tests that define acceptable outcomes across thousands of prompts. Everyone, from engineers to designers, contributes to this shared library of expectations. Evals, Jimmy explains, become a “quality moat,” a living reference that guides improvement as models evolve. The second challenge is performance. After years of optimising for instant results, AI forces teams to design for latency. Users might wait seconds, even minutes. That means making waiting feel active. Airtable’s Co-builder streams its progress live, revealing tables and connections as they form. It’s still slow, but it’s engaging, transparent, and human.

Then there’s cost. Each model call has a price. A single user session can cost a dollar or more, a trivial sum in isolation but huge at scale. Jimmy’s framing is simple: value must outweigh cost. When one of his directors spent $700 in API fees to prototype an idea that would’ve cost tens of thousands in engineering time, the economics were obvious. What matters is outcome, not optics.

Finally, there’s enterprise reality. Leaders want AI innovation fast; legal and compliance teams say “not yet.” Meeting both sides means flexibility. Airtable allows customers to choose deployment models, maintain their own keys, and set zero-day retention policies. It’s not just good security practice; it’s the cost of adoption. And if the value is clear enough, Jimmy says, the business side will fight to make it happen.

Strategy Before Novelty

AI makes it easy to chase what’s impressive instead of what’s important. Jimmy warns against that trap. Teams should ground their experiments in company strategy in the unique problems their customers face and the distinctive data or workflows that make their products valuable.

It’s tempting to build “me too” features just because the technology allows it. Resist that. The differentiator isn’t that you use AI, it’s where and why you use it.

The Opportunity in Front of Us

Jimmy ends with an invitation, not a prescription. Every best practice in this space is still being written. For those willing to rethink old rules, this is a rare moment of creative freedom.

“Everything we’ve known is changing,” he reminds the audience. “You get to invent the next way people interact with products.” The challenge, then, is to play, prove, and productionise. Play with the tools until you uncover what’s truly possible. Prove where they deliver real value. And then productionise with the care and discipline that turns novelty into reliability.

If teams follow that loop, AI stops being a buzzword and becomes an invisible engine beneath meaningful work. The question is no longer whether your product needs AI; it’s where AI can help your users do their best work. That’s where the real opportunity begins.

Want to watch the full talk?

You can find it here on UXDX: https://uxdx.com/session/bridging-the-gap-how-product-ux-and-dev-can-build-ai-native-products-together/

Or explore all the insights in the UXDX USA 2025 Post Show Report: https://uxdx.com/post-show-report/