Howie Liu - Building Airtable - [Invest Like the Best, EP.375]
What this covers
Patrick interviews Howie Liu, co-founder and CEO of Airtable, which began in 2013 and now serves over 300,000 organizations. The conversation moves between software strategy, AI forecasting, and Liu's record as an operator. He sorts software into three buckets: broad-but-shallow suites like Office, verticals such as Procore and Veeva, and deep horizontal platforms. His founding belief, taken from his time at Salesforce, is that most software splits into a data model, business logic, and an interface layer. He argues Salesforce beat Siebel and SAP through a Metadata-Driven Platform Model that let every customer customize, not through the best prepackaged CRM.
The AI section holds the most contested claims. Liu says current LLMs are both broad and deep, that reasoning quality matters more than new capabilities, and that big platforms like Microsoft can arrive years late and still win on distribution. He lays out an Accuracy Threshold Ladder running from 20 percent (useless) to 95 to 99 percent (top of class). He also contrasts humans chaining functions with autonomous agents that drift, as with Auto-GPT. Patrick presses on non-determinism and what it demands of builders. Liu expects AI talent to arrive in waves, ending with industry operators paired with technologists.
The last stretch is about running a company. Liu calls his two reductions in force, made with about $1 billion on the balance sheet, his hardest decision. He treats constant hiring as a check on employer brand, much like a funding round. He also describes starting customer engagements by asking what is motivating change, and notes how streaming is upending media companies. The closing portion covers early product-market fit, first monetization, and the first unicorn round.
Liu argues that Airtable's defensible position in the AI era comes from a horizontal-but-deep platform (data model, business logic, and workflow interface) that lets each customer build its own AI-enabled processes, rather than from wrapping one LLM use case or competing in vertical niches.
- LLMs need structured data and a human-facing workflow layer to deliver value, and a platform supplies both
- Platform customizability (as with Salesforce) beat hard-coded vertical products, and AI accelerates that pattern
- Enterprise adoption is gated by behavioral change, so the winners will be companies that motivate and guide process transformation
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Salesforce beat Siebel and SAP not by building the best prepackaged CRM but by building a metadata-driven application platform where every customer can customize schema, objects, logic, and page layouts, starting from a template that codifies best practices.
“Salesforce wins and has gotten to... was not because they went and built the best possible CRM all hard coded into a prepackaged app, but rather that they built a platform.”
The latest LLMs can reason about advanced topics, which makes them both broadly and deeply capable, unlike earlier two-layer neural nets that were only narrowly applicable to tasks like simple classification.
“these LLMs are both broadly capable but also deeply capable”
Neural nets and LLMs represent the ultimate meta-approach to software development, because instead of hand-coding every if-this-then-that rule, developers train the system to perform the behavior, and LLMs show emergent capabilities even model researchers did not anticipate.
“the ultimate meta-approach to software development is instead of having to go and literally code out to the specific if this than that statement”
Streaming has become the massive change driver for media companies, forcing them to rethink their value chain, since the pace and volume of content for owned streaming services upends how they produce content and distribute it on linear TV.
“streaming has become the massive change driver across every media company”
Large platform companies like Microsoft can make multiyear bets and be a few years late to a technology shift, because their distribution advantage lets them win market share even with a mediocre late product.
“you can afford to be a few years late, but you're going to have such a distribution advantage by the end of it that even if you build not that great of a product and it's a little late, you'll still be able to win a lot of market share”
Salesforce started as a non-customizable product similar to SMB CRMs and only a few years after founding, helped by product-led growth, evolved into a much more flexible platform foundation.
“Salesforce started out as a not customizable product. It was a true product... they evolved to having a much more flexible platform foundation.”
Current LLMs already have the pieces of a much more advanced intelligence; the main gap is learning to chain them together, and reasoning quality, not novel capabilities, will be the most important advancement.
“I don't think we need dramatically novel capabilities”
Most people will not spend hours experimenting to get value from LLMs, so if models give consistently good results on the first try without prompt expertise, adoption will become broad, and sticky recurring use cases will drive durable adoption rather than novelty uses.
“if the prompting gets easier where it will more consistently give you good outputs even without you having to be overly clever”
Software falls into three buckets: broad but shallow horizontal products like the Office suite, verticalized solutions like Procore, Veeva, and Dandy, and horizontal-deep platforms that give building blocks so users can build very specific applications, which is where Airtable plays.
“And I think the third bucket, which is the one we play in, is actually both horizontally applicable, but also pretty deep.”
Airtable does not aim to be a foundation model developer; it plans to plug in the best models, whether open source like Llama 3 or Anthropic and OpenAI models, and uses its large customer base as a petri dish of real-world LLM experimentation.
“we're not trying to go and be the leading foundation model developer. In fact, we want to just ride the wave of all the best models”
Humans can chain primitive functions into a sequence without drifting from the original intent, while fully autonomous LLM agents like Auto-GPT progressively drift off course, which is why they have been ineffective on real-world problems.
“it just progressively drifts more and more off course, which is why those completely automated LLM core agents were and still are pretty ineffective”
Some small business owners, such as restaurant operators or solo lawyers, have an innovative operational mindset, often already using tools like Zapier and Webflow, and they adopt Airtable to customize operations for an edge, while others prefer simple out-of-the-box products.
“there are some people who just naturally have an innovative and operational mindset, and these are the people who might also be using other low-code products.”
The hardest decision was doing two reductions in force despite having about $1 billion on the balance sheet, because Liu wanted to keep hiring over time and to run a leaner, more agile company focused on AI.
“we made a tough decision to actually do two RIFs. And now we're done with them because we're cash flow positive and growing.”
Keeping a recruiting flywheel going acts like a funding round for a private company, because continuous hiring checks employer brand and brings fresh external market perspectives into the business.
“it's almost a little bit like doing a funding round. For companies that aren't public and aren't constantly fundraising”
Talent moves into new technology with a lag: first people closest to foundation models, then SaaS veterans and founders building AI-first software, and the third wave will be industry operators who know their sector and use AI as an advantage to win best-in-breed media companies, retailers, or construction firms.
“I think the third wave is going to be the real business operators who know that industry, but are also savvy enough”
Airtable's combination of a system of record, the ability to integrate with existing systems of record, and a workflow layer on top are the ingredients necessary to unlock value from LLMs, and the platform was almost made to plug in off-the-shelf improving models.
“And it turns out those are all the ingredients that I think are completely necessary to actually unlock value out of these LLMs.”
Developers building applications must adapt to probabilistic rather than purely deterministic programming, and this non-determinism is a feature of LLMs that builders have to respect and design around.
“it's just going to take some time for people developing applications to get used to a probabilistic way of programming versus just purely deterministic.”
The best-positioned future AI-first industry companies will be founded by a technologist together with a domain operator who understands the industry and is an aggressive executor, rather than by pure technologists alone.
“But I think the third wave is going to be the real business operators who know that industry, but are also savvy enough”
An LLM at 20 percent accuracy on human tests is not useful, 60 percent starts to be interesting, 75 to 80 percent is a decent average performer, and 95 to 99 percent is top of class, so moving up this ladder is where value unlocks.
“A 20% accurate LLM is not that useful... But you get to 60%, you're like, okay, this is starting to become interesting.”
Most software can be platformized into building blocks of a data model, business logic, and an interface or workflow layer, analogous to model-view-controller programming, and this founding belief was inspired by Liu's time at Salesforce.
“most software can be platformized or can be reduced into these building blocks of basically data model, business logic and then an interface or a workflow layer.”
When engaging a customer, start by understanding what is motivating change, because the worst software businesses change things for the sake of change, and shiny new products without a strategic driver are not durable.
“I first start with trying to understand what's motivating change because I think the worst software businesses are changed for the sake of change.”
It is philosophically important for companies to always be hiring, because it would be arrogant to assume you will never need new skill sets as markets and strategy change, and hiring also brings fresh external perspectives that act like a funding round.
“it would be arrogant to assume that all the skill sets you ever need to execute even as the market changes”
LLMs could already be exploited for 1,000x or 10,000x more economic value than they currently generate, and the big gap in the market is the data, workflow, and human interface layer that Airtable is positioned to supply.
“these LLMs could already be exploited for 1,000x or maybe 10,000x the economic value impact that they are today. And the big gap is really around having that data and workflows and human interface layer”
AI will be the big tailwind for platform plays over verticalized solutions because it can gently guide less innovative customers toward specific use cases without requiring them to build everything from scratch, which may let Airtable win the down-market in parallel.
“AI is going to be the big tailwind for platform plays versus verticalized solutions.”
Generative video models like Sora could replace much CGI and stitch scenes into full movies, and the winners in media will be industry operators fusing technology and storytelling, as Pixar combined technology with filmmaking, rather than pure technologists.
“Pixar was a fusion of both technology and filmmaking and storytelling, and you couldn't have had just the technologists.”
Airtable chose two reductions in force even though it had about 1 billion dollars on the balance sheet, because it wanted to keep hiring continuously, get outside perspective, and execute in a smaller, more agile way, and it is now cash flow positive and growing.
“we made a tough decision to actually do two RIFs.”
Airtable beat competitors that had raised more money earlier by shipping a better, well-architected product in about 2.5 to 3 years, which Liu considers the fastest it could ship a good product, and that product showed real product-market fit through retention and organic virality.
“we just outexecuted them on building a better product and actually shipping it”
Very abstract building-block primitives leave many users stuck because they do not know how to build what they want, so platforms need templates, tailored onboarding, and showing users how to build rather than telling them.
“If you give them very, very abstract primitives, building blocks, sometimes they get stuck, and they don't know how quite to build the thing they want.”
Most successful platform companies eventually reach a high level of platform customizability, as Salesforce did and Atlassian's Jira did, though these tools are hard for lay users to customize, which is the gap Airtable fills.
“most successful platform companies eventually get to that level of platform customizability”
Throwing thousands of people at AI will not solve it faster; smaller, tight-knit teams that work directly with customers, with very low iteration cycles, execute better, consistent with Amazon's two-pizza rule.
“I don't think this is a problem that throwing thousands of people at it will solve faster.”
AI will accelerate customization in two ways: category A, where AI builds a customer-specific app from public information about the company, and category B, where AI is embedded into the app's workflow steps so the apps themselves are AI-enabled.
“I think that AI will for sure accelerate it, both because, a, the way that you build apps can be accelerated with AI”
One of the fundamental bottlenecks to AI value is behavioral change and enterprise willingness to adopt, with SMB and consumers adopting faster than large enterprises.
“I think that actually one of the fundamental bottlenecks is going to be behavioral change and the willingness of enterprises.”
A human doing a research project breaks it into preliminary questions, searches and synthesizes sources, forms a thesis, drafts, critiques, and revises, and the LLM capabilities for each step already exist; what is missing is chaining them together to get the most value.
“we didn't quite figure out how to chain them together to get the most value out of them.”
Even with AGI-level autonomy, a human interface where people can see data, review AI-generated output, edit it, and hand it off is probably the most important part of unlocking LLM value.
“That's still going to be a really, really important, in fact, probably the most important part of unlocking these LLMs' value.”
As an investor, Liu would favor generally horizontal but deep platforms over vertical narrow products or shallow horizontal products, with platform customizability as table stakes and AI integrated across many parts of the workflow rather than as a single feature.
“I would definitely be a much bigger believer in generally horizontal but deep platforms as opposed to either vertical and narrow or horizontal and shallow products.”
Industries facing existential change, such as media moving to streaming and retail moving to omnichannel, are where companies will spend time and money on operations change, creating opportunity for disruptors like Airtable and new startups to capture the flow of new value.
“when you have a motivator that's flowing from existential change or impact from the very top, the entire industry is being upended.”
Using Roelof Botha's crucible moments framework, Airtable's history shows that the first three years were a prelaunch build during which market timing, bottoms-up adoption, and browser performance gains mattered more than funding.
“there was something happening in the world at that moment, both behavioral change on behalf of end users within companies, bringing in their own products to work”
After rates moved away from ZIRP, customers shifted from default spending to asking first why they use a product, how it is differentiated, and what its business value is, reevaluating every seat and vendor, which led to contraction even with little logo churn.
“customers went from a default spend and then ask questions later to ask those questions first. And in fact, reevaluate every seat, every vendor”
Describing the role as capital allocation is misleading because capital is fungible; the best scaled operators like Elon Musk and Mark Zuckerberg stay hands-on in product details, and the biggest, best-resourced team often does not win.
“I hesitate to use that term too wholeheartedly because capital is fungible.”
Airtable launched around the same time as Figma because bottom-up adoption of employee-introduced tools and improved browser performance made a rich desktop-grade web experience possible, and Liu believes market timing strongly determines an idea's success.
“I do believe market timing has a lot to do with the success of an idea.”
It is really challenging to build a pure agentic LLM experience that does a bunch of work behind the scenes without human input or without the ability to present well-modeled data, so a platform with a system of record and workflow layer is needed to unlock LLM value.
“I think it's really challenging to build a pure, for instance, agentic LLM experience that just does a bunch of stuff behind the scenes without human input”
Vendor AI features such as Slack AI summaries are useful but strategically less important than products that rethink an end-to-end business process and the job to be done, rather than adding AI as a feature on top of an existing product.
“what I care more about is what I would call process transformation.”
Airtable's biggest evolution has been moving from pure LEGO kits to productized solutions that sit between a template and a full app, such as sales or support CRMs, while still built on the same extensible platform.
“finding that blend of it's both platform and also application and solution and the way that we engage with our customers”
During the demand surge Airtable grew headcount more than 100 percent per year for several years, from low hundreds to over 1,000 employees, and Liu said he most neglected building the organizational foundations and management structure needed for scale.
“during that stretch, we grew headcount at more than 100% per year for multiple years. So we went from low 100-ish employees to over 1,000”
Airtable saw very little logo churn during the spending reset because it was embedded in sticky, important processes with high-value data sets, even though seat and revenue contraction still occurred.
“we were embedded into sticky enough use cases. I mean, these are really important processes, high-value data sets”
Airtable reached product-market fit through durable retention, team expansion, and organic virality before it began monetizing, which was about a year after launch when it first pushed paid conversion.
“And so this next phase was about what is the business of Airtable. And we had always advertised pricing plans. We didn't really try to monetize or convert people up to the paid plans until maybe a year in.”
During COVID, every company had to become hybrid or fully remote overnight, which pulled forward demand for collaboration software, and Airtable grew as a genre of structured collaboration tools beyond basic chat and video.
“Every company had to become a hybrid or a fully remote company overnight.”
Over a ten-year horizon Liu is at the extreme bullish end, arguing that even if AI research froze at today's models such as GPT-4 and Claude 3, many trillions of dollars of GDP could still be created by better applying them to business and personal use cases.
“even today's LLM capabilities, ask every AI researcher... to stop and just freeze time... I still think there are many trillions of dollars of GDP that can be created”
Legacy media companies are launching owned streaming services to avoid disintermediation and gain subscriber data, and this requires rethinking the entire value chain of content production and distribution, which creates opportunity for operations platforms.
“every big existing media company is trying to launch their own owned streaming service for a number of reasons”
Although multimodal inputs and larger context windows will advance, the most important advance is improvement in reasoning quality, the gap between a five-year-old and a graduate-level or expert researcher.
“I think the most special advancement, in my opinion, is going to be the quality of the reasoning.”
Airtable's large enterprise business, which is now more than half of total revenue and the biggest growth driver, is concentrated in media, retail, financial services, and tech, because these industries have undergone the most change.
“our large enterprise business, which now is far more than half of our total revenue and the biggest driver of growth for us. But that's really in media, in retail, in finserve, and then in tech.”
Airtable advertised pricing plans from the start but did not really try to monetize or convert users to paid plans until about a year in, and the first $10,000 customer felt like a major milestone.
“We didn't really try to monetize or convert people up to the paid plans until maybe a year in.”
Netflix built sophisticated content production systems on Airtable because Airtable provided building blocks that customers could combine with their proprietary process knowledge, rather than being shown a pre-engineered content production system.
“So that's why we have companies like Netflix early on building these content production systems that are actually quite sophisticated.”
The Apple Macintosh was developed by about 50 people while Apple put thousands on the Lisa computer, which Liu argues was less innovative and less good, illustrating that small renegade teams produce better products.
“Mac was made by 50 people when Apple, the massive behemoth put thousands of people on, I think, it was their Lisa computer”
The shift to digital commerce and omnichannel retail, where stores become experiential and transactions happen online, is a big change driver that makes retailers willing to spend time, energy, and money rebuilding operations.
“This shift to digital commerce and omnichannel and stores becoming more of an experiential place”
Instacart's predecessor Webvan tried the same idea about 15 years earlier and failed, likely because it was the wrong time, while Instacart later succeeded once timing was right.
“Instacart tried 15 years prior through Webvan, didn't work. It was probably the wrong time.”
A movie with a massive budget and top talent can still be a flop, so the executive producer role is about making sure teams operate in harmony and that the story and feel are right.
“a movie that just has a massive budget and you throw the best talent at it might still be a flop.”
Even the current models' capabilities across every industry function, use case, and team or individual level remain a relatively undiscovered frontier, so much of their applicability is still unknown.
“the frontier of even what the current models are capable of today... is still a relatively undiscovered frontier.”
Airtable's latest phase requires a repeatable go-to-market engine aligned with its platform vision, including going to customers instead of waiting for them and acting as a strategic consultant to large customers spending more than 5 million dollars.
“And you're spending $5 million-plus with us, we better understand your business”
Because LLMs are both broadly and deeply capable, companies must find every point in a business process where an LLM step can automate work that once took a human hours, and can sometimes yield better results.
“you got to be really imaginative to find every single place in a business process where you can introduce an LLM step”
Enterprise AI adoption will be a sanctioned and methodical process with lengthy behavioral and process implementation timelines, so not everything will change within a year.
“So I don't think a year from now, everything is going to change everywhere.”
ChatGPT has grown mostly through consumer use because individuals do not face the safety and copyright issues that enterprises must address, so consumers and small businesses experiment faster.
“you're already seeing that rapidity of adoption with ChatGPT being mostly consumer-driven because you don't have to worry about all the safety issues and copyright issues.”
Many big companies spent large sums on the vague promise of digital transformation without a strategic thesis tied to a specific application motivated by a changing environment, though digitization is a good use case when motivated by real change.
“there were a lot of dollars spent by a lot of very big companies on the vague promise of digital transformation”
During COVID, Airtable saw churn from SMB customers who downgraded because they were shutting down or laying off staff, so it raised another round led by Thrive to gain certainty and remain long-term oriented.
“We ended up raising another round in the midst of that, just to give ourselves a little bit more certainty, stability”
Wrapping an LLM around one narrow application is like taking a broadly intelligent human and having them perform one rote task every day, so platforms should let each customer design how to exploit LLM capabilities in their own work.
“having a platform that doesn't just wrap around the LLM and use it for one narrow application, that would be the equivalent of taking the beautiful human intelligence of a person”
A second concentration is in dynamic front-office processes, such as marketing operations, digital product operations, content production, and retail store openings, where the process is changing quickly and teams need agility.
“a lot of front-office use cases where the actual process itself is quickly evolving.”
Early expectations that new direct-to-consumer brands would take down large retailers proved too strong, because existing retailers fought back well and B2C was not as rapid a disruption as predicted.
“I think the retail companies fought back well and B2C ended up not being quite as big and as rapid a disruption”
After monetization began, Airtable took three to six months to reach $500,000 in revenue, a few more months to reach $1 million, and between one and one and a half years to go from $1 million to $10 million.
“it probably took us three to six months to hit the first $500,000, maybe only another few months to hit $1 million”
Models like OpenAI's Sora will replace much CGI in the near future, and in the not-too-distant future enough generated scenes can be stitched into entirely end-to-end movies, which some filmmakers are already attempting.
“you can replace a lot of CGI with models like Sora from OpenAI”
Airtable runs its own product operations on Airtable with AI steps that capture customer feedback, synthesize it into insights that triage the product roadmap, help generate product requirement documents, and produce summarized updates.
“which we actually run on Airtable and have introduced lots of different AI steps as part of that”
Lukas of CrowdFlower gave Liu a small personal check when he left to start his own company, extending his runway enough to found a company, go through YC, and reach his first fundraise and exit.
“he basically wrote me a small personal check, didn't expect to have it paid back on any particular time frame.”
Airtable priced its first unicorn round around the time revenue reached roughly $10 million, with Thrive, Benchmark, and Coatue as the main investors, which shifted the company from an indie darling to a company needing to scale into the mainstream.
“that was Thrive, Benchmark, Coatue were the main investors.”
The first 10,000 dollar customer felt like a major milestone because it showed that pure software, delivered from the cloud, was valuable enough for someone to pay real money, a feeling rooted in Liu's modest upbringing.
“I remember the first time we got a $10,000 customer”
After the first unicorn round, scaling a team of 30 to 40 people into a real company meant the business moved from being a cult-like indie favorite to needing to become mainstream, which was a phase shift.
“it felt like a phase shift in the company again”
Airtable's customers range from individual cattle farmers managing livestock to small nonprofits, legal firms, and very large enterprises running end-to-end processes.
“we have everyone from literal cattle farmers managing cattle in Airtable to small nonprofits or legal firms”
Liu would draft the big hyperscalers highly, with Microsoft and Google at the very top because of their role in AI, while expecting many low-draft upstarts and industry operators to produce major new companies.
“Microsoft and Google actually get a very top pick in my mind because of the role they're playing in AI specifically.”
Mark Zuckerberg remains intimately involved in product details, including spending time with individual Oculus engineers and physically touching and feeling the product experience during his focus on it.
“he would actually go and spend time individually with Oculus engineers”
A single model that accepts full video, long reports, documents, text, and audio and outputs any of these is coming very soon, but it is less profound than reasoning gains because value can be piecemealed today through specialized models integrated with data and workflows.
“But to me, actually, that's a little less profound because you can piecemeal get some of those pieces of value today”
Airtable has more than half the Fortune 500 as paying customers, around 100,000 total paid customers including SMB and self-serve, and hundreds of thousands of free organizations, giving it a sizable distribution base versus a new startup.
“We have more than half the Fortune 500 as paying customers of some kind. We have around 100,000 total paid customers”
If he could run any non-tech business for ten years, he would pick a theme park business such as Disney's parks because it would be fun and about creating joyful experiences, despite operational and political challenges.
“if I can take off, let's say, just the theme parks part of Disney”
Liu's parents supported his decision to leave stable income for entrepreneurship despite a family with little money and few entrepreneur role models, even offering what they could spare.
“I think my parents have just both been so incredibly supportive of my journey as an entrepreneur”
Patrick closed by encouraging listeners to build something with Airtable, describing it as a general-purpose computer that is not constrained by technical ability.
“I would encourage everyone listening to go try to build something with Airtable.”