400,000 Customers, 100 Million AI Tasks: Zapier’s CEO on Making Computers Do More Work

Nathan Labenz welcomes Wade Foster, co-founder and CEO of Zapier, the automation platform that has been a no-code pioneer since 2011. The product’s reputation is still “connect two apps”: a Stripe payment that posts to Slack, a Gmail attachment that lands in Dropbox. Foster says the company is chasing a larger role. Y Combinator president Gary Tan recently tweeted that Zapier could be the AI-powered knowledge worker of the future, and Labenz wants to know whether that is how Foster now thinks about the business.

Foster’s first answer is competitive, not cosmic. The market around Zapier has filled in, which he thinks is good for users, but the company still wins on being easy to start and then unexpectedly deep. A new user can wire up two simple apps and feel the wheels turn. Stay a while, and Tables and Interfaces—launched in the last couple of years—let people build full applications instead of hunting for another tool. Easy to start, more depth over time: that, he says, is why people stay in the ecosystem.

The bigger line has been sitting on Zapier’s about page for years: “We’re just some humans who think computers should do more work.” Foster and the people around him keep having ideas for apps that turn out to be more hassle than they looked, because the distance from idea to a living thing is still large. Zapier’s job, as he frames it, is to shrink that distance until the product feels like a sidekick at work. The numbers already look like a knowledge-worker story: about 400,000 customers have delegated more than 100 million tasks to AI, and AI-related workflows are the fastest-growing category on the platform.

Foster’s favorite depth story is domestic. A customer’s wife opened a gardening business, loved the plants, and refused to send invoices. The husband, wanting her to get paid, built what he calls voice-to-invoice on Zapier. A Jotform on his phone is just a record button. He talks: he met Sally, $150 for tulips, $100 for daisies. Whisper transcribes the clip, ChatGPT pulls out the line items into a Google spreadsheet, and QuickBooks generates and sends the invoice. Ten years ago, Foster says, you could have raised a million dollars on that idea. The husband built it in 30 minutes to an hour.

A higher-stakes version comes from Siqi Chen, CEO of Runway, a financial-planning product Labenz describes as Notion meets finance—and not the AI-video company of the same name. When Runway’s new site went viral, dozens of people signed up every minute. The engineering team stayed on the product; sales was two people. Chen wired a Zap that dumps new leads into Zapier, asks ChatGPT to drop anything that looks like a free email domain, enriches the rest with Clearbit, then has ChatGPT draft a follow-up with a prompt that says, roughly, make it witty but not cringe. The drafts land in Gmail. In the morning the reps scan and hit send.

Asked about impact, Chen told Foster he would have had to hire another 10 or 20 people to do that job; instead he built it in 30 minutes on Zapier, ChatGPT, and a few other tools. Stories like that, Foster says, are common. The bottleneck is rarely the stack. It is coming up with the idea. After the novelty of demos wore off, Twitter filled with people asking what anyone actually uses AI for. Publishing Danny, Siqi, and the other messy, repeatable cases is how Foster thinks you give people a version they can copy even if their problem is not identical.

Labenz recites a rule making the rounds: use AI only where it is strictly required; anything that traditional code can do should stay in code, because it will be better, faster, and cheaper. Foster subscribes. At the current state of the art there is a real last-mile problem, so you have to be more prescriptive than people hope. LLMs earn their keep on jobs you could not do the other way, or only with exceptional pain: extracting information from unstructured text, categorizing a pile of items, generating novel copy. Inject the model into the middle of a deterministic workflow and you can build something that did not exist before GPT-4.

Zapier had to learn that the hard way. It launched a product called Central whose idea was “inference everywhere,” trying to let AI delegate a whole stack of tasks. Central is having some success, Foster admits, but the bulk of those 400,000 users and 100 million tasks still run through the traditional workflow engine. In that engine you can mix and match deterministic steps with a single AI step and delegate exactly the slice you need. That pairing, he says, is what improves end-to-end reliability.

Are users fine-tuning models to squeeze the last point of quality? Almost never. Most people get there with basic prompting, maybe a few examples, and they are not spending hours on the perfect prompt. Nobody on the platform, in the use cases Foster sees, is training their own models. The failures he does see are usually vague instructions—the same vagueness that would confuse a human assistant. The magic he is waiting for is a model that can take a messy request, rewrite it into a better prompt, or come back and ask whether you wanted A or B, the way a good assistant would.

Labenz is trying to get a model to write as him. He is exporting Gmail, Twitter, and the rest of his locked-up data, summarizing it into tens of thousands of words of factual background—he mentions a 50,000-word context pack—so the model does not have to invent his life, then training on his actual output for style. Tyler Cowen’s line, which he quotes, is that context is that which is scarce. GPT-4o fine-tuning results on a software-engineering benchmark have him thinking there is still juice in that approach.

Foster and his assistant tried a cheaper version. They fed the model a pile of his emails, internal posts, and Slack messages and asked it to describe the writing style, then iterated until they had a short, standalone prompt. They ran an internal test: employees submitted questions, Foster answered them, an AI Wade answered them, and the company had to guess. People still picked the real Wade a bit more often, but it was not easy. Labenz jokes that the CEO Turing test may be the next thing to fall. Foster wonders how many employees would love CEO advice without CEO judgment—maybe a perk.

The recipe is copyable today: dump samples into ChatGPT or Claude, ask for a prompt that would write in that style, loop a few times, and set the result as a system prompt. Labenz finds Claude best at imitating voice when you just paste samples. The failure mode is also copyable. Ask about something specific to Zapier that is not in the training data and not in the samples—what is the deal with the team retreat?—and the model will hallucinate a confident, correct-looking answer that is wildly wrong. Style transfers. Internal facts do not.

The other hard skill on the platform is decomposition: taking something you do by muscle memory—Siqi’s pipeline, Labenz recalls, was 16 steps—and turning it into triggers, actions, and field maps. Zapier now lets you describe what you want in natural language and have AI draft the Zap. People are attracted to it, Foster says, and it is still fairly immature, but it works often enough that the team gets daily notes of “holy cow that actually worked.” Without AI, even an easy product still asks you to click a lot of knobs. With it, you get a thought partner, a better rubber duck.

He sees two improvement vectors, and he thinks both are required. One is a better coach: the human stays in control, the model suggests directions and gets creative juices flowing. The other is a contractor: you said you wanted this, I already built it, should I drop it in? If you only do the building, you succeed with people who arrive with ideas. If you only do the ideas, the configuration is still sophisticated enough that people trip. Labenz points him at Gamma, a former guest, for a guided outline-then-slides loop with tasteful accept-or-reject. Foster is more taken with Claude Artifacts: AI on one side, a builder motif on the other.

That pattern, he suspects, is the next application layer—including disposable experiences you spin up, use, and throw away. How often are you on your phone hunting a utility in the App Store when it would be better to make the thing in under five or ten minutes? That is the time budget Zapier is gunning for. The same last-mile problem shows up when the draft Zap leaves the simple cases: it feels like self-driving, where the demo works and then edge cases that a human spots instantly trip the model left and right. Existing models still have room, he says, but some of this is waiting on a step-function jump.

Foster’s co-founder Mike Knoop, a previous guest, has argued that progress toward AGI has stalled. Labenz does not quite share that view and asks what the Zapier platform shows. Foster splits the difference. Knoop is focused on the G: the ability to learn something new from scratch and top domains. By that definition, today’s LLMs are trained on a set of stuff and then pattern-match. Foster thinks there is truth in the critique. He also thinks bigger, cheaper, multimodal models have made the commercial utility of automation a night-and-day change from 18 months ago. Years ago Zapier’s own email parser was the bane of the team’s existence. Now a basic prompt beats any parsing engine from two or three years ago. HTML email, Labenz adds, is so gnarly you might need to be AGI-complete just to deal with it.

Does Foster want a mid-scale intern-agent you can hand a project, or a more reliable local step inside existing Zaps—and does a godlike assistant become a competitor? Everyone wants to describe tasks in vaguer language and have an assistant just do them, he says. Under the hood he still expects a workflow engine routing data across services, because jobs differ wildly on latency, cost, and accuracy. He does not see much evidence that one magic model that secretly routes everything is next. In the meantime Zapier’s bet is to inject smarter, more flexible models into the toolkits people already have, with easier builder interfaces.

Labenz’s number-one tip for pushing those experiences is to fine-tune on chain-of-thought: capture the internal monologue between ticket and result so the model imitates the reasoning, not just the output. Foster’s version is almost the opposite of Google search. With a search box you are lazy and type the shortest keyword; with a model you want the fullest description of the task. More examples help. Narrowing the task and chaining those narrow steps helps more. Try to do several things in one prompt and the model trips. Humans do not walk around thinking in the smallest bit of a task. It is a nerdy concept, and it is the one that most improved performance inside the workflow engine.

He reaches for the elementary-school exercise: one child instructs another to make a peanut-butter-and-jelly sandwich, and the listener may only do what they are told. Chaos always ensues, because describing a muscle-memory job is much harder than doing it. Working with AI is that drill every day. Clear communication, he says, is only heading upward in importance. The people with the best communication skills may end up exceptional managers in this chapter of the economy. They are already used to delegation and to describing intent. They have tuned the engine. The rest of us are still figuring out what delegation even means.

Internally, the step-function was a founder-mode week. Around the GPT-4 release, some 15 or 16 months before this conversation, Zapier stopped the whole company—not just engineering—for an AI hackathon. In the weeks after, daily use went from about 10 percent of the workforce past 50, and Foster thinks it settled around 60 or 70 percent. It is hard to measure in shipping velocity. Qualitatively, AI is now a common tool. Engineers use copilots; there is a lot of buzz around Cursor; Claude Artifacts are fun. Data analysts, asked what they would do if ChatGPT were taken away, say you would have to pry it from their cold dead hands. HR put handbooks into chatbots, answers repeat Slack questions that used to occupy a fleet of humans, and drafts performance reviews from a sloppy list of accomplishments. The rough draft, Foster says, is quite good, and it beats wasting time on polish that does not matter.

Not every function is a winner. Customer service is where many companies claim huge gains; Klarna had just made waves by saying AI replaced the equivalent of 700 jobs. Zapier is finding those gains harder, Foster suspects because its support was already technical and highly automated, so the easy fruit was gone. Mileage varies. The only sure thing is a lot of experimentation. Culturally they started bottom-up at ChatGPT’s launch—“this is cool, you should try it”—and switched to a top-down mandate when GPT-4 arrived. Before the hackathon they sat with legal, engineering, and data-privacy teams and wrote rules of engagement around customer data. He would not suggest going wild west without that environment.

To keep the drumbeat, Zapier runs all-hands every Thursday and, about once a month, an AI showcase of half a dozen use cases on video or live. There is a Slack channel for showing off builds, and a handful of engineers plus non-technical but sophisticated AI builders host weekly office hours anyone can attend. The advice Foster would give other leaders is not to be shy about saying this is important. That, he thinks, is why founder-led companies have done better in this period: they can walk in and say let’s go. Sleep on it and it may not be a problem today, but give it a year, give it five—and five years is a long time in this space.

Labenz asks what a 10x more expensive Zapier looks like if they spare no expense on AI. Foster’s answer is a move from a suite of tools to packaged solutions: Tables, Interfaces, chatbots, and a code surface that is coming, each reinvented with AI as a first-order primitive rather than a chat box bolted on. ChatGPT made the world think chat was the new UI; the further he gets from that, the more he likes tables and workflows, rebuilt. Enterprise customers should not have to configure everything from scratch. DIY is still a haven, but marketing, HR, and finance teams want done-for-you. Professional services, priced at a premium while AI takes the labor out, are the obvious market—until, as Bezos put it, your margin is my opportunity. Capitalism competes the premium away. Consumers win. Come check out Zapier, Foster says on the way out: it is not the simple integrations tool people remember.