Person
3 episodes · 54 quoted claims
Zapier CEO Wade Foster says he points an AI agent at the company's Slack, email, meeting notes, project management, CRM tools and code base to get summaries of company activity.
“I can point an agent at all the data sources inside of a company, whether it's, you know, our Slack instance, email, meeting notes, project management, CRM tools, code base, et cetera.”
Wade Foster says meeting recorder tools can mix up who said what, so AI summaries are often only directionally correct.
“One of the challenges with a lot of the meeting recorder tools is speaker attribution. It can mix up things like that.”
Wade Foster cites Zapier co-founder Brian's line that leaders can delegate work to AI but not accountability for its quality.
“my co-founder Brian has this line, you can delegate the work, but you can't delegate the accountability.”
Wade Foster lists as a Zapier communication rule that the author of a document should put more effort into it than the reader.
“I think one, the author should put more effort in than the reader.”
Zapier runs AI bots in customer support with humans reviewing at several steps of the workflow.
“we have, uh, AI bots that work in customer support, and there's human in the loops on, you know, a variety of steps in those workflows”
Foster says Zapier's customer support bots are measured at every step with metrics such as customer satisfaction (CSAT) and internal evaluations.
“those bots are measured very specifically. And, like, we have metrics every step of the way that are man measuring, you know, CSAT”
Wade Foster says his custom AI writing skills, trained on his writing, only get 80 to 90 percent of the way to sounding like him.
“Even still, I'd say it only gets, like, eighty, 90% of the way there.”
Wade Foster says some newer AI models are less good at communication tasks than earlier ones, and his co-founder says they sound 'fried'.
“some of the, like, newer models, I find to be even, like, less good at communication tasks.”
Wade Foster says he now uses voice-to-text heavily and that reading his own transcripts made him think more before speaking.
“I am a huge voice-to-text fan now because of AI.”
Wade Foster says AI models are heavily goal-seeking because of how they are reinforced, which makes them tend to tell users what they want to hear.
“these models are all goal seeking, um, like, heavily goal seeking. That's how they get reinforced.”
Wade Foster says he changed the system prompts on the AI models he uses so they debate and critique him and challenge his assumptions.
“changed the system prompts on all the models and things that I use to say”
Wade Foster recommends connecting a large language model to an organization's own communication systems and data, saying results are much worse without that context.
“if you are not hooking up your LLM to the actual data that you have inside your organization, you are getting a much worse experience.”
Wade Foster grew up in Jefferson City, the capital of Missouri.
“I grew up in central Missouri, Jefferson City, it's the capital of the the the state.”
Wade Foster says the financial crisis at the end of his university years was his wake-up call that he had to make his own path.
“what really kind of was the wake-up call for me was the tail end of university when the uh, financial crisis hit.”
Wade Foster says Stripe and Twilio both launched in 2011, the year Zapier began.
“So this is 2011, SaaS is on the rise. This is the year that Stripe launched, Twilio launched things like that.”
Zapier's founders built their first prototype over a weekend at a local Startup Weekend hackathon, working with Mike.
“we've got this local hackathon coming up, startup weekend. We go, we team up with Mike, we hold up into, like, a like a garage in the back of this building all weekend.”
The Zapier prototype demoed well and won the hackathon.
“came out and had a prototype. It worked, it demoed well, uh, so ended up winning the hackathon.”
The Zapier founders initially worked on the product on nights and weekends while keeping their day jobs.
“We're having fun. So we said, we'll do it nights and weekends, kept our day jobs”
Wade Foster's earlier job was at a local software company selling a natural language processing product that automatically graded student papers.
“it was a professor trying to, like, automatically grade student papers, which now seems quaint.”
Zapier's first customer was Andrew Warner, who ran a podcast, after Wade Foster emailed him about a PayPal and Highrise integration he was seeking.
“I remember the first customer we had, guy Andrew Warner. I'd emailed him, he ran a podcast”
Wade Foster says he quit his day job in January, about a month after the first customer's reaction in early December.
“It was like early December that this happened, and, uh, by January, I'd quit the day job”
Wade Foster recalls that GPT-4 came out roughly six months after GPT-3.5.
“So first, it came out roughly six months, I think, after 3.5.”
Zapier paused normal work for a week and held a company-wide AI hackathon that included non-engineers.
“We paused the company for an entire week and ran a hackathon. Didn't matter if you were in engineering or not.”
After Zapier's AI hackathon week, daily AI use at the company rose from about 10% of employees to over 50%, according to Wade Foster.
“And at the end of that week, we went from about 10% of people using AI as part of their daily job to over 50%.”
MCP stands for model context protocol, which Wade Foster describes as a way for agents to talk to other agents and tools.
“now there's this new thing, uh, MCP, you called it stands for model context protocol, but effectively it's just a way for agents to talk to other agents.”
Zapier's business model is a subscription combined with usage-based credits.
“Yeah, I mean, for Zapier, it's pretty simple. It's a subscription with usage-based credits associated with it.”
Wade Foster argues the leading AI labs keep leapfrogging each other with models that are good at different tasks, so companies shouldn't commit to one AI stack.
“Like they're constantly leapfrogging each other, their models are good at different tasks, different capabilities.”
Zapier converts agent tasks into deterministic workflows, which Wade Foster says avoids the cost of an agent burning expensive tokens continuously.
“we're turning those workflows into deterministic workflows. Now, why is this important? Is if it's running agentically the full time, you're burning tokens left and right, and tokens are very expensive.”
One of Zapier's three AI transformation efforts is a 'software factory' that builds the machinery that writes code, rather than engineers writing it directly.
“So we have a software factory that's working on building the products of the future.”
Zapier uses go-to-market agents, particularly in its small-business funnel where it previously could not afford sales staff.
“We've got these go-to-market agents that are doing the same thing, you know, a lot of that is working at our, like, our small business funnel”
Wade Foster says Zapier employees now use meeting prep agents for nearly everything, a task that was previously not done well.
“we've all got meeting prep agents for pretty much everything we do when nobody was really doing that before.”
Wade Foster suggests companies may give employees token budgets that cap how much they can spend on AI.
“Maybe employees get token budgets where it's like you're only limited as uh allowed to spend a certain amount.”
Wade Foster says Zapier doesn't have many outside investors, so it faces less investor scrutiny than a public company.
“we we're fortunate enough that we don't have like a ton of outside investors. And so we don't have the same scrutiny”
Zapier CEO Wade Foster says he mostly uses the Cursor AI coding tool as his daily driver.
“Yeah, so I mostly use cursor every single day.”
Wade Foster says the newest top model scores about 40% on Zapier's AutomationBench, the highest result on that benchmark.
“It performs about 40% of the tasks accurately, which is the highest that there is.”
Wade Foster says Zapier's AutomationBench contains roughly 600 tasks resembling normal knowledge-work workflows across disciplines such as marketing, sales, HR and operations.
“So we have probably 600 tasks that are of that variety that describe like a normal knowledge work workflowy task”
Wade Foster says about 80% of what Zapier customers use AI agents for would be better done with deterministic code.
“the vast majority of what people are using an agent for 80% in fact, they probably should be using actually old-fashioned deterministic code.”
Wade Foster says building workflows with no-code boxes feels antiquated and that AI now does the building, summing it up as 'no code is code'.
“the idea of building a no code, it feels antiquated to me. It's like the new code no code is code.”
Zapier's automated support email program runs five independent AI agents to evaluate each troubleshooting case, according to Wade Foster.
“it spins up five agents, independent agents that evaluate the troubleshooting situation.”
Wade Foster says the automation market is far larger than he imagined when he started Zapier about fifteen years ago.
“The market for automation was like orders of magnitude bigger than I ever thought it was when we started the company fifteen years ago.”