Zapier stopped work for a week in 2023, put every employee in front of a keyboard and told them to build something with AI. Daily AI use inside the company went from about 10% of staff to over 50%.
Most companies met the same moment with a memo from the chief executive or a standing AI committee. Foster said every company he has since watched run the same week-long pause reports that the hands-on time mattered more than either of those.
"It is not hard to find individuals that have like quote unquote 10x their productivity with AI. It is much harder to go look around and find the companies that are 10x productive because of this."
Foster co-founded Zapier in 2011 with a prototype built over one weekend at a Missouri hackathon, and has spent 15 years connecting software that was never designed to connect. He is now selling that same service to AI agents.
I listened to the full interview so you can skip it. 36 minutes of audio, 18 minutes of reading.
Here are the 11 takeaways that matter.
👤 Guest: Wade Foster, Co-Founder and Chief Executive of Zapier, the app-integration company he started in 2011 and is now rebuilding around AI agents
🎙️ Host: Rana el Kaliouby, an AI scientist, investor and author who co-founded the emotion-recognition company Affectiva
📰 Published: 9 September 2026
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 36 min | ✅ Time saved: 18 min
Key Takeaways
Individuals who 10x their own output do not add up to a company that 10x's anything
His test for a 10x company is revenue up ten times or costs down to a tenth, and he does not see it
Pausing the whole company for a week took daily AI use from about 10% of staff to over 50%
He says hands-on building time beats the chief executive's memo and the AI committee
Only three places in a company can produce a 10x gain: products, go-to-market, and coordination cost
Going all in on a single AI provider is a strategic error, because the frontier keeps moving
Open-source models are close behind at a fraction of the cost, and the labs keep leapfrogging each other
An agent left running all the time burns tokens; the same job written as fixed code is cheap and reliable
Fragmented agent setups across a company leave security holes nobody is watching
Meeting prep is work that was never done properly before, because it cost somebody else's selling time
The token-maxing era of AI spending is ending and ROI questions are coming back
Likely answers: per-employee token budgets, and moving workflows off models onto code
Zapier's org chart still looks much as it did, and he does not expect that to last
1. Integrate With X
Foster did not grow up expecting to start a company. He grew up in Jefferson City, in central Missouri, where the jobs he could see were the state government, medicine, teaching and the police. The financial crisis hit at the end of his time at university, and he said that was the moment the calculation changed for him.
Being a good student stopped counting when nobody was hiring. "And I think that was when I realized like no one's waiting to give a hand out to me"
He took an internship at a small software shop in Columbia, Missouri, and decided building software was what he wanted to do
The idea for Zapier came out of a repeated pattern in software company support forums in 2011, the year Stripe and Twilio launched. Users kept asking for integrations that never arrived: "And the customer request would always be when will you integrate with X, right?" — followed by a queue of other customers agreeing, and then a product manager promising to take a look
"And if you've ever worked at one of these companies, that's code for yeah, probably not going to happen"
His co-founder Brian's insight was that the integration could be a simple interface rather than an engineering project, so a user would not have to touch the APIs. Foster's own day job made the case for him: "My day job, I am working with a Marketo API doing email marketing and I'm a bad engineer and so I'm having a tough time."
They built the prototype in a weekend at a Startup Weekend hackathon and won it. "It worked. It demoed well."
There was no local investor base for three recent graduates in central Missouri, so they kept their jobs and built it at night and on weekends over takeout
2. How Much Do I Owe You?
Before Zapier, Foster had been trying to sell software for a local company doing natural language processing — a professor's system for grading student papers automatically. "In fact, it was a professor trying to like automatically grade student papers, which now seems quaint," he said. "It's like ChatGPT could do that."
He could not sell that product at all, and concluded the problem was him. The software was well built and well designed, and he still could not find a buyer
Zapier's first customer was Andrew Warner, who ran the interview site Mixergy, found through a forum post asking for a PayPal-to-Highrise integration. Foster cold-emailed him, then looked up which other tools Warner used — Wufoo and AWeber among them — and offered to connect those too
The product barely worked and needed a Skype call to set up. Foster walked Warner through it step by step, and when the test form submission finally landed as an email address in AWeber, Warner's reaction was: "Oh, this is incredible, Wade. Like, how much money do I owe you? Like, this is going to save me hours every single day."
That contrast — an unsellable polished product against a barely-working useful one — was the signal. "And so I was like, man, if we could just get this product to not totally stink, like I think we're on to something."
It happened in early December 2011; he quit his job in January
3. The 2023 Code Red
Foster called an internal code red when GPT-4 launched. He said three things about it were true at once: it arrived roughly six months after GPT-3.5, the capability jump was obvious in ordinary use, and the cost had come down meaningfully.
The code red was about the trend line, not the model. He said the question was what happens when the next version ships, and whether the releases get faster, better and cheaper each time
Any one of those three being true would matter; all three would make it the most important technology of his lifetime. "If all three of those are true, then this is the most important technology that has maybe ever been invented"
Zapier had no code-red procedure and had never used the term before. Foster said nobody knew what steps a code red was supposed to trigger — it was a way of making people pay attention while the company worked out what to do
He drew up a list of changes to the product and to operations, and said in hindsight one action out of that list did more than everything else on it
The week nobody did their job
Zapier paused the entire company for a week and ran a hackathon in which everyone, engineer or not, had to build something with AI
Daily AI use went from about a tenth of the company to more than half. "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%"
The point was building an intuition for the technology, not shipping anything. More of the company came out of it with a view on what the models could do, how Zapier might build alongside them, and where they could be used on Zapier's own operations
Foster said he has since seen the same exercise work at other companies, and that all of them report the hands-on learning time being more useful than a memo from the chief executive or an AI committee
4. Agents Talking to Agents
The host said her team runs Claude Code and a set of AI agents, and came to Zapier to connect them to email, calendar, Airtable and PitchBook. Foster's account of why that became Zapier's business is that the integration work was already done.
Zapier's integration story is 15 years old, and he claims it as the company's edge. "We've just gotten really good at integrating any tool that you have on the internet, far better than anyone else"
Models are strong on general knowledge and weak on the specifics of any one company, so a question about a marketing campaign or a product decision gets an answer that is broadly reasonable and not useful
Give the model the company's own context and the answer changes character. Foster listed CRM records, recorded sales calls in Gong, email and Slack: "Stuff that you start to go, that's as good as I'd get from like a good teammate internally or maybe better in some cases because it's able to read over all this stuff that most humans can't do and hold in context"
His plain definition of the model context protocol is that it lets agents call each other. "But effectively, it's just a way for agents to talk to other agents." A person writes a request in ordinary language and the protocol converts it into the structured call a piece of software needs — what used to be a hand-written API integration
His own example is a chief-of-staff agent running on Claude that he asks to research a prospective investment, find the founder and the website, and add the record to Airtable
The Zapier version of this connects one agent to every app the customer already uses, so a request such as pulling the five largest customers out of the CRM and drafting an invitation email runs end to end
The step beyond a request is a standing instruction. Rather than re-asking each time the same event occurs, the user tells the agent to build a workflow that watches for it and acts
He frames the market in three eras: chatbot, then assistant, then delegation. "I think the next era is like, oh, AI truly taking action, delegating it to these agents, these bots, these workflows, etc."
5. Charging By the Credit
Asked about the business model, Foster said Zapier sells a subscription with usage-based credits attached, and that this is now standard for AI products.
The pricing follows a change in who does the work. With conventional software the customer still writes the email, logs the call and does the data entry, and the subscription pays for the interface
With an AI product the software does the work, which is why consumption pricing fits. "I think with these AI tools, with things like Zapier, the value is it can do the work for you"
He described the product as closer to an entity that acts on the customer's behalf whenever something happens, so charging per event follows
6. The Middleman Question
The host put the vulnerability directly: Zapier sits in the middle, and when her team wanted PitchBook they could not find it on Zapier and PitchBook told them to use its own MCP server instead. What stops the middle layer being cut out?
Foster's answer was that the middle layer is worth paying for only if a company refuses to standardize on one AI vendor — and that refusing is the correct strategy.
He said a lot of organizations are going all in on one stack, and that the sophisticated ones are not. "So, right now, I see a lot of companies, organizations going all in on one AI stack." The reason not to: "And the reason why is that the frontier is changing super duper fast"
Whichever lab leads at a given task will not lead for long, because the labs leapfrog each other and their models are good at different jobs
Open-source models are the other pressure on any single-vendor commitment, which he described as nearly as good at a fraction of the cost
The conclusion is a stack built for swapping. "I need to build an AI stack that allows me to rotate the best tools for the best jobs" — and if every connection terminates at Zapier, swapping means repointing the connections instead of rebuilding them somewhere new
Governing the agents
The second argument for a central layer is oversight. "The second thing you really get out of it is you get a safe place to observe and govern all these agents"
A company with agent setups scattered across teams cannot see what is going in and out of its tools. Foster's warning was that fragmented setups leave "vulnerabilities hanging out all over the place" that will, in his words, "bite you in the butt"
Turning agents into code
Deploying an agent is still awkward, and a lot of them run on somebody's laptop, which has to stay open for the agent to keep working. Hosting the workflow removes that
The bigger saving is converting an agent into a deterministic workflow — a fixed sequence of code rather than a model deciding each step. "Now, why is this important is if it's running agentically the full time, you're burning tokens left and right"
Code costs less and behaves the same way every time. "Whereas a workflow, it runs on code and that code is reliable. It's low cost and it's trustworthy." His advice is to make that conversion wherever it is possible
7. The 10x Company Gap
Foster's framework for AI inside an organization is the move from individual AI to institutional AI. The host said most companies she sees have the first and not the second, and called the current state a "hot mess."
The gap he points to is between personal productivity gains and company results. "It is much harder to go look around and find the companies that are 10x productive because of this where their revenue is like 10xed or their costs are like a tenth of the price or whatever"
One person's tenfold improvement does not travel through the system. Nothing about it automatically produces ten times the customers at the other end
He reduces a company to two customer-producing functions plus overhead. "If you want to be reductive about it, there's really two things that generate customers" — research and development, which makes the product, and sales and marketing, which sells it. A very small company spends effectively all its effort on those two; a larger one accumulates coordination work that supports them without doing them
That leaves exactly three places a 10x gain can come from: the efficiency of the go-to-market engine, the quality or speed of product development, or cutting the coordination cost to a tenth and redirecting the time into the other two
He says almost nobody is approaching it this way. "And I don't see a lot of companies yet at that stage where they're sort of thinking through this in that same sort of like scientific way." The common approach is handing individuals AI tools and hoping something useful comes back
His summary of the moment is that the technology is ahead of the organizational design. "And so I think we're in this stage where it's like we have the technology, but we don't know how to build the organizations around it quite yet"
8. Inside Zapier's Rebuild
Asked what has actually changed internally, Foster said the org chart still looks much as it used to, and that he does not expect it to stay that way. The three transformation programs the company is running map onto the three targets in his framework.
A software factory, aimed at product development. "So we have a software factory that's working on building the products of the future." Engineers work on the machinery that writes the code rather than writing it directly, with the goal of getting code generation running in a loop
Go-to-market agents, aimed at customers Zapier could never afford to sell to. "A lot of that is working at our like our small business funnel where they never had sales deployed to help them in the first place because we couldn't afford to do it" — agents are being put into that funnel to see how it changes the sales engine
A company-wide program to make the business readable to its own AI, which Zapier internally calls a hive company for want of a better term. The aim is that the systems are legible enough for AI to operate the company
"And so that means like when you go into a meeting the agenda is crafted by the AI"
"When you're looking at action items, the action items are generated by the AI. Some of those action items are taken by the AI."
The target is the coordination layer, not the product teams. Work that used to mean a person carrying information between departments and connecting one stakeholder to another is being encoded into these systems
9. The Meeting-Prep Agent
Asked for a task that was not automatable before and is now, Foster picked preparing for his own customer calls. Those calls are open-ended and exploratory, and may lead nowhere.
The first pre-AI option cost the sales team its selling time. "If I ask that, I would be pulling the account team away from active deals, active prospects, things that could like actually put money on the board for them to sort of help me out with this like quote unquote like more speculative thing"
The second option was doing it himself, badly. "Maybe I put as much effort into it. Probably don't, to be honest."
The task now gets done, by an agent, for every meeting. "And so all of a sudden we've all got meeting prep agents for pretty much everything we do when nobody was really doing that before"
The general claim is that a category of work was simply skipped, because it was too hard, too expensive or impossible to do consistently well — and that agents pick up that work and leave people the parts above it
10. When Tokens Have to Pay
The host raised the cost of running AI at scale, citing Microsoft canceling Claude Code licenses over cost, and companies spending about $7,500 per employee on tokens. Both claims were the show's, not Foster's.
His answer starts from ordinary cost discipline. "Look at some point in time any cost that a company spends has to translate into ROI," he said, adding: "Like none of us show up to our jobs and are allowed to spend more money than the company earns."
He said companies deliberately suspended that test for about a year. The reasoning was that the experimental phase was worth overspending on, on the understanding that some of it would work and some would not
He thinks that period is now ending. "And I think that's what we're seeing right now is there was a lot of token maxing going on in the name of experimentation, in the name of figuring a lot of this stuff out." Companies are now sorting the results into what worked and what was waste
The two fixes he expects are budgets and code. Employees given a capped token allowance, and expensive workflows moved off live model calls onto deterministic ones
This is also the commercial case for Zapier. A workflow that cannot justify its cost today may be worth keeping at half or a tenth of it
His last point was about model selection. The wasteful pattern is sending a basic task to a top-end reasoning model: "you don't need that kind of power to do a basic task"
11. 10/10 Fun, 10/10 Anxiety
The host said it is a hard moment to run a software company — a valuation can move on a single headline, and the "SaaS is dead" argument is everywhere. She noted that Box's chief executive Aaron Levie had told the show he went back into founder mode.
Foster's own summary of the job came from another chief executive rather than from himself. "I had another CEO friend say 10 out of 10 fun and 10 out of 10 anxiety right now," he said, and agreed it fit
The fun is that everything is open to redesign. He described the moment as "a period of intense creative disruption" and said of the technology: "You get to rethink everything from first principle."
He contrasted it with calmer periods in Zapier's history, which were more stable and less chaotic, and duller: "you can say a lot about this time period, but boring it is not"
Zapier is insulated from one pressure and not from others. He said the company does not have "a ton of outside investors" and so does not face the scrutiny a public company would, but it still answers to customers and employees, whom he described as stakeholders in what the company does
The hard part is communicating a direction he cannot fully describe yet. "And so there is a lot of work you have to do to sort of make sure that you're painting a clear picture of the future when you yourself are trying to figure out like what exactly is this going to look like" — he knows the direction and not the final form
The volume of AI coverage is its own management problem. Employees cannot easily tell fact from fiction: "There's a lot of disinformation. There's a lot of people pitching their own agendas."
The other hard part is reversing course in public. When a hypothesis turns out wrong, he said the job is to call it quickly and turn the company, repeatedly — jarring, and in his account part of what makes the period enjoyable
Bonus Insights
Foster's own framing of the AI capability question was a set of rates of change rather than a level — whether releases arrive faster, whether each is better, and whether each is cheaper. He treated any one of the three as enough to matter to Zapier's business
The host adopted the friend's line as a description of the whole industry, saying that "10 out of 10 fun, 10 out of 10 angst or anxiety" summarizes most people's experience of AI right now
The host closed by naming what she took from the conversation: his theory of change for AI inside an organization, built back to first principles — the people who build the product, the people who sell it, and the support functions such as HR and finance around them. She said she preferred that to telling everyone to use AI
Foster's bottom line is that the constraint on AI inside companies is no longer the technology or individual skill, but organizational design: unless a company can name which of its three functions it is trying to improve tenfold, and can move the routine work onto cheap deterministic code rather than expensive always-on agents, its employees' personal productivity gains will not show up in its revenue or its costs.
Products, Companies & Tools Mentioned
Zapier (Foster's company: 15 years of app integrations, now repositioned as the connection and governance layer between AI agents and a company's other software, sold as a subscription with usage-based credits)
Model Context Protocol (MCP) (The standard behind the strategy. Foster described it as simply a way for agents to talk to other agents, converting a request written in ordinary language into the structured call a piece of software needs)
Claude and Claude Code (The host's team runs Claude Code with a set of agents connected through Zapier; Foster's own chief-of-staff agent runs on Claude)
ChatGPT and GPT-4 (GPT-4's launch triggered Zapier's code red — six months after GPT-3.5, obviously smarter and meaningfully cheaper)
Gemini and Microsoft Copilot (Named alongside Claude and ChatGPT as the stacks companies are wrongly standardizing on, in his view)
Airtable (The destination in his own agent example: research a prospective investment and add the record)
PitchBook (The host's counter-example — her team could not find it on Zapier and was told by PitchBook to use its MCP server directly)
Gong and Slack (The company context he says turns a generic model answer into a useful one, alongside CRM records and email)
Microsoft (The host cited it canceling Claude Code licenses over cost)
Box (Its chief executive, Aaron Levie, had told the show he went back into founder mode; Foster was asked to respond to that)
Stripe and Twilio (Both launched in 2011, and their support forums were where Foster saw the unmet integration requests that became Zapier)
Marketo (The API he was fighting with in his day job when the idea came up — his own evidence that integration was too hard)
Wufoo, AWeber, Highrise and PayPal (The apps in Zapier's first paying setup, for Mixergy's Andrew Warner)
Mixergy (Andrew Warner's interview site; he was Zapier's first customer, and his "how much money do I owe you" reaction was the moment Foster believed in the product)
Startup Weekend (The hackathon where the prototype was built and won, in a garage in central Missouri)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: