Zapier CEO: 80% of agent tasks should be deterministic code
Wade Foster says 80% of what customers do with agents should really be old-fashioned deterministic code; no code is over, and the new no code is code.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Knowledge work is moving into people's everyday tools
Foster recaps his biggest shift in thinking over the past 18 months: most people have already picked a daily AI tool — maybe Cursor, Claude Code or ChatGPT — and that is where work mainly happens. So getting context and data into that daily tool (an MCP server like Zapier MCP, for instance) is where knowledge work is heading. He splits the market in two: one camp wants to keep people on its own platform building agents, the other, like Salesforce, goes headless and lets you take it wherever you want. He clearly thinks the latter is the winning strategy, because that is what tool consumers want and where the growth is.
— Wade FosterThe strongest model gets only four in ten automation tasks right
Zapier's Automation Bench tests models on roughly 600 real knowledge-work tasks, for example ‘just signed the Meridian Core platform contract, mark it as won per the routing policy and route it to the right team, confirm the account from the account hierarchy table, convert currency if needed, and check whether there is an open support escalation’. Foster says the newest strongest model is about 40% accurate, the current high, but far more expensive than Gemini 3.7, which creates a curve of ‘how much are you willing to pay for incremental performance’. He stresses the benchmark is nowhere near saturated, and what Zapier wants to prove is that putting Zapier next to an agent produces better output than the model alone.
— Wade Foster80% of agent tasks should use deterministic code
Foster's core judgment: the vast majority of what customers do with agentic products — his figure is 80% — should really use old-fashioned deterministic code. The approach is to have the agent first decide which steps to write as code (cheaper, more reliable) and only call AI or build agents where reasoning is genuinely needed. He concedes models will keep getting better, but finds it hard to imagine a world where deterministic code is no longer more reliable and cheaper. At the same time he keeps the visual layer: people still need to see what the workflow looks like, to verify ‘is this what I want it to do’, and to use it as documentation.
— Wade FosterThe real rival is the product director who leaves in two years
Asked what kind of competitor he worries about most, Foster brings up what PG said during YC: back then the question was not ‘what if Anthropic or OpenAI does this’ but ‘what if Google does this’, and what PG wanted to instil was — you rarely fight Google directly; you face a mid-level product director who wants a promotion and will be gone in two years. He says OpenAI and Anthropic are now big companies and will be strongest on models, but they cannot do everything. On the other side, ordinary AI users have actually done nothing and may only have used ChatGPT or Gemini, so most companies are not competing with each other but with ‘do people even know what to do with these tools’.
— Wade FosterLet AI watch you work and find automation opportunities for you
Foster thinks the hardest part has always been recommendation: what should you do with this thing. This year he started running a workflow where AI looks at his daily activity in Gmail, Slack, the browser and chat, then tells him ‘you should do this differently’. Most Zapier employees plug their tools into Zapier MCP and run an automation once a week, collecting a week of work signals and producing ‘you did this and this, here is a tool I suggest you build’. He says half the battle is just getting you to react to a suggestion — the idea does not need to be perfect, 50% good is enough to start brainstorming. He confirms this will very likely be productised.
— Wade FosterThe bottleneck in AI transformation shifted from tech to people
Foster says phase one was company-wide AI fluency, and within a year nearly 100% of employees used AI daily, so technical adoption was not the bottleneck. After that the question became: how do you turn individual success into company-level production-grade workflows, and so the challenge looks more and more like a people problem — rewriting job descriptions, rethinking compensation, reorganising teams, retraining people from teams you do not need into teams you do. The people officer at the time, Brandon, was good at this and the team was at the frontier, so he was asked to help the whole company. Foster stresses this is not a ‘the people officer must own AI’ argument: a CMO or CPO could do it too; the key is to see where your bottleneck is and then find the right person.
— Wade FosterSome people burn $30,000 of tokens a month
Foster says a few individual engineers spend $30,000 a month on tokens, which is an outlier. Zapier currently sets no per-person budget, but has built tools so people can see their own spend and see the cost difference between choosing a strong model and a cheap one across different workflows. When he sees someone spending especially much, his first reaction is to talk: ‘what are you doing, I'm curious’. It turns out some are doing extremely high-output work, while others are using Fable or Astro where it is not needed. He expects token budgets will eventually become a real thing, and that part of AI fluency is some people getting a higher budget because they use it better.
— Wade FosterUsing AI is not the problem, low quality is
Foster says he is not against employees using AI to write communications; the real problem is low-quality communication: people with poor judgment can produce enormous amounts of low-quality content extremely fast and drown readers. He offers a few rules: you are responsible for what you send, and the author should spend more time than the reader; you must understand what you send and not be unable to answer follow-up questions; write requests clearly, saying whether you want a decision, feedback, or are just marking it as a draft; verify details, because AI hallucinates and may also pull up an old project from three months ago as relevant material, and summaries of summaries passed down the chain turn into misinformation. He also calls out that AI voice — ‘not this, but that’, em dashes, honest truth, load-bearing point.
— Wade FosterIn their own words · checked verbatim
I definitely think that latter is the winning strategy
Wade Foster5:21
the idea of building a no code, it, it feels antiquated to me. It's like the new co- no code is code
Wade Foster18:01
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 Foster23:02
you're going toe-to-toe with a potential mid-level product director who's trying to get a promo- promo, might be there for two years and then bounce
Wade Foster30:34
most of us, our competition isn't each other. It isn't the, the tools that you talked about. It's do people actually know what to do with these tools yet?
Wade Foster30:34
You need to look at what are your bottlenecks, what are your constraints, and go identify the person who is the right fit for that job
Wade Foster46:09
using AI is not the problem. It's like low quality. That's the fight at the end of the day
Wade Foster55:46
Figures
| Automation Bench strongest model task accuracy | about 40% | 10:20 |
| Automation Bench task count | about 600 | 10:20 |
| Share of customer agent usage that should use deterministic code | 80% | 23:02 |
| Share of Zapier employees using AI daily | close to 100% | 46:09 |
| Zapier employee count | close to 800 | 55:46 |
| Years Zapier has held customer credentials | 15 years | 55:46 |
Glossary
- MCP / Model Context Protocol
- An interface standard for plugging external tools and data into your everyday AI tool.
- headless
- A product that does not require users to stay in its own interface; its capabilities can be called from any tool.
- Automation Bench
- Zapier's evaluation set that tests models' automation ability on real knowledge-work tasks.
- Pangram
- A detector that judges the probability a piece of text was generated by AI.
- harness
- The everyday interface wrapped around a model that provides context and tool calls.
How to listen
Founders building agent or automation products, engineering leads weighing model price-performance, and managers pushing AI adoption inside their companies.
Sponsor reads and the closing recommendations can be skipped; the rest has even information density.