Zapier CEO: Best models still fail 60% of real work tasks

Wade Foster says most knowledge workers have already picked a single AI "daily driver," and the winning strategy for tools like Zapier is to plug into that driver rather than compete for it. He also argues that despite frontier-model gains, 80% of what companies run through agents should really be deterministic code, not AI reasoning.

• Headless tools thesis — Foster says knowledge workers now settle on one daily driver — Claude Code, ChatGPT, Grok, or Cursor — and tools like Zapier MCP win by bringing context into that tool rather than pulling users onto a separate platform.
• AutomationBench score — Zapier's AutomationBench, roughly 600 knowledge-work tasks spanning marketing, sales, HR and operations, put model Astra at a new state-of-the-art of about 40% task accuracy, while Gemini 3 performs nearly as well at a fraction of the cost.
• 80% should be code — Foster estimates that 80% of what people delegate to agents today would be better handled by deterministic code, with AI reasoning reserved only for the steps that truly require it.
• Support troubleshooting — Zapier's auto-email support program spins up five independent agents to evaluate a failure, and when four of five agree, Foster says that's usually a reliable signal of the actual problem.
• No-code is code — Foster argues visual, box-based Zapier building is largely obsolete; users now describe workflows to an agent, which writes deterministic code and only invokes AI where reasoning is required.
• Competition framing — Citing YC's Paul Graham, Foster says the real competitive risk isn't OpenAI or Anthropic building everything, but a 'sea of sameness' among startups — while most everyday users still haven't adopted any AI tool beyond ChatGPT or Microsoft Copilot.
• Weekly self-review — Foster personally runs a weekly automation that reviews his Gmail, Slack, browser and Cursor activity and proposes new tools or workflows to build, and says Zapier is likely to productize this for customers.
• Pricing shift — Foster calls seat-based software pricing 'dead or dying,' predicting a split between usage-based pricing for commodity AI and outcome-based pricing, such as per resolved support ticket, for more specialized enterprise products.
• AI transformation lead — Zapier put chief people officer Brandon in charge of AI transformation not because of a general rule but because the company's bottleneck shifted from technical AI adoption, already near 100% employee usage within about a year of ChatGPT's launch, to people issues like job descriptions, compensation and reskilling.
• Public-by-default Slack — Zapier pushed communications further into public Slack channels partly because AI agents perform better with more visible context, keeping HR incidents and active security vulnerabilities private but sharing post-incident learnings.
• Security posture — Foster says Zapier has held large volumes of user credentials for 15 years, and while frontier-capable models make vulnerabilities easier to find, they also make them easier to patch, citing the Hugging Face incident as a 'sci-fi' warning sign.
• Token spend — Foster noted some Zapier engineers spend as much as $30,000 a month on tokens; the company has no formal budgets yet but built self-monitoring dashboards and expects token budgets tied to individual AI fluency to become real.
• Co-authorship standard — Foster says he has no problem with AI-assisted writing or high Pangram AI-detection scores, but insists people must own, understand and verify anything they send rather than just forwarding an AI draft unedited.

In their words

Astra came out GPT6. It's the new state-of-the-art on that model. It performs about 40% of the tasks accurately which is the highest that there is.10:29

The idea of building a no code it feels antiquated to me. It's like the new co no code is code.17:57

When we look across like even our own a like in customers usage of agentic products the vast majority of what people are using an agent for 80% in fact they probably should be using actually old-fashioned deterministic code.14:54

most folks seem to have adopted their own daily AI driver tool.7:00

Disclosure · Wade Foster is CEO of Zapier and the interview promotes Zapier's MCP, agents, and AutomationBench products; the episode also carries paid sponsorships from Mercury, Athena, OutSystems and Anthropic.

One thing to add — One thing to add — the 40% AutomationBench figure is a useful reality check against more triumphant AI coverage, since it comes from Zapier's own benchmark of realistic enterprise tasks rather than an academic leaderboard. Readers should weigh Foster's incentive to frame Zapier's tooling as the thing that closes that gap, since the AutomationBench v2 results he references, meant to show exactly that lift, haven't been published yet.

One thing to try tonight
Install Zapier MCP (or a similar connector) in your daily-driver AI tool tonight and ask it to review a week of your Gmail/Slack/browser activity to propose one workflow to automate.