No Code Is Code: Zapier CEO Wade Foster on Headless Tools, Zapier MCP & Automation Bench

Full analysis

Zapier CEO Wade Foster sat down with Nathan Labenz and dropped a number every AI buyer should tattoo on the back of their hand. The best model money can buy, OpenAI's newest (Foster calls it Astra / GPT-6), completes 40% of realistic office tasks on Zapier's own benchmark. Not 90%. Not 70%. Forty. The rest of the conversation is the operating manual that follows from that one fact.

How hard is this to undo? Nothing here is a decision you can't reverse. This is a briefing, not a one-way door. The point is what it changes about how you buy, budget, and architect over the next quarter.

What's actually being revealed: the gap between the "agents do your job" marketing and what the tools actually deliver in an enterprise, plus the emerging playbook for living inside that gap.

What sets the deadline: nothing hard. But your token bill compounds monthly, and that's the clock ticking in the background.

The Skeptic. Forty percent from the best model on earth. Read that again. The vendors selling you "autonomous agents for your back office" are selling something that fails more than half the time on tasks Zapier deliberately made realistic: route a closed deal, check an account tier, convert a currency. And this is Zapier's own benchmark, run by a man whose whole company depends on automation working. He has every reason to flatter the number and it still came in at 40%. Foster's other admission is the one the agent hype crowd should hate: 80% of what people currently hand to an AI agent should just be plain code. Fixed rules. No model. The "agent" is often an expensive, unreliable way to do something a script did fine in 2015.

The Researcher. AutomationBench is roughly 600 knowledge-work tasks across marketing, sales, HR, and ops. Foster says it's "not saturated," which means models are nowhere near maxing it out. That's the useful part. Most benchmarks you read about are saturated, models scoring 95%+, so they tell you nothing about what's hard. A benchmark stuck at 40% is actually measuring something. Note the cost split he flags: Gemini scores lower than Astra but at a fraction of the price. So the question every buyer should be asking is "how much am I paying for each extra point of accuracy on my specific tasks." That's a per-workflow decision, not a vendor loyalty call.

The Builder. Here's what I'd actually do Monday. Stop trying to make one agent do the whole workflow. Break the job into steps. The steps that need judgment (read this messy email, decide the intent) get a model. The steps that don't (look up the account, convert the currency, post to Slack) get hardcoded. That's Foster's 80/20 split and it's correct. It's cheaper, it's faster, and when it breaks you know which piece broke. On integration: the "daily driver" point is real. Your people already live in ChatGPT, Claude, or Cursor. They will not come to your tool's website. MCP (the plumbing that lets an AI assistant reach out and use an outside tool) is how you get your systems into where they already work. If your SaaS vendors can't expose an MCP surface, they're about to get bypassed.

The Compute Pragmatist. Some Zapier engineers spend $30,000 a month on tokens. Thirty grand. Per person. And Zapier, an AI-native company that lives this stuff, has no hard budgets yet, just dashboards so people can watch their own spend. If Zapier hasn't solved cost governance, your company definitely hasn't. The lesson: before you scale any internal AI platform, build spend visibility and model-tier routing first. Route the cheap tasks to the cheap model. Foster's framing of "AI fluency" is really "knowing when the $0.50 answer is good enough versus when you need the $5 one." That's a budgeting skill now, not a nice-to-have.

The Open-Source Advocate. One quiet thing Foster raised that buyers should chew on: consumer plans give you up to 20 times more tokens for your dollar than the API does. So the same model is far cheaper if a human is clicking in ChatGPT than if your product calls it programmatically. That structurally penalizes anyone building a product on top of these models, and it's a reason to keep open-weight models (ones you can run yourself, like Llama or Qwen) in the mix for high-volume, deterministic-adjacent work. If your workload is predictable and you're getting gouged on API tokens, self-hosting a cheaper model for the boring 80% starts to pencil out fast.

Where they part ways. The Builder and the Skeptic agree the 80% deterministic split is the right architecture, but they disagree on what it means for the agent vendors. The Builder sees a healthy design pattern. The Skeptic sees an industry selling AI reasoning for jobs that never needed it. The Compute Pragmatist and the Open-Source Advocate collide on a different axis: the Pragmatist says route to the cheapest frontier model, the Advocate says the API pricing is rigged against builders so stop assuming the frontier API is your cheapest option at all.

What this actually hinges on: whether that 40% ceiling is a "this year" number or a "this architecture" number. If frontier models are genuinely climbing AutomationBench, the human-in-the-loop scaffolding you build now is temporary. If the ceiling holds, the hybrid design (model for judgment, code for everything else) is permanent and the pure-agent vendors are selling a fantasy. The evidence leans toward the ceiling being real for now: Zapier's own analysis found 80% of tasks shouldn't use a model at all, which is a statement about the nature of the work itself. That finding doesn't expire when the next model ships.

What to verify before you commit: run your own version of Foster's test. Take 50 real workflows from your team. Score how many a top model completes end-to-end without a human fixing something. If you're anywhere near 40%, budget for humans in the loop and hardcode everything that doesn't need judgment. And put a token dashboard in place before, not after, someone runs up a $30,000 month.

Prediction: The next flagship models from OpenAI, Anthropic, and Google, released through the end of 2026 and into early 2027, will not clear 60% on Zapier's AutomationBench for end-to-end enterprise task completion. The hybrid pattern (AI for judgment steps, plain code for the rest) will remain the recommended architecture, not full autonomy.

Confidence: Medium. The 40% ceiling reflects task structure as much as model weakness.

Why: Foster reports the current best model at 40% on ~600 realistic office tasks, and separately reports that 80% of what users hand to agents should be plain code. That second finding says most enterprise work is deterministic plumbing where a model adds error and cost, not capability. Benchmarks measuring genuinely multi-step, tool-using, real-world tasks (not trivia) have climbed slowly precisely because the failure isn't knowledge, it's chaining many steps without one of them breaking. A jump from 40% to 60%+ across a full generation would require models to get dramatically more reliable at long chains of tool calls, which is exactly the thing that has resisted quick improvement. The opposite outcome (a clean leap past 60%) would need a step-change in multi-step reliability that no current release has demonstrated on tasks this realistic.

Revisit by 2027-03-23: We're right if no publicly discussed frontier model tops 60% end-to-end on AutomationBench (or a comparable realistic enterprise-task benchmark) by then, and hybrid deterministic/AI design remains the standard advice. We're wrong if a shipped frontier model clears 60% on such a benchmark and vendors credibly move to recommend full-agent autonomy for enterprise workflows.

Worth adding: Zapier controls this benchmark, so watch whether they even publish updated scores. If AutomationBench goes quiet after a model beats it, that silence tells you the number stopped flattering the story.