How to Get a ‘Transformative’ AI Fluency Rating as a PM, with Wade Foster | CEO of Zapier

Context

Wade Foster co-founded Zapier and wrote one of the first public AI fluency rubrics, which many companies later copied. Zapier published a V2 of that rubric on March 31, 2026, and has already tweaked it internally since. In this episode Aakash Gupta shows Wade three versions of the same PM project, an "IDE harness" that connects Claude Code to Zapier's integrations, and Wade grades each one live. It matters to PMs because AI fluency is now part of performance reviews at Zapier, Shopify, Meta and Microsoft, and the episode shows exactly how a CEO reads PM work against that bar.

The Big Idea

In 2026, a well-written AI-assisted PRD is the floor, not the bar. What earns a high AI fluency rating is going further: a prototype, real customer evidence, and eventually a system of agents that does the work while the PM audits it.

Wade's point is that AI fluency and good product thinking can no longer be separated. The same AI that makes the extra work cheap also makes skipping it look old-fashioned.

Key Insights

Zapier's four-level fluency rubric

Zapier grades AI fluency per role on four levels:

• Unacceptable: the bar Zapier will not hire at.
• Capable: using AI for simple tasks like summarizing, drafting, lookups and basic PRDs. In 2026 this is table stakes at most companies.
• Adoptive: a structured, reusable approach to specs, prototypes and research. Using AI to mine quant and qual user insight at a scale humans couldn't manage before, and generating many working solutions quickly. PMs here also stretch into adjacent skills (data analysis, marketing, sales) to "80 or 90% good".
• Transformative: building systems, for example a pipeline of agents that takes in customer feedback, writes specs, prototypes and ships. Wade compares this to "software factories": building the system that builds the product.

Most PMs should sit at Adoptive

Wade does not want everyone at Transformative. Zapier's target is PMs who sit at Adoptive most of the time, have some normal gaps back at Capable, and can point to one Transformative example. His reasoning: someone living at Transformative full time is constantly tuning their system instead of getting work done. He also frames the rating as "most of the time I'm this", not a fixed label on a person.

Transformative means an n of one

Wade's bar for Transformative is very high: you are the first person doing a new thing and figuring it out as you go. On a team of 30 to 40 PMs, he agreed that means roughly one Transformative rating per half. He rates himself Adoptive in most places and admits his Transformative tactics are mostly borrowed from his team.

Never stop at the PRD

Aakash's first artifact was a ChatGPT-written PRD, iterated a few times, with an overview, problem, target user, features, MVP and a north-star metric. Wade liked the product intuition but called it "a little old-fashioned" for 2026, somewhere between Unacceptable and Adoptive. His two fixes:

• Build the prototype. The goal is not production code. It is that "seeing is believing". A picture is worth a thousand words, and a prototype is worth ten times that.
• Go find the evidence. Sift Gong calls, Zendesk tickets, subreddits, LinkedIn and X for people complaining about the problem or asking for the solution. This also tests what is uniquely Zapier's to own, versus what Claude Code could just do alone.

Evidence got cheap, so buy-in moves later

In the past a PM would ask for buy-in before spending two weeks gathering evidence, and couldn't build a prototype at all. Now both are fast, so the buy-in conversation happens after you have a mountain of proof. That makes a yes more likely, or it lets you shelve a bad idea yourself before anyone else spends time on it.

The second artifact: prototype plus evidence

Aakash's second version opened with a working prototype, then a tighter PRD with evidence (agent tooling came up unprompted in 34 of 47 Gong calls, 312 Zendesk tickets tagged MCP or AI agent, a six-partner pilot where five of six enabled a pre-approved write action). Wade rated it solidly Adoptive, the caliber he'd be happy to see across Zapier. His one push: why wait for a Q4 beta that is five weeks away? AI tools let you get from zero to design partners much faster.

Be ahead of your CEO in your discipline

Wade's rule of thumb: inside your own discipline, your AI fluency should be ahead of your CEO's. PMs breathe this work every day and he doesn't. If he spots a technique that is obvious to him and new to the PM, "that makes me nervous".

The job shifts to auditing the factory

Aakash's third artifact was a five-layer agent stack (skills, parallel sub-agents including a "skeptic", tools and MCP, deterministic hooks that reject unsourced memory, a sourced memory layer) plus a nightly "triage swarm" and an auto-build chain from two-page PRD to draft code. Wade said this is starting to hit Transformative. In this model the PM's job is not to do each task. When a factory produces a bad widget, you don't fix the widget by hand. You ask why the machine made it and fix the machine.

Nobody has solved where the gates go

Wade said there is no consensus yet on how to design these loops. The open question is where to put the human gates. Removing an early gate lets agents run on everything, but burns tokens on wasteful work. Keeping it slows the system but saves cost. He sees many teams "scratching their heads" on this.

Not every PM should build a factory

At a small company, the only PM has to build the system. At Zapier's size, with 30 to 40 PMs, Wade wants one shared "product factory" (or a few variants for different jobs), not 30 personal ones. A few PMs redesign the system while others keep running the standard playbook and act as user testers for the new way of working.

Mental Models & Frameworks

Slop cannon versus turbo brain

A 2x2 from Dan Hockenmeier: judgment on one axis, AI use on the other. Good judgment plus AI makes a "turbo brain". Poor judgment plus AI makes a "slop cannon". Wade agreed and added a twist: the slop cannon is harder to work with than someone with poor judgment who uses no AI, because "at least you're quiet". People without judgment can generate far more output than others can review.

You can delegate work, not accountability

Zapier's internal line on slop. Sending unreviewed AI output to a manager hands your judgment to them, which is the opposite of what you want when moving up. Aakash added: people above you want filtered, quality information, not "what do you think of this?"

Label the effort level

Wade's etiquette for sharing AI-assisted work: say up front how much you reviewed it. For example, "AI drafted this, I skimmed it and removed the obviously bad parts, sharing for speed", versus "I stand by every sentence". Sharing rough work is fine when speed matters. The painful case is passing off zero-effort output as polished, which makes the reader do your job.

Deterministic where you can, agentic where you must

Wade runs his personal agents as Zapier workflows rather than inside Claude Code or Codex. Most steps (fetching calendar events, posting to Slack) are plain code that runs the same way every time and burns no tokens. Only the step that needs reasoning, like drafting the digest, calls a model. The benefits are reliability, much lower cost, and no need to keep a laptop open.

Decision Principles

Principle: Match quality bar to audience

• When: deciding how polished AI-assisted work needs to be before sharing.
• Why: Wade said the standard for an external customer-facing product is different from a quick internal task. What matters is that your judgment is switched on and the reader knows what to expect.

Principle: Design software for agent users

• When: building B2B software.
• Why: Wade argues the user of work software is increasingly the agent, not the human. If your product can't run headless (APIs, MCP servers), "you're swimming against the current". He points to Salesforce going headless. He sees consumer software as a different case.

Case Studies Mentioned

Zapier's company brain

Many Zapier staff had built personal "second brains": markdown files of goals, working preferences, design system and strategy that a coding agent reads. One product leader argued this should be a company brain, not an individual one, and built a V1 for product people. Wade said every product person at Zapier "got a jet pack" when it launched. It happened right after the late-2025 model jump, when engineers largely stopped writing code by hand.

Wade's personal robot staff

• Morning brief (6 a.m.): reads calendar, to-do list and linked docs, then posts to Slack. Wade calls it the easiest place to start but only modestly valuable.
• Scribe (5 p.m. daily recap): loops over unanswered email, saved Slack messages and Granola meeting transcripts, then drafts follow-ups and action items. It cut his end-of-day cleanup from about two hours to about 15 minutes. This is his favorite.
• Agenda builder: scans a week of meetings, Slack, email, AI chat sessions and metrics to propose exec and board agenda items. It beats people's "pet items", for example by flagging a bad conversion rate nobody raised.
• CEO CRM: a vibe-coded CRM (despite Wade saying you shouldn't vibe-code one) that runs every Saturday night and leaves 10 to 20 drafted emails to enterprise customers, each with a reason: an upcoming renewal, or usage up or down. The hardest part was getting the drafts to sound like him, and he has a separate agent tracking which subject lines work.

Zapier after the no-code "apocalypse"

Zapier hasn't raised primary capital since 2012, so its 2021 $5B mark matters little to Wade, and he says the business has grown since then. His view: "the new no-code is code". Zapier is exposing its triggers, actions, runtime and governance to coding agents. Its bet against newer competitors (Lindy, n8n, Cowork) is running automations deterministically in the cloud, with an agent handling setup and recovery when things break. It also still has 9,000+ integrations that no competitor has matched.

Anthropic's focus on code

Wade used Anthropic as an example of focus winning in a crowded market. OpenAI chased many things early, while Anthropic chose to be great at code and stood out, which pushed OpenAI to sharpen its own strategy. Wade's lesson: in a huge market with many winners, you have to choose who you are for.

Practical Application

Grade yourself on Zapier's rubric

Take Zapier's four levels (Unacceptable, Capable, Adoptive, Transformative) and mark where your last three projects actually sat. Aim for mostly Adoptive with one Transformative example to point to, which is the distribution Wade wants from Zapier PMs.

Ship a prototype with every PRD

Before your next product review, paste the PRD into a coding agent and build a clickable prototype, even a rough one. Wade said he would do this himself with any PRD handed to him, so arrive with it already done.

Mine evidence before asking for buy-in

Before pitching a new bet, have AI sift support tickets, sales call recordings and public forums (Reddit, LinkedIn, X) for the pain point. Aakash's second PRD cited counts like 34 of 47 sales calls raising the topic unprompted. Numbers like that turn a pitch into a decision.

Add an effort label when sharing

When you share an AI-assisted doc, add one line saying how much you reviewed it: a quick skim for speed, or fully reviewed. This follows Zapier's line that you can delegate the work but not the accountability.

Start with a daily recap agent

If you are new to agents, build Wade's "scribe": at the end of each day, loop over unanswered email, saved Slack messages and meeting transcripts, and draft follow-ups. It is more valuable than a morning brief. For Wade it replaced about two hours of end-of-day cleanup with about 15 minutes.

Questions to Consider

• If your CEO reviewed your last PRD, would they spot an AI technique you should have used but didn't?
• Where in your team's workflow would you put human gates if agents drafted specs, prototypes and first-draft code, and what would removing each gate cost in wasted tokens versus speed?
• When you last shared AI-assisted work, did the reader know how much you had personally reviewed it?
• Could your product be used headless, by an agent through an API or MCP server, instead of a human clicking through the UI?

Bottom Line

A polished AI-written PRD now reads as old-fashioned. Zapier's CEO expects PMs to arrive with a prototype and customer evidence, and to spend most of their time at Adoptive fluency with judgment switched on. Transformative is rare: it means building the system that does the work, then auditing that system instead of doing each task by hand.

Resources Mentioned

Tools & Products

People to Follow

Wade Foster

Co-founder and CEO of Zapier, which reached a $5B valuation on roughly $1.4M of venture money. He wrote the AI fluency rubric that many companies modeled their own on.

Dan Hockenmeier

Creator of the "slop cannon versus turbo brain" framework on judgment and AI use that is discussed in the episode.

Notable Quotes

"You can delegate the work, but you can't delegate the accountability." (Wade Foster, quoting Zapier's line on AI slop)

"A picture says a thousand words. Well, a prototype is that times 10." (Wade Foster)

"If I'm spotting things that are obvious to me and not to you, I'm like, ooh, that makes me nervous." (Wade Foster)

"The new no-code is code." (Wade Foster)