How I AI: Zapier CEO Wade Foster's Playbook for AI-Powered Recruiting and Culture Building

Wade Foster, co-founder and CEO of Zapier, uses AI in recruiting, culture work, and company operations rather than limiting it to product features.

Wade calls it a delegation trap when leaders write an AI memo and leave everyone else to figure it out. He instead runs hackathons, shares examples, and uses the tools himself. One Zapier value captures the approach: Don't be a robot, build a robot.

In this episode of How I AI, Wade demonstrates how he turns authorized meeting transcripts into a description of working culture, evaluates interview evidence with an advisory agent, and searches X for candidates through Grok.

Describe working culture from meeting transcripts

The detailed setup is in Wade's unspoken-company-culture workflow.

The valuable output is not a clever summary of company values. It is a set of observed behaviors with enough evidence that leaders and employees can say where the analysis is accurate, where the sample is thin, and which tensions should remain visible. Wade uses the document as material for a hiring rubric only after that review.

Wade used Granola to turn patterns from recorded meetings into a concrete description of how Zapier makes decisions and communicates.

Move from stated values to observed examples

A values document describes intent. Meeting examples can show what those values look like in practice and provide evidence for a hiring rubric.

Use Granola's unspoken-culture recipe

Granola records and summarizes meetings under the organization's recording, consent, retention, and access policies. Wade uses those existing transcripts as the source for the analysis.

1. Select a representative set of authorized meeting transcripts rather than collecting conversations solely for this analysis.
2. Run the Granola recipe named Build the unspoken company culture handbook.
3. Review the generated document against the underlying examples and correct anything that does not match how the company works.

When I first ran this, I was shocked. We spent a lot of time thinking about our culture, writing about our culture... but as I read through it, I was like, wow, this actually gets at the specifics in a way that even I hadn't figured out quite how to do.

Turn the culture description into interview criteria

Wade then asks a model to translate the culture description into criteria that can be used during interviews.

"Hey, can you take this unspoken culture and actually generate a set of scoring prompts for how to evaluate somebody in an interview against these traits that match Zapier?"

The result is an input to a hiring rubric, not an automatic employment decision. People still define the criteria, review the evidence, and make the decision.

Build an advisory interview-evaluation agent

The full configuration appears in the interview-evaluation agent workflow.

Wade wants a consistent second opinion, not an automated hiring decision. The agent must point to interview evidence, compare it with a role-specific rubric, and make uncertainty legible. Its recommendation gives the interviewer another view to interrogate; it does not turn a transcript into an employment verdict.

Wade built an agent to review interview transcripts, compare evidence with the role and company values, and provide a structured second opinion.

Apply a consistent evidence structure

Zapier Agents gives Wade a repeatable format across roles, but it does not remove bias or replace the hiring team's judgment.

Configure the interview agent

The agent uses the following structure:

1. Trigger when an authorized Granola interview note is added to the designated New Interviews folder.
2. Provide instructions that define the evidence to consider and the output format.

You are an expert hiring evaluator at Zapier. Your task is to review the interview transcript and notes provided by Granola. You're reviewing the job description provided as a knowledge source and Zapier's company values provided as a knowledge source to determine whether a candidate should advance in the hiring process. You want to evaluate the candidate's functional expertise, their values alignment... your goal is to recommend yes, no, or maybe to this candidate and provide your reasoning. Give me three to five sentences on why you are recommending this. Then go ahead and email me the evaluation.

1. Ground the review in two documents:

• The Zapier values rubric.
• The job description for the role.

1. Send Wade an email containing the recommendation and supporting reasoning for human review.

Revise the agent with Copilot

During the demo, I suggested two additions:

• Evaluate whether the interviewer gathered evidence for every part of the rubric.
• Put the advisory yes, no, or maybe recommendation in the email subject for triage.

Wade used Copilot in Zapier Agents to update the instructions in plain language during the recording.

The result: a structured second opinion

The email includes a recommendation and reasoning tied to the job description and values. Wade remains responsible for reviewing the evidence and making the hiring decision.

Source candidates through Grok and X

The sourcing sequence is documented in the under-the-radar talent workflow.

The search is useful because public work can reveal communities and candidates a conventional title search misses. Every result still needs manual verification against transparent, job-related criteria. Noise, bots, stale profiles, and protected-trait proxies are reasons to inspect the list, not reasons to let the model rank people invisibly.

For a social-media role, Wade used Grok's access to X to search for people active in online communities who might not appear in a conventional recruiter search.

Look for demonstrated participation in online communities

Wade wanted candidates who understood internet culture and had already posted tutorials or built an audience.

Refine the search through conversation

He started with a broad search, inspected the results, and adjusted the criteria.

Begin with a broad role and community query.

help me find posters on X that are fans of Zapier, no code, agent building, automation, and related topics. I want posters that share tutorials and education related ideas. These posters should have modest followings... I'm looking for diamonds in the rough... We are on a budget, so look for folks outside the Bay Area. Give me 10 ideas.

Remove obvious bots and narrow the location or type of work after reviewing the first results.

• let's do not a bot
• give me people with real faces as for profile photos
• let's do United States located folks

Change the source criteria when the role calls for a different medium, such as video.

how about finding 10 YouTubers

The result: another candidate-discovery channel

The results included noise and bots, but also surfaced regional no-code and creator communities. Wade said he had used the same approach to search for specialized AI talent.

A connected operating loop

The three workflows connect observed examples of culture to interview criteria, structured candidate review, and sourcing in communities relevant to the role.

AI increases the amount of material a leader can review, but the company still owns consent, criteria, access controls, and employment decisions.

The sequence also makes the handoffs inspectable. A meeting transcript supports a culture claim; an accepted culture claim becomes a rubric; an interview transcript supplies evidence against that rubric; and a public profile can explain why a person entered the sourcing conversation. When the evidence is weak, the system should expose the gap instead of manufacturing confidence.

That traceability matters most in hiring, where a concise answer can look authoritative while hiding a bad assumption. Wade's system is useful only when reviewers can return to the source and disagree.

Wade's playbook is a leadership loop: observe how the company actually works, translate the accepted patterns into explicit criteria, use those criteria to review evidence, and keep looking for talent in places the standard funnel overlooks. AI expands the material a leader can examine. It does not transfer responsibility for consent, fairness, or the final decision.