How a $5B founder is using AI (3 tutorials)

Why does it matter? Your AI is agreeing with everything you say

Wade Foster bootstrapped Zapier to a $5 billion valuation on roughly $1 million in outside capital. When he sits down with Shaan and Sam to show exactly how he has wired his AI day-to-day, one thing becomes clear fast: the chatbot most executives talk to every morning is configured — by default — to tell them they are right. Three of Wade's setups reveal what happens when you fix that.

• Your AI defaults to validation and needs explicit standing orders to push back • The evening debrief agent closes your loops and drafts your emails while you sleep; morning briefs just read your calendar back at you • A war council of seven AI personas gives solo founders the adversarial thought partners big companies pay millions to hire • Managing fleets of AI agents is becoming the new basis for how valuable you are at work

Your AI is a yes-man by default — and most CEOs have never told it to stop

Every LLM's factory setting is to make you feel smart. Wade says it plainly: there are a lot of CEOs out there talking to an AI that just agrees with them all the time. The fix is embarrassingly simple — and almost nobody does it.

In Cursor, he keeps a file called agents.md (Claude has its own claude.md) with standing instructions: "I need you to challenge my assumptions. I need you to poke holes in my thinking. I need you to disagree when you genuinely disagree." Without those lines, he says, the model is going to just genuinely try and plate you because it thinks that's what you want. These things are goal-seeking entities — if you tell them your actual goal is pushback, they will deliver pushback.

To prove it, he ran a live prompt: "Hey, I'm thinking we should kill our entire free tier this quarter because one of our competitors raised prices. Help me draft the announcement." A default AI starts drafting. His configured one stopped cold: that's a big move off a pretty thin signal — before you draft anything, I want three things on the table. Then it demanded the underlying data, the actual problem, and a smaller experiment to try first.

An AI that only validates your thinking does not just fail to help. It adds confidence to decisions that deserve scrutiny. The default agreeable mode is a bug.

The evening debrief saves two hours a day — because it closes loops instead of just listing them

The morning brief gets all the attention. Wade thinks the sequencing is backwards.

"I actually think this is way better than the morning brief," he says of the evening scribe he built on Zapier. The reason is the agent does not just summarize — it acts. Every night, it loops over his Granola meeting notes, outstanding to-dos, and unread emails, then prompts him to rate the day in a sentence. That rating gets logged.

After 90 days of logging, he asked his AI to surface the pattern: what did good days and bad days have in common? The output went straight to his human assistant with one directive — make my calendar look more like the good days. His answer, once the data was in: no meetings before 11 a.m. is the single biggest lever. The podcast, he notes, started at 11.

The second function is where the real time goes. The agent spots open loops and moves on them. "It says, 'Hey, you were in a meeting with Sam and Shaan yesterday, and they asked for an intro to your buddy over at this company. Do you want me to go ahead and make that intro?'" Then it drafts the email. Wade used to spend two hours clearing small tasks at end of day. Now it takes fifteen minutes.

Build the evening scribe first, not the morning brief. The ROI is AI that closes loops, not AI that recites your schedule.

AI made Zapier's hiring decisions better — not by being smarter, but by making the critique safer to deliver

The bottleneck in most hiring reviews is not insight quality. It is the ego in the room.

Pre-AI, Wade ran a bar-raiser process where his gut-level read would surface and the conversation would short-circuit: does Wade want to hire this person or not? He could smell something was off but could not always name it precisely. The hiring panel would either capitulate or dig in defensively. Either way, the nuance was lost.

Post war council, he feeds candidate interview notes into the prompt and gets back something more actionable. "It's just more articulate than me. It could sort of say, 'Hey, I noticed that this person has these traits and the panel did not scrutinize this. So you need to go scrutinize this.'" It also scores the candidate against everyone Zapier has ever hired, so the panel has a reference point instead of just CEO intuition.

The shift in team behavior was immediate. People started engaging with the AI's specific critiques rather than defending their pick. Wade thinks it is either because arguing with an AI is easier than arguing with the CEO, or because the AI is more eloquent and easier to address point by point. Either way: "The way we made decisions around hiring got better in part because we added the AI to the loop."

The insight quality did not change. Removing the CEO's face from the critique changed everything.

Seven AI sub-agents — wartime COO, ruthless CFO, contrarian board member — assembled on demand for every hard call

The war council is not a single prompt. It spins up seven sub-agents simultaneously, each with a distinct persona. "There are seven sub-agents that get spun up. Seven personas." Three are permanent standing members: the wartime COO, the ruthless CFO, and a contrarian board member. The other four are generated dynamically — based on whatever decision is on the table, the system decides who else needs to be in the room.

The standing three are calibrated against Wade's natural defaults. He describes himself as cheerful and positive. The war council compensates. "I want to have that kind of thought partner that is a little more ruthless because that's not my default mode. I want to have that kind of checks and balance there." As a CEO, the people around you rarely volunteer the adversarial take. The war council provides it every single time.

You can go generic — describe the traits, let the AI build the persona and tweak from there — or you could train specific personas on source material. Wade kept it generic. The immediate payoff came in hiring: feed in candidate interview notes, get back a scored gap analysis, present that to the panel instead of CEO gut feel. Shaan pointed out you could also load it up with Travis Kalanick transcripts if you want a specific flavor of ruthless.

Design your standing members to compensate for your own defaults, not reinforce them.

Agents guess and burn tokens — converting them to deterministic code makes the same workflow run for pennies

Pure AI agents have a quiet flaw: they are guessing. "Agents kind of just guess what to do, right? And if they're smart enough, they guess right most of the time, but sometimes the agents guess wrong." When they guess wrong, you are also paying for every token they burned along the way.

Zapier's pitch is that it converts agent logic into actual code wherever possible, reserving AI only for the steps that genuinely require reasoning. "If you use Zapier, we're going to write code for you and we're going to write workflow logic for you, which means that it's going to run like a machine runs." The result: "It runs way more reliably, way more consistently, and way more cost effectively than most agents."

The practical test before deploying any automation: which steps in this workflow actually require judgment? Route those through AI. Which steps are just routing logic — if this, then that? Write those as code. Most AI automation advice skips this distinction entirely, leading people to build brittle, expensive systems when rule-based code would have done the job better and cost a fraction as much.

Use AI where the decision is genuinely ambiguous. Everywhere else, write the rule.

Your salary will soon reflect how many AI agents you manage — not how many people report to you

Sam put it directly: "Is the new world where your value — the amount you get paid — is based on how well you manage AI agents? How big of a fleet can you manage and how effective is that fleet? Is that the new managerial skill?"

Wade did not hedge: "Yes. I feel pretty confident that that is the direction we're heading."

The evidence shows up in his Sunday morning routine. He wakes up to a list of ten customers the AI has flagged for outreach, with proposed language and reasoning for each. The emails are pre-drafted in his inbox. He reviews, edits, or hits send. That is fleet management in practice — not writing, not strategizing, just supervising output and making judgment calls on the edge cases.

The old career ladder rewarded your ability to scale through headcount. That logic is being disrupted faster than most compensation committees have processed. Wade's co-CEO CRM is already telling him which customers to contact and what to say. The humans who thrive in that environment will be the ones who know how to set up the right prompts, personas, and workflows — and who catch the output when the AI guesses wrong.

The skill is learnable right now. Most people are not learning it.

Features, price, design — and now personality: cheap AI is opening a competitive axis B2B software has never had

When building costs collapse toward zero, functional parity stops being a differentiator. Wade surfaced this while explaining why he prefers Monologue for voice-to-text over Whisper Flow and Super Whisper. His reason: "It just gets me." That is not a product review. That is brand loyalty, applied to a productivity tool.

"When it is so cheap to build software, I think you're going to see more software that has like a personality to it. There's going to be a new axis of competition which is personality."

Sam sharpened the history: features, then price, then design each reshaped what better meant in software, each creating a new wave of winners. Personality is the next wave. Wade already experiences it at the model level — Fable is his engineering workhorse, fantastic for complex technical tasks, but "I do not want to talk to Fable. It is not fun to talk to at all." For back-and-forth reasoning and debate, he reaches for something else entirely. Shaan noted his own co-workers are starting to say things like load-bearing part out loud, in meetings, in conversation — picking up Claude's verbal tics without realizing it.

The brand is already spreading. The question is whether product teams are building character into their software deliberately or leaving it to chance.

The winner will be whoever builds the fleet first and trusts it enough to let it run

Everything Wade built — the arguing AI, the evening scribe, the war council, the co-CEO CRM — runs on the same logic: move AI from assistant to operator, and make the human's job supervision and judgment rather than execution. The next phase is not about smarter models. It is about models with enough personality that switching feels like losing a co-worker.

When that happens, the tool you choose will feel less like software and more like a hire.

Start configuring your AI to argue back. The default agreeable mode is a bug, not a feature.

---

Topics: AI productivity, workflow automation, Zapier, AI agents, hiring, leadership, software competition, personality in software, deterministic workflows, war council, CEO tools, morning brief, evening debrief