# The Most AI-Pilled CEO We Know

[Watch on tuber](https://tuber.guzus.xyz/video/mPAHvz8kW24/summary) · [YouTube](https://youtu.be/mPAHvz8kW24)

**Score: 67/100 · skim**

Brex CEO Pedro Franceschi argues that we are only about six months past AI's "electricity moment" (Claude Opus 4.5 in December). His advice is to stop treating LLMs as scarce resources to micromanage and instead give agents tools, context and tokens. Founders should keep for themselves only what models can't do: choosing which problems matter and pulling unspoken signal out of customers. He says CEOs of non-AI-native companies should treat this as a turnaround that they lead personally.

- **0:00-3:03** — **Intro and "free the claw."** Garry Tan says Pedro's lunch visit to YC pushed the team into building their own agent setups, which led to "GBrain." Garry admits his mistake was writing about half a million lines of Rails code to tightly control what the LLM sees, a "Foxconn factory" for agents. Both agree that good AI products are mostly just an agent loop with tools, skills and a model. Over-engineering the harness adds little.

- **3:03-5:56** — **How Pedro got "AI-pilled."** He played with GPT-3 during the pandemic, but it felt like a research toy. ChatGPT, reasoning models and tool use helped, but he calls everything before December a blip. His framing is that "electricity was invented in December" with Opus 4.5, the first time coding harnesses really worked. Claude Code had existed for about a year before then without being very valuable. OpenClaw was the next step, showing how far you can get just by configuring markdown files. He once bought a movie ticket entirely through OpenClaw using a Brex card provisioned by API. His team said they could do it in 10 seconds by hand, and he told them they were missing the point.

- **5:56-9:38** — **Securing agents with CrabTrap.** His first personal setup gave the agent read-only OAuth access to email, Slack and other tools, and it got surprisingly far. Getting write access approved at Brex meant solving security first, and he spent about four weeks on it himself. He concluded that controlling the tool layer (like NVIDIA's NemoClaw-style forks) isn't enough, because an agent can still send a bad HTTP request. So Brex built **CrabTrap**, which it open-sourced about two months ago. It is an HTTP proxy around an agent's whole network boundary, so every request can be audited. After recording a day of an agent's traffic, another model writes a policy for it. Models read HTTP traffic unusually well because they were trained on the web. For example, Brex's recruiting agent "Jim" gets about 98% of its requests approved automatically, and the remaining 2% go to an LLM judge. Caveat: Brex still doesn't use this on customer data as aggressively as it eventually wants to. Pedro's view is that if a regulated fintech can do this, YC startups have no excuse not to be at the bleeding edge.

- **9:38-14:17** — **Why only the LLM judge, and how adoption splits into three tiers.** They dropped a credential vault from CrabTrap to do one thing well. The LLM-as-judge was what got Brex's strict security team to say yes. "We're not in the business of building HTTP proxies," Pedro says; they built it because they needed it, and he'd happily use a better YC-built version. He sees AI adoption inside companies in three tiers:
  - token-maxing engineers who live in coding harnesses;
  - average engineers, at roughly a tenth of that productivity;
  - everyone else, using AI in "Google search mode" (a chatbot plus a few MCPs).

  Brex's bet is to give non-technical teams a harness of their own: an OpenClaw-like "virtual employee" with Slack, email and meeting access that bootstraps new skills from markdown without anyone hand-coding them. Garry gives an example: YC planned 60 Startup School dinners by talking to a claw in Slack, and nobody opened Claude Code. Pedro's caveat is that Claude Code isn't magic. It's a harness around the same API models, and a claw can call Claude Code or Codex itself (ACP helps with this).

- **14:17-18:05** — **Why token-maxing hasn't spread, and the "AI-pilled test."** Founders are shy about spending on tokens. Pedro says that as a teenager today he would token-max cheaply with Chinese open models or a home GPU rig. A tell: far more people should be complaining about Max plan limits, but only about 0.1% of Twitter does. His test is whether you default to AI first for any problem in your life, until your brain is rewired to think that way. If he started a company today, his premise would be "why can't it just be me?" The real unlock isn't cost savings. It's that the company's internal boundaries become type systems, interfaces and agents talking to each other instead of people. Solve problems with AI first even when it feels suboptimal, because that learning compounds.

- **18:19-22:02** — **Keep the MVP surface small; AI doesn't fix choosing.** Successful companies started with tiny customer surfaces: early Stripe was just an API, early Brex was a terminal with no UI, and Airbnb and DoorDash were basically forms with founders doing the work behind them. The risk with AI is that cheap building erodes discipline about what matters. If you can't solve the problem within clear, minimal boundaries, you haven't found the right problem. "Intelligence is compression": a good idea fits on a napkin. On Brex's pivot process (two-week explore/exploit cycles), he would still use it. Execution is now handled by models, so the bottleneck is the wisdom to choose. That depends on signal the models weren't trained on, mainly customer conversations, where customers describe local-optimum problems rather than prompt-ready answers.

- **22:02-28:01** — **Signal that isn't in the models, and the out-of-distribution problem.** Pedro would explore more ideas in parallel early on but wouldn't parallelize deep customer understanding. Synthetic customer personas only help once you already know the customer well. Brex learned this when it moved from selling to founders like themselves to selling to finance teams with different values. His heuristic is to spend time on what only you can do and models can't. The biggest pitfall of LLMs is that you can't tell how much training data sits behind a given answer. A model that reported how often something appeared in its training data would change how much you trust it, and Garry calls that a startup idea. Data companies like Mercor hunt for these blind spots, but you need expertise to see the gaps, which founders exploring a new space lack. Garry fills gaps by having agents deep-research everything into a git repo, GBrain, for retrieval. Brex does the same for customers with a "customer world model" that takes in every touchpoint, from dashboard clicks to emails and calls, to predict what each customer needs next.

- **28:01-31:41** — **Model bias and being "long inference."** Model behavior reflects the people who build the models. When Pedro asked a model for accounting-category examples, the first one it gave was "AI capex," because model builders think about AI capex. His case for being "long inference" uses a chart of 2,500 dots, each about 3.2 million people. About 84% of the world has never used AI, 16% have tried a free chatbot, about 0.3% pay $20 a month, and one dot uses agents. He expects tokens to become companies' biggest expense: even if prices fall 10x, usage will rise 10x. Brex built an internal tool called **Magpie** that attributes every dollar of token spend to a product, customer, internal tool or employee, and is now adding analytics to measure ROI.

- **31:41-36:40** — **Token-spend data, and redesigning from scratch instead of bolting on AI.** Brex data shows heavy token use, and faster revenue growth, within roughly a 10-mile radius of SF plus New York. Elsewhere, even large companies with big budgets spend around $10K a month when they should spend 10-100x more. About two and a half years ago Brex leadership asked what they'd do differently if they started in 2024, and the answer was "everything." The KYC example: KYC used to be about 80% automated and 20% manual. The obvious move was to build an agent for the manual part. Instead they redesigned onboarding end to end and realized that with near-free KYC, they can screen leads rather than signed customers. That moves risk assessment to the top of the funnel and changes who they target, and the same partly applies to credit. Pedro says competitors who "latch AI on top" of old processes don't get these step changes. Diana compares OpenClaw/Hermes power users to Arch Linux users, versus ChatGPT users as Ubuntu users: heavy customization is what lets you build something like Valve's SteamOS.

- **36:40-39:00** — **The ROI debate and the electricity analogy.** A CFO of a large public company told Pedro they see the token spend and more lines of code, but question the ROI. Pedro agrees ROI tracking matters ("don't bankrupt your company on tokens") but says it misses the historical moment. A host adds that early electricity also had poor ROI, and people stuck with it out of curiosity. Pedro was up until 2 a.m. the night before experimenting with Opus 4.8. He says he'd do it unpaid, and treats that curiosity as the real litmus test. Over the very long run he expects tokens to become nearly free, the way we rarely think about electricity costs today.

- **39:00-46:26** — **Leading AI change at a large company (Brex plus Capital One).** "The CEO needs to be the chief AI officer." Only the CEO sees the whole system; a KYC team would never think of using KYC to score leads. The key exercise: take today's technology back to the day you founded the company, design how you'd build it, compare that with what you have, "suffer in silence," then act. He calls this re-founding the company's identity, effectively a turnaround. Brex splits AI into three agendas:
  - **product AI**, shipped to customers;
  - **operational AI**, which serves customers at scale (support, risk, onboarding);
  - **corporate AI**, for how employees work internally.

  Organizations build "antibodies" against disruption. The CEO can break glass about 10x more easily than an executive, who can do it about 10x more easily than an employee. A blocked AI project takes him 10 seconds to unblock versus 10 hours of escalation for someone else, and without him it often just gets built the old way. On Jack Dorsey's idea of a "company AGI," Pedro prefers domain-specific virtual employees with clear boundaries, such as a customer world model with a separate roadmap agent built on top of it. He only counts usage that actually replaces a hire or saves hours. Brex's client sales team runs on the customer world model, and it found things before a customer lunch the team didn't know about, such as an executive's card problem at an airport.

- **46:26-53:56** — **Evals, personal agents and closing advice.** Brex turns every human intervention into an eval case. When the expense agent has a bad conversation, it files a bug, and an agent edits code and prompts until the eval passes. Engineers step in only if that fails. The goal is a self-improving system; Garry compares it to GBrain's nightly "dream cycle." Garry's GBrain holds 350,000 markdown pages. He also describes an "LSD" (lateral synaptic drift) mode that forces orthogonal concept combinations and ranks them for coherence, and he ingested a 60GB Google Takeout, filtered with Haiku down to about 4,000 meaningful emails. Pedro's main developer interface is voice memos to OpenClaw. He also says organizing context is the main bottleneck. His closing advice to founders:
  1. Remember you're six months into electricity.
  2. Keep a sticky note asking why AI can't solve each problem. Chatbots handle about 80%; build something for the other 20% to learn what the technology can do.
  3. Measure your token use and start from "why can't it just be me?", spending your own time only on choosing problems and covering for what LLMs can't do.

  In a big company that means a turnaround with the LLM placed almost as founder or CEO. The video ends with brief thanks.

**Practical takeaways**
- **Default to AI first:** for any new problem, ask why AI can't solve it. Treat suboptimal early attempts as an investment that compounds, not a cost.
- **Secure at the network layer, not by caging the agent:** an auditing HTTP proxy, a policy learned from recorded traffic, and an LLM judge for edge cases is how Brex got security sign-off. It approves about 98% of requests automatically.
- **Redesign, don't bolt on:** ask how you'd build each process if you founded the company today, as with KYC moving up the funnel. In established companies the CEO has to drive this personally and clear blockers quickly.
- **Keep scope and judgment human:** keep the MVP surface small, make ideas fit on a napkin, and spend founder time on customer signal the models don't have.
- **Make agents improve themselves and track spend:** turn human interventions into evals, and attribute token costs by product, customer and employee, without letting ROI anxiety stop experimentation.

## Audience signal

- No high-signal audience themes available.