Why CEOs Must Lead AI Transformation Themselves

The CEO must be the chief AI officer

Pedro Franceschi, co-founder and CEO of Brex, says something most founders won't want to hear: the only person standing between your company and a genuine AI transformation is you. Not your engineering team. Not your product leads. You. And if you are not spending your days token-maxing, breaking glass, and redesigning core processes from scratch, you are already falling behind. Franceschi makes the case that we are six months into a platform shift as significant as the invention of electricity, and most people are still lighting candles. The hidden consequence: the companies that will dominate the next decade are those willing to suffer short-term inefficiency and organizational discomfort now, because the cost of inaction compounds faster than token prices drop.

Why only the CEO can break the antibodies

Franceschi's central argument is deliberately uncomfortable. He insists that the CEO must be the chief AI officer, not as a symbolic title but as a daily practice of pushing technical boundaries and understanding what the models can and cannot do. The reasoning is brutally practical: nobody else in the organization has the authority to override the natural immune response that kills transformative ideas.

"I think the CEO needs to be the chief AI officer. Like it's not an engineering team thing. It's not like a product team thing is like you have to understand the bounds of technology better than anyone."

The problem isn't technical. It's organizational. Franceschi describes companies as systems that build antibodies against any disturbance to social cohesion. A senior engineer might know exactly what needs to change but lacks the political capital to push it through. An employee sees an opportunity but faces ten hours of meetings to get a simple approval, and probably gives up. The CEO can break that in seconds. That asymmetry is a superpower, but only if the CEO is actually in the trenches, feeling the friction of the technology. Most aren't. They delegate AI strategy to a VP of Innovation or a Chief Digital Officer, which ensures the transformation stays safely within existing processes.

The system responds predictably: incremental AI applied to old workflows. Franceschi calls this the "Foxconn factory" approach, treating the LLM like a constrained worker that needs endless guardrails and if-then logic. It feels productive. It creates a dashboard. It doesn't change anything that matters.

Redesign the process, don't just add AI

The most dangerous move a company can make right now is to bolt AI onto existing workflows. Franceschi illustrates this with Brex's KYC (Know Your Customer) process. The obvious approach: automate the 80% that can be automated and leave 20% manual. That's what most companies do. It feels smart. It feels low-risk.

Here's what they actually did: they redesigned the entire onboarding process from scratch. Which meant asking a deeper question: what changes when you have on-demand intelligence? The answer: you can now qualify leads as early as the top of the funnel using KYC data. You know who will pass credit checks before you even engage them. That changes your entire go-to-market strategy, not just your operations.

"When you redesign the entire onboarding process what you realize is there's a very important thing that happens in the beginning of the funnel which is deal qualification like is this customer even remotely qualified to be a Brex customer? But when you have KYC for free, you can KYC a lead versus the customer."

The downstream effect ripples outward. Customer acquisition becomes targeted. Risk becomes a predictor. The company's identity shifts. Franceschi calls this "refounding" the company, reimagining what it would look like if you started today with the current technology. That exercise is painful and requires founder-level energy, but it's the only way to create a real discontinuity. Most competitors won't do it because it requires admitting that your current processes are obsolete. And that hurts.

Security as the unlock, not the gate

Every enterprise AI conversation eventually hits a wall: security. Franceschi's team at Brex faced the same problem. They wanted to let agents write into production systems, not just read. The security team's natural inclination was to say no, for all the right reasons. Instead of negotiating from that position, Franceschi spent four weeks solving the hardest problem: how to secure the network boundary of an agent.

The result was Crab Trap, an open-source HTTP proxy that treats every agent request as auditable. The key insight: because LLMs are trained on hundreds of billions of web documents, HTTP traffic is actually the language they reason in best. So they built a system where another LLM acts as a judge, analyzing traffic patterns and approving or denying requests based on policy. After a day of operation, 98% of requests pass automatically. The remaining 2% get adjudicated by the model.

"We ended up realizing that the only way to actually do something about it was to do something in the network layer and if you treat the agent like you know the agent has its own will's desires and you know they go to the Esalen Institute for agents."

This is classic second-order thinking. Instead of adding more tool-level controls (which the model can bypass via HTTP requests), they changed the environment the agent operates in. The system now teaches itself what good behavior looks like. The immediate cost was four weeks of engineering and some uncomfortable conversations. The lasting advantage is an enterprise that can experiment aggressively with AI on customer data while maintaining compliance.

Token maxing: the competitive moat nobody wants

Here's where Franceschi's thinking becomes most contrarian. He argues that the biggest determinant of future success is not efficiency but consumption, specifically how many tokens you burn and how deeply you integrate AI into every problem you face. Most companies are still in "Google Search Mode" with AI: ask a question, get an answer, move on. The token maxers, the top 0.3% of users, are building custom harnesses, feeding their entire email history into context windows, and spending hours each night debugging agent workflows.

The gap between these two groups is widening exponentially. Franceschi points to data showing that companies in tech hubs like San Francisco and New York are consuming orders of magnitude more tokens than similarly resourced companies elsewhere. The cost difference is real but irrelevant. The ROI of early electricity adoption was terrible. The people who stuck with it did so because they saw the possibilities. Same here.

"I keep trying to picture myself. Imagine if I was like 14 or 12 when I started coding for real and I had the technology we have now, I would be token maxing the cheapest way possible."

The signal is clear: the companies building their own "customer world models," running automated evals every night, and designing self-learning agent loops are the ones pulling away. Everyone else is optimizing for a cost that will soon be negligible, while missing the system-level advantage that compounds daily.

Key Actions

• Become a token maxer yourself (immediate). Spend at least an hour a day pushing the boundaries of what you can do with AI: custom harnesses, voice-to-agent pipelines, full takeout data ingestion. Don't delegate this to your team. You need first-hand frustration to see the possibilities.
• Redesign one core process from scratch this quarter. Pick a department like onboarding, customer support, or product development. Ask "What would this look like if I started from zero today?" Then build that, not a patched version. Expect it to feel slow and painful. That's the point.
• Solve security boundaries for agents (next 3-6 months). Implement network-layer controls like Crab Trap or equivalent. The goal is to make it safe enough to let agents write into production systems. Most companies stall here because they over-engineer tool-level controls. Start with the proxy.
• Measure token consumption, but don't let cost dictate strategy (ongoing). Track usage per team and per use case, but resist the urge to cut budget. The correct response to high token spend is usually "what else should we be spending on?" not "how do we use less."
• Create an escalation path that bypasses organizational antibodies (immediate). Make it trivially easy for any employee to surface an AI opportunity that requires breaking existing rules. Remove the friction of ten meetings. If you're the CEO, tell people to come directly to you.
• Build a "dream cycle" for your agents (12-18 months). Every agent should improve itself automatically overnight by analyzing failures, generating evals, and updating its own codebase. This turns manual maintenance into a self-learning system and is the difference between a demo and a production capability.
• Spend time on things only you can do (immediate, forever). Franceschi's framework: what can't the models do? Right now, that's choosing which problems matter and extracting unspoken customer signal. Everything else, execution, code generation, data synthesis, is getting cheaper by the week. Your job is to stay where the alpha is.