⚡ Six Months After Electricity: Building Companies in the Age of…

⚡ The Electricity Analogy

Pedro Franceschi, co-founder and CEO of Brex, opened with a powerful historical lens: "Electricity was invented in December" — referring to the launch of GPT-4.5 and the reasoning models that followed. Six months after that inflection point, most companies are still "playing with candles," while a small fraction have begun to harness the breakthrough.

The implication is profound. Standing at this moment in the 200-year arc of technological progress, founders and executives face a choice: continue optimizing legacy workflows or rebuild from first principles, knowing what's now possible with AI.

"You're standing in the timeline of history and it's six months after electricity was invented. Thinking about your electricity bill being so high — like, yes, of course, don't bankrupt your company on tokens — but the question is: what do you do differently knowing everything that will be true about electricity?"

For Franceschi, this isn't theoretical. Brex has moved aggressively to rewire its entire company around AI — from product to operations to corporate workflows — and the lessons are both tactical and philosophical.

🧠 Why Token Maxing Hasn't Gone Mainstream (Yet)

Despite the hype, token consumption remains surprisingly concentrated. Franceschi estimates that roughly 84% of the world has never used AI, 16% have tried a free chatbot at least once, 0.3% pay $20/month for premium access, and only a single box out of 2,500 (representing 3.2 million people each) actively uses agents.

Even within high-growth tech companies, token spend is lower than expected. Brex's data shows significant variance: companies in the 10-mile radius of San Francisco and New York are burning tokens at multiples of their peers elsewhere — and correlating that with faster revenue growth. Meanwhile, "very large companies with very large budgets that could be token maxing spend like $10,000 a month" when the economically rational amount might be 10x to 100x higher.

Why the gap? Franceschi identifies three barriers:

Risk aversion: Companies overengineer "harnesses" to control LLMs, treating them like precious resources rather than abundant inference engines.
Legacy mental models: Teams default to inserting AI into existing processes instead of redesigning from scratch.
Lack of CEO-level commitment: Without top-down mandate and curiosity, organizations build antibodies against the disruption AI requires.

His prescription? "The CEO needs to be the chief AI officer. It's not an engineering team thing. It's not a product team thing. You have to understand the bounds of the technology better than anyone."

🦀 The Crab Trap: Solving Security at the Network Layer

One of Brex's most important internal innovations is Crab Trap, an open-source HTTP proxy designed to secure agentic workflows in production. The challenge: how do you "free the Claw" (give agents autonomy) without exposing financial services data to uncontrolled behavior?

The solution was counterintuitive. Rather than building complex control layers inside agent harnesses — what Franceschi likened to "putting the agent inside a Foxconn factory" — Brex focused on the network boundary.

Here's how it works:

All HTTP requests from an agent pass through Crab Trap.
After recording a day's worth of agent traffic, the system uses an LLM as a judge to build a policy.
~98% of requests auto-approve based on learned patterns; ~2% require LLM review against policy.
The proxy is auditable, adaptable, and surprisingly effective — because "models are trained on hundreds of billions of web documents, so HTTP traffic is actually the way models reason more than anything else."

This approach allowed Brex's traditionally rigorous security team to greenlight agentic experimentation in production — a unlock that enabled the company to move to the bleeding edge of enterprise AI deployment.

"We're not in the business of building HTTP proxies. We're in the business of being at the bleeding edge of what you can do with AI. To get to the bleeding edge required us to build this proxy."

🔄 The Three Pillars of AI Adoption at Brex

Franceschi breaks down AI transformation into three distinct agendas, each requiring different strategies and timelines:

Product AI: Features shipped directly to customers (e.g., expense categorization, underwriting automation).
Operational AI: Tools that improve the company's ability to serve customers at scale (e.g., KYC, customer success, risk operations).
Corporate AI: Internal workflows — how employees collaborate, make decisions, and execute.

The critical insight: these three layers are interdependent, and focusing on just one (usually Product AI) leaves massive value on the table. Franceschi emphasized that "people will sometimes pigeonhole themselves in one of the three, but you have to take a step back and ask: why can't you solve everything with AI?"

Example: Redesigning KYC from scratch. Instead of automating the existing 80/20 manual/automated KYC process, Brex reimagined the entire onboarding funnel. The realization: "When you have KYC for free, you can KYC a lead versus a customer." By moving risk evaluation earlier in the funnel, Brex gained deal qualification superpowers — changing not just operations, but go-to-market strategy.

🧪 Building Company-Specific AGI: The Virtual Employee Framework

Franceschi subscribes to a domain-specific view of AI rather than a monolithic "company model." His framework: virtual employees — agents with well-defined scope, clear APIs, and functional expertise.

Examples include:

Customer World Model: An agent with "total information awareness" of every customer touchpoint — emails, support tickets, product usage, meeting notes. Brex's client sales team now runs on this model. Franceschi used it to prepare for a customer lunch, surfacing insights "including things the team didn't know about, like an executive traveling who had an issue at an airport with their card."
Jim (Recruiting Agent): Operates under a specific policy managed by Crab Trap, with 98% of requests auto-approved.
Expense Agent: Handles employee inquiries, with every flagged conversation automatically converted into an eval case — triggering another agent to modify code and prompts to ensure the issue doesn't recur.

This architecture mirrors an "executive team" rather than a single superintelligence. "Domain specificity still matters. Functional knowledge still matters. You should separate the agent emitting code from the one talking to customers from the one reasoning about product roadmap."

🔁 The Dream Cycle: Self-Improving Systems

One of the most forward-looking ideas discussed: how do you make agents that improve every day?

Franceschi's answer: bake evals into every human interaction. When a Brex employee manually intervenes in an AI-driven workflow — say, resolving a KYC exception the model couldn't handle — that interaction becomes a breaking change and an eval case. An agent then attempts to modify the codebase or prompts to handle it autonomously next time. If it can't, an engineer steps in.

The vision: "A self-learning system where the agent dreams every night — 'What's going on? Is there a pattern? How do I recause this?'"

This closed-loop design is inspired by the best AI products in the wild, which Franceschi describes simply: "Every good AI product is an agent loop with tools. There's not really much else."

🎤 The Personal AI Stack: Free the Claw

Franceschi's own setup is legendary among YC partners. After a casual lunch demo, it reportedly "sent the entire team down a rabbit hole of building on their own."

His journey started conservatively: read-only access to email, Slack, and other services via OAuth tokens. Even without write permissions, the utility was shocking. From there, he tackled the hardest problem — security — leading to Crab Trap.

Today, his workflow emphasizes:

Voice memos to OpenClaw via Telegram: "My most-used developer UI right now." This forces him to make agents smarter rather than build more UI.
Ingesting personal context: A 60GB Google Takeout archive, filtered down to ~4,000 consequential emails. "Those emails capture your thinking and the consequential moments of your life."
LSD Mode (Lateral Syntactic Drift): A feature in GBrain (built by YC partner Jared Friedman) that deliberately combines orthogonal concepts from a knowledge base, ranking novel combinations for coherence. "The top five tend to be banger tweets."

The underlying philosophy: tokens should be abundant, not rationed. "Let it rip. Free the Claw. Give it tokens."

📉 The Trap of Legacy Thinking

Franceschi offered a sobering exercise for later-stage founders: "If you started your company today, with the AI tools available now, what would you do differently?"

The answer at Brex? "Everything."

This creates a painful realization: companies are operating with "a completely old way of thinking about the fabric of the company, the way we build product, the way we build processes." The natural response for most organizations is incremental adoption — "latch AI on top" of existing workflows.

But that's the wrong move. Franceschi's prescription:

"You have to refound the very concept of what the company's self-identity is. It's almost like a turnaround. If you're a large company that's not AI-native, you're doing a turnaround to some degree."

This requires founder energy. Only the CEO has the authority and context to redesign cross-functional systems. "The KYC team would never think of using KYC technology to score a lead. The only people who can think about the organization of the system itself are those with context of the whole."

And critically: escalation paths must be desensitized. Companies naturally build antibodies against disruption to social cohesion. Franceschi's approach: make it trivial for teams to escalate AI experiments directly to him. "It takes me literally 10 seconds to solve a problem that would take someone 10 hours — or maybe never."

🧩 The Missing Signal: Customers and Compression

Despite all the progress, Franceschi is clear-eyed about AI's limits. LLMs are trained on a specific corpus, optimized for specific benchmarks — and they have no intrinsic sense of how much training data they've seen for a given query.

"Imagine if every time you asked an LLM a question, it gave you a sampling frequency of how much this appeared in the training data. You would trust answers very differently."

The unspoken signal — customer needs, local constraints, implicit worldviews — doesn't exist in the training corpus. This is why "you can't prompt your way into a billion-dollar company." The founder's job is to extract that signal, compress it into clear problem definitions, and then unleash AI.

His rule of thumb: "Great ideas fit on a napkin. Intelligence is compression."

The risk of AI abundance is "lack of discipline on what matters to solve." Franceschi warns against using AI's speed as an excuse to avoid choosing a minimal surface area. "If you can't compress the problem into a smaller surface area, you haven't found the right problem to solve."

🌍 The 10-Mile Radius and the Long Inference Thesis

Franceschi is bullish on long inference — the idea that inference compute will vastly exceed current expectations. The bull case isn't just about model improvements; it's about adoption curves lagging capability.

Using Brex's payment data, he sees a clear pattern: companies in the SF and NYC 10-mile radii are spending aggressively on tokens and growing faster. Everywhere else? "Very large companies with very large budgets spend $10,000/month when they should be spending 10x or 100x more."

Even as token costs decrease 10x, usage will increase 10x, keeping absolute spend high. Brex has built an internal tool called Magpie to track every dollar of token spend, attributing it to products, customers, internal tools, or employees — and layering in analytics to understand ROI.

Franceschi's conclusion: "Tokens will be the biggest expense in a company, easily. And yet, we're just starting."

✨ Advice for Founders: The AI-First Playbook

Franceschi closed with three directives for founders:

Marvel at the moment. "You're standing at a 200-year timeline. Six months after electricity was invented, what do you do differently?"
Start with a daily habit. "Put a Post-it on your computer: 'You wake up, whatever problem you have in your life, why can't you solve it with AI?' Start there." Push the limits daily, not to solve every problem immediately, but to develop texture and feel for what's possible.
Measure token consumption. Track it like revenue. Ask: "Why can't it just be me? Why can't I build the whole thing alone?" Spend your time on things only you can do: choosing what problems to solve, talking to customers, and injecting the signal that models lack.

Above all, adopt a founder mindset even inside large companies. Treat AI transformation as a turnaround. Redesign from scratch. And remember: "The biggest risk is not taking the risk — it's missing the opportunity to rethink the problem from 'what would you do if you started the company today?'"

🔮 The Bottom Line

Pedro Franceschi and Brex represent a rare breed: a high-growth fintech that has moved beyond experimentation to rewiring its entire operating system around AI. From Crab Trap to customer world models to self-improving agent loops, the company is building infrastructure and culture for a post-2024 world.

The lessons are universal:

Tokens are cheap. Inaction is expensive.
Security is solvable — at the network layer.
Domain-specific agents beat monolithic models.
Evals should be baked into every workflow.
The CEO must be Chief AI Officer.

And perhaps most importantly: we're only six months in. The steam engine is still 20 years away. Those who treat this moment as a marginal improvement rather than a refounding event will look back with regret.