Going all-in on AI: 7 lessons from Brex's CEO

Brex CEO Pedro Franceschi is one of the most AI-forward founders in tech. A self-taught coder from Rio de Janeiro, he built Pagar.me (the "Stripe of Brazil") before dropping out of Stanford to start Brex in 2017. In a recent Y Combinator Lightcone interview, he laid out how he rebuilt his own day-to-day (and is rebuilding a multi-billion-dollar company) around AI.

Most of it applies directly to you, whether your team is one person or fifty. Here's the distilled version.

1. Make "can AI solve this?" your default

Franceschi describes a simple test: whatever problem shows up in your life, your first move is to try solving it with AI. About 80% of the time, a chatbot gets you there. The other 20%, where it falls short, is the part that matters.

That 20% is your build list. Not because solving it once saves you an hour, but because it gives you a feel for the edges of the technology you can't get any other way. You only learn the limits by hitting them every day.

2. The CEO is the Chief AI Officer

AI isn't an engineering project or a product initiative you can hand off. "The CEO needs to be the chief AI officer," Franceschi says. "You have to understand the bounds of a technology better than anyone."

There's a structural reason for this. Only the person with context across the whole company can see that the tool you built for one team could reshape another. And only the founder can move fast when the organization pushes back. As he puts it, it's far easier for a CEO to break through internal resistance than an executive, and far easier for an executive than a regular employee.

3. Don't bolt AI on. Refound the company.

The single most valuable question to ask: if I had today's technology the day I started, how would I build this? Then compare that answer to what you actually have.

Bolting AI onto an old process gets you a modest improvement. Redesigning the process end to end is where the step-changes come from. Brex did this with customer onboarding: instead of automating their existing compliance checks, they redesigned the whole flow, and realized they could run those checks much earlier, which changed who they even target as customers.

Keep the old way in a corner. Then ask how you'd design it from scratch today.

4. Stop micromanaging your agents

Here's the mistake Franceschi thinks most people in software still make: treating the model as precious and expensive, so they wrap it in hundreds of rules to control exactly what it sees. He jokingly calls it putting the agent in a "Foxconn factory."

Every good AI product, he argues, is the same simple thing: an agent loop with tools. A model, a few tools, and the context and skills you give it. Your job is to organize the context well, not to handcuff the model. Give it room and let it work.

5. Be less shy about burning tokens

Founders are oddly frugal with tokens. Franceschi's counterpoint: this is the worst these models will ever be, and the cost will keep falling.

His analogy is electricity. We're roughly six months past the invention of electricity, and most people are still arguing about the price of their candles. Don't bankrupt your company, but don't optimize the wrong number either. If you've never hit a usage limit, you're likely underusing the tools that matter most.

6. Your edge is judgment, not execution

Execution is increasingly handled by the model, often better and faster than you'd do it yourself. What's left, the real bottleneck, is the wisdom to choose what matters and to feed the model the signal it was never trained on.

That signal lives in customer conversations: the unspoken needs a customer will never hand you pre-shaped as a prompt. The models can't manufacture it. So spend your time on the things only you can do: choosing the right problem, and talking to the people you're solving it for.

7. Start small and safe, then expand

One more lesson, especially if your agents will touch customer data. Brex didn't flip every switch at once. They started by giving agents read-only access (email, Slack, internal tools), and were surprised how far that alone got them.

Only later did they solve the harder "let it write" problem, and they did it at the network layer: auditing the agent's traffic, with another model acting as a judge for what to allow. You almost certainly don't need that level of infrastructure yet. But the pattern is the one to copy: least access first, a human in the loop for anything irreversible, and more autonomy as trust grows. It's the same discipline Anthropic uses to ship fast without getting burned.

What to do this week

1. Pick one problem from your week and try to solve it entirely with AI.
2. Note where it breaks. That 20% is your build list.
3. Write the one-line answer to: "If I started today, how would I build this?"
4. Check your token spend. If you've never hit a limit, push harder.

Going all-in on AI isn't a single tool or a one-time project; it's an operating system for how you run the company. That's exactly what we help founders build in AI Operating System for Startups.

Sources

• Pedro Franceschi on Y Combinator's The Lightcone, "The Most AI-Pilled CEO We Know", the interview this article distills.
• Background on Pedro Franceschi and Brex: CNBC, TechCrunch, and PYMNTS' "The Stripe of Brazil."
• LinkedIn: Pedro Franceschi · Brex.