Pedro Franceschi

1. Key Themes

The CEO as Chief AI Officer

Pedro argues that AI transformation cannot be delegated to engineering or product teams — it must be owned at the top. The CEO is uniquely positioned to see the whole system and redesign it.

"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. It's like you have to understand the bounds of the technology better than anyone." [00:00:00]

"The KYC team would never think of using the KYC technology to score a lead. The only people that can think about the organization of the system itself is if you have the context of the whole." [00:39:49]

Stop Treating LLMs Like Precious, Scarce Resources — Free the Claw

The dominant failure mode for most AI builders is over-constraining the model — building elaborate if-statement harnesses to control what the LLM sees, rather than just giving it tokens and letting it operate.

"Most people in software are still getting it wrong. They've been treating the LLM like this very precious thing that's very expensive. And so as a result, you have to literally put the agent inside a Foxconn factory... I need to control what the LLM sees because it's about really, really — I only want the context from here. And let me write all the if statements to make sure, like a Foxconn engineer, you're waking up at 6 a.m." [00:01:53]

"Every single good AI product you've used is an agent loop with tools. That's it... it's skills, tools and a model. There's not really much else." [00:02:46]

Security Is the Real Unlock for Enterprise AI Adoption

The actual blocker for aggressive enterprise AI experimentation isn't model capability — it's security. Brex solved this at the network layer, which opened the door to everything else.

"Where I spent, I don't know, probably four weeks of my time was, okay, let's solve the hardest problem, which is security. And we ended up realizing that the only way to actually do something about it was to do something in the network layer." [00:06:51]

"We built this thing called Crab Trap, which we open sourced probably about two months ago... you HTTP proxy the entire network boundary of an agent... you basically can use another agent to analyze the traffic and create a policy to let traffic go through or not... 98% of requests go through automatically, 2% use an LLM." [00:07:17]

Redesign the Entire Process — Don't Just Latch AI Onto Old Workflows

The biggest wins at Brex came not from adding AI to existing processes, but from scrapping the process entirely and rebuilding it from scratch with AI-native assumptions.

"A lot of our competitors had this approach of saying, oh, I have this entire old process. Let me go and latch on AI on top of it or latch on AI on top of our product. And I think the biggest discontinuities in a positive way that we've had were when we said, hey, let's keep this old way here, put it in a corner, and how would we design it if we started the company today from scratch?" [00:34:33]

"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... when you have KYC for free, you can KYC a lead versus a customer. So you start to have risk orientation up in your funnel and that changes who you even target." [00:34:05]

The Three-Tier AI Adoption Problem Inside Companies

Most companies have a huge gap between power users and everyone else, and the answer isn't giving people more MCPs — it's building proper harnesses for non-technical employees.

"There's tier number one, which is your token maxers... And then you have the sort of average engineer... probably a tenth of the productivity. And then you have like the entire rest of the company. And the entire rest of the company typically is interacting with AI in what I call like Google search mode way, which is a chatbot with a few MCPs." [00:11:00]

"Our thesis was, if you think about the value that AI creates for like a token maxer, a lot of the value comes from the harness. The thesis was how to actually build an equivalent harness for other teams that are non-technical." [00:11:50]

Token Spend Management as the Next Big Frontier

Token spend is on track to become the largest expense in most companies, yet almost no infrastructure exists to manage it intelligently. Brex is building this internally and sees it as a massive white space.

"It will be the biggest expense in a company, like, easily... even token costs decrease by 10x, they're going to have 10x more usage. So it will be still a large cost. And we're spending a lot of time thinking how to help companies actually manage token spend." [00:30:35]

"We ended up building our internal version of this. We call it MagPi, where the idea is you can effectively, you know, every dollar of token spend in the company, you can attribute to a product we have to customers, an internal tool that we use to serve, or an internal employee, and understand model usage... we're now figuring out how to build analytics on what are we trying to do with the tokens to start to get a sense of ROI." [00:31:05]

AI Adoption Is Still Shockingly Nascent — Be Long Inference

The data on actual AI usage is staggering in how early we are. 84% of the world has never used AI, and only one box out of 2,500 (representing 3.2 million people each) actually uses agents.

"84% of the world never used AI. 16% have used at least once a free chatbot. Then 0.3% pay 20 bucks a month for AI. And one box out of the 2,500 actually use agents in whatever capacity. So that's the argument to be long inference." [00:30:06]

"You look into like very large companies with very large budgets, and that could be token maxing. And the economic thing for them to do would be to token max. And they spend like $10,000 a month. And you're like, you should probably be spending 10 times more or 20 times more or 100 times more." [00:32:15]

Agents Need a Dream Cycle — Self-Improving Systems Are the Real Unlock

Getting an agent working is table stakes. The unlock is building the feedback loop so every human interaction with the agent becomes a training signal that improves it automatically.

"I think a lot of what I see with companies is they spend a lot of time getting an agent working, but never thinking how to make the agent improve every day. And I think that's like always the biggest unlock of this model." [00:27:27]

"Whenever someone has a conversation with the agents that flags an issue or a bug or something that feels like the conversation didn't go as smoothly, that creates a bug. So that bug triggers an agent that's going to go and modify the code base and the prompts and everything to make that eval pass." [00:46:57]

Minimal Surface Area Remains the Founder's Most Important Discipline

Even with AI enabling you to build more, the most successful companies start with the smallest possible surface area. AI tempts founders to abandon this discipline.

"I think the risk with AI is that the agency behind choice goes away. So you have this lack of discipline on what matters to solve... I always tell people, like, I think if you can't minimize your surface area and solve the problem with a very clear set of boundaries, you haven't found the right problem to solve." [00:19:05]

"Intelligence is compression. So when someone comes to pitch me an idea in the company, I'm like, it has to fit in a napkin. Like, great idea fits in a napkin. What's your napkin?" [00:20:01]

2. Contrarian Perspectives

The Founder's Job Is No Longer Execution — It's Choosing What to Build

Most people still think of founders as people who both choose and execute. Pedro argues execution is essentially gone, making pure wisdom-of-choice the scarce resource.

"A lot of the job is — the job now is to have the wisdom to choose what you want. Because before, the wisdom was not just to choose, was to choose and know how to execute it. The execution is out, right? The execution is gone and the models are going to do that better. The wisdom to choose is still, I think, the missing bottleneck." [00:21:39]

LLM Outputs Are More Biased by Their Builders Than Anyone Admits

Models aren't neutral tools — they reflect the mental models of the people who built them in ways that are subtle and systematically misleading for most use cases.

"I was playing with AI for accounting categorization. And then the first example of an example is AI CapEx. And I'm like, oh, why is it AI CapEx the first example it comes up with? Because the people that are building the models fucking only think about AI CapEx, right? So there are things like that that I think is kind of interesting to think about — the mental models of the models I think are out of the box are more biased than we may give them credit for." [00:28:49]

ROI Analysis of AI Is the Wrong Frame Entirely

Measuring AI ROI right now is like measuring electricity ROI six months after it was invented — you will get the wrong answer and make the wrong decision.

"Imagine someone saying in the 1800s, like, oh, my electricity bill is so high now. Like, gosh, let's use a little less. Let's push this steam engine to come, like, maybe 20 years later because the cost savings." [00:37:35]

"It wasn't the cost savings. It was just because people were curious about it." [00:38:07]

Most Companies Are Doing AI Turnarounds, Not AI Transformations

Pedro's framing is harsher than the typical "transformation" narrative — if you're a large non-AI-native company, you are in a distressed situation requiring a full turnaround, not incremental improvement.

"I think it's a turnaround almost. I think you have to assume that if you're a big, large company that's not an AI native, you're doing a turnaround to some degree." [00:41:42]

Synthetic Customer Models Only Work When You Already Know the Customer Well

Most people assume AI can replace customer discovery. Pedro argues the opposite — synthetic models are only valuable after you've built deep customer understanding, not before.

"We did a lot of exploration with synthetic customers and building customer world models and things like that. And those are really valuable. Once you know a lot about the customer. But when you don't know enough yet, I think there's this very basic thing." [00:22:39]

3. Companies Identified

Brex

Corporate spend management and financial services platform for startups and enterprises. Mentioned throughout as the primary case study for enterprise AI adoption — building Crab Trap (open-source agent security proxy), MagPi (internal token spend management), customer world models, KYC agents, recruiting agent Jim, and expense agents. Pedro is co-founder and CEO.

"We ended up building our internal version of this. We call it MagPi, where the idea is you can effectively, you know, every dollar of token spend in the company, you can attribute to a product we have to customers, an internal tool that we use to serve, or an internal employee." [00:31:05]

Anthropic (Claude / Claude Code)

AI lab, maker of the Claude models and Claude Code coding harness. Identified as the key enabling technology — Claude Code specifically called out as the moment agentic loops became real.

"Cloud Code existed for probably a year before, but it wasn't that valuable yet... that was the tip of the spear where you could say, yes, like coding harnesses actually work." [00:04:03]

Stripe

Payments infrastructure company. Cited as a canonical example of minimal surface area in early company building.

"You look at Stripe, for example. Stripe early days was, like, literally an API." [00:18:37]

Airbnb

Home-sharing marketplace. Cited as another canonical minimal surface area company.

"You look at Airbnb, it's like the website was a form. And the form was just like literally where you inputted what you needed." [00:18:37]

DoorDash

Food delivery platform. Cited as another minimal surface area example.

"DoorDash in the early days, similar, right? Like, it was just, like, literally. So the surface area was so small with the customer." [00:19:05]

NVIDIA

Semiconductor company. Mentioned in context of NeMo Guardrails as an alternative approach to agent security that Brex decided not to pursue.

"A lot of folks were, and we saw NVIDIA and others on NeMo — let's build these like open shell forks that have controls over tools that the model calls. And the reality is, yeah, you can do all that, but you can also just make an HTTP request wrong." [00:07:17]

Mercore

Data and AI company. Mentioned as working on identifying blind spots and out-of-distribution gaps in LLMs.

"That's what Mercore and a lot of the out-of-data companies are doing. Like, a lot of the jobs for them is to say, well, what are the blind spots for LLMs?" [00:25:41]

Valve (Steam Deck)

Gaming company. Cited as a powerful analogy for what deep customization of a foundational platform (Arch Linux) can produce — compared to deep customization of Claude Code.

"The operating system that makes it feel like a Nintendo Switch is actually built on top of Arch. They customize all the drivers, over-the-air updates. It works with all consoles. It works with all sorts of hardware out of the box. But they super-duper customized it." [00:36:05]

AgentVault

Credential brokering company in the Claude ecosystem. Mentioned as doing important work on credential management for agents.

"I think you had mentioned the first version of Crab Trap included like credentials vault... I think your intuition around like the proxy at the network level ended up being quite prescient. Like a lot of the stuff that I'm seeing around the Claude ecosystem at the moment is essentially doing that. Like we're seeing that with credentials, credential brokering, like AgentVault is doing a lot of that." [00:09:38]

4. People Identified

Co-founder and CEO of Brex. Identified as one of the most practically advanced enterprise AI operators among large company CEOs — personally building agents, architecting company-wide AI infrastructure, and redesigning core business processes from scratch with AI.

"You are by far one of the most AI pilled, farthest out on the edge, but also very practical CEOs who is playing with this stuff and actually building it yourself." [00:51:12]

Jack Dorsey

Co-founder of Twitter and Square. Referenced approvingly for his framing that every company is trying to build its own internal AGI.

"Do you buy into sort of like the Jack Dorsey view of every company is essentially trying to like build its own little company AGI?" [00:43:45]

Garry Tan

CEO of Y Combinator. Cited for two distinct insights: that tokens are expensive (explaining why founders underspend), and for the framing "make the implicit explicit" about customer discovery.

"Garry mentioned this point, which is tokens are expensive." [00:14:35] "I think Garry says this, make the implicit explicit — of what are all those desires." [00:23:51]

5. Operating Insights

Every Human Touchpoint With an Agent Should Automatically Become an Eval

Rather than treating evals as a separate engineering task, Brex has wired their production systems so any friction moment in a human-agent interaction triggers an automated eval creation and a code/prompt fix cycle.

"We have an expense agent in Brex. Whenever someone has a conversation with the agents that flags an issue or a bug or something that feels like the conversation didn't go as smoothly, that creates a bug. So that bug triggers an agent that's going to go and modify the code base and the prompts and everything to make that eval pass." [00:46:57]

Break Glass Fast — CEO Glass-Breaking Is 10x Easier Than Executive, 100x Easier Than Employee

Organizational antibodies kill AI experimentation at the employee and even executive level. The operating tactic is for the CEO to personally unblock escalations that would otherwise die in meetings.

"It takes me literally 10 seconds to solve that problem. And it would take someone 10 hours to go into the meetings and escalate and understand, okay, can we do this with AI? Or maybe never. Literally never." [00:42:41]

"I think the escalation paths need to be like desensitized in the system because the company builds antibodies against any sort of disturbance to the social cohesion of the company typically gets like rejected by the antibodies." [00:43:11]

Build a Customer World Model as Your First Company-Wide AI Agent

Rather than a monolithic company knowledge base, start with one well-scoped agent that aggregates every customer touchpoint — clicks, emails, calls, support tickets — and makes it queryable. This is the foundational building block for a broader company model.

"I had lunch with a customer tomorrow. And I don't know the state of that account as well as I probably should. My customer world model answered the question for me. And I now have a report including things that the team didn't know about that came through support tickets." [00:45:45]

Separate AI Into Three Distinct Agendas — Product, Operational, Corporate

Conflating these three causes companies to make uneven progress. Each has a different ROI profile and timeline.

"AI is an umbrella that I think has like three things in the way we talk about it internally. There's product AI, the product we actually ship to customers. There's operational AI, which is things that directly affect our ability to serve customers at scale... And then there's corporate AI, which is how people work internally." [00:40:45]

6. Overlooked Insights

LLM Training Data Frequency as a Product — The Model Doesn't Tell You What It Doesn't Know

Pedro dropped a highly specific and valuable observation: the biggest hidden failure mode of LLMs is that they give you confident answers regardless of whether that topic was well-represented in training data. He proposed a product that shows sampling frequency per answer — and the hosts agreed they'd pay for it.

"The biggest pitfall of LMs is you have no sense of how much training data the model has seen for the exact thing that you're asking it. Imagine if every time you're asked an LM a question, it gave you, like, yeah, the sampling frequency of this in my data set was X. And on this other answer was 0.00001X. You would trust it very differently, right? The distribution is so different." [00:24:51]

"Oh, I would pay for that. That's a great startup idea." [00:25:20]

This is a genuine product gap. Every enterprise deploying AI for domain-specific tasks (legal, medical, accounting, compliance) is flying blind on this. A model or wrapper that surfaces confidence calibrated to training data density — not just output probability — would be a significant trust and safety unlock. No major lab currently exposes this cleanly, and the comment was passed over in seconds.

Token Spend Management Is a Brex-Sized Opportunity — And Brex Knows It

Pedro mentioned in passing that Brex built an internal token spend attribution tool called MagPi, and that they are now "figuring out how to build analytics on ROI." Given that Brex's entire business is spend management and that Pedro explicitly says token spend "will be the biggest expense in a company," this is a strong signal that Brex is preparing to launch a commercial token spend management product. No one in the conversation flagged this explicitly.

"We're spending a lot of time thinking how to help companies actually manage token spend... we call it MagPi... we're now figuring out how to build analytics on what are we trying to do with the tokens to start to get a sense of ROI. But anyway, it's a fascinating topic that I think has a lot of early, early, early work compared to what it will be one day." [00:30:35]

If token spend becomes the largest line item in corporate budgets — and Pedro's inference thesis suggests it will — the company already managing corporate cards and expense management is the natural incumbent to own token spend management. This is either a major new Brex product line in development or the precise thesis for a competitor to build against Brex before they move.