The Most AI-Pilled CEO We Know

Franceschi says GPT-3 felt like a research demo, while reasoning models and tools made AI useful for real work. He describes discovering coding harnesses during a holiday break as a turning point, then exploring OpenClaw and using Markdown to configure agents. A movie ticket purchase through OpenClaw convinced him that the point was to explore what agents could do, not just save time on a familiar task.

At Brex, Franceschi first gave personal agents read-only access to his accounts, then worked on the harder problem of letting them write to company systems. Brex built CrabTrap, an HTTP proxy that audits agent traffic and uses policies, with an LLM judge for uncertain requests; he says 98% of requests from its recruiting agent Jim pass automatically. For broader adoption, he argues companies need virtual employees with useful tools and context, rather than chatbots limited to search.

Franceschi suggests using AI first when a problem arises, then learning where the models fall short. For founders, he says customer conversations reveal needs and perspectives that are missing from training data, especially when a company moves beyond customers like its founders. A practical guide for allocating time is to focus on what only a person can do, such as choosing which problems matter and understanding customers.

Franceschi says Brex leaders concluded they would build the company differently if starting again, and that the next step is to change how it works. In a KYC redesign, Brex moved checks earlier in the funnel so it could assess leads before they became customers. He divides company AI work into product, operational, and corporate uses, and argues CEOs should understand the technology firsthand and reconsider company processes from scratch.

Franceschi favors separate, domain-specific agents, such as one that builds a customer view and another that uses customer patterns to inform the product roadmap. At Brex, human reviews and agent conversations can become evaluation cases; reported bugs can trigger agents to adjust code or prompts, with engineers stepping in when needed. He says agents need a recurring improvement cycle, not just an initial launch.