Pedro Franceschi on Building Specialized AI Employees at Brex

Brex co-founder and CEO Pedro Franceschi talks with Peter Yang about product management, narrowly scoped “virtual employees,” agent security and AI spending. The video description says the data Pedro shares is Brex demo data. Riverside sponsors the episode; Linear is also listed as a sponsor.

Franceschi judges a product organization by how quickly it turns customer signals into a useful shipped solution. He says AI is blurring traditional product, engineering and design roles, but argues that the judgment of all three remains important.

As building becomes easier, he says prioritization matters more: teams should identify the actual bottleneck and solve a smaller number of problems deeply. Franceschi says he spends about half his time reviewing work to set a quality bar.

Brex distinguishes problem alignment from solution alignment. Product teams use customer calls and other unstructured signals to formulate hypotheses, then validate them with prototypes. A shared repository of review skills encodes recurring product principles and questions, leaving reviewers to focus on the specifics.

Franceschi describes an internal agent-building approach centered on bounded roles rather than a single open-ended assistant. His example is Jim, a recruiting agent built on OpenClaw with skills, persistent state, connections and additional infrastructure.

According to Franceschi, Jim synchronizes with the applicant-tracking system, gathers candidate information, evaluates it against role-specific criteria and surfaces people for recruiters to review. The interface presents a score and supporting context, and the team compares the agent’s judgments with human decisions. Those scores are model outputs, not proof of a candidate’s merit or a substitute for fair human review.

The recruiting system also provides analytics and can report to staff through Slack. Franceschi says the team improves its behavior from feedback. He adds that open-source contributions can be a positive signal in hiring, but their absence is not automatically a negative one; Brex’s engineering interviews include a building exercise using AI tools.

For network controls, Franceschi demonstrates CrabTrap, an open-source HTTP proxy. He describes static rules for known requests and an LLM-based judge for other traffic, plus an audit trail. A suggested policy can be generated from observed traffic and reviewed; network policy and model judgments have limits and require testing rather than guaranteeing security.

The next tool, Magpie, tracks AI consumption and costs across corporate, operational and product use. Franceschi shows how a team might break spending down by use case or model and investigate changes. He describes plans to connect usage with actual company spend; future availability is his stated intention, not a promise.

Asked about token budgets, he says Brex initially allowed extensive experimentation, with some caps, before optimizing usage. Examples of later optimization include caching and choosing a less expensive model where appropriate. High token use by itself does not demonstrate productivity or return on investment.

Franceschi predicts changing jobs and more people building with AI, while acknowledging uncertainty about employment. His estimates of productivity and future adoption are opinions, not established outcomes. He expects companies to need people to help implement these changes.

He shows a personal OpenClaw system called Autopilot that collects signals from Slack, email and meeting notes. It organizes them around people and projects, produces action items and retains links to source interactions. The underlying project files are Markdown stored in a repository, with scripts providing a searchable interface.

On software businesses, Franceschi argues that AI changes how providers deliver value: a product can sell completed work as well as a tool. He describes agents, tools and feedback loops as a common pattern. This is his strategic view of the market, rather than a guarantee that any particular model will succeed.