Pedro describes product work as shortening the path from customer signals to shipped solutions. At Brex, AI helps teams analyze years of customer calls and other unstructured data, develop hypotheses, and test them with inexpensive prototypes. As building gets easier, he says teams need stronger judgment about which problems are worth solving.
Stop Building AI Agents. Build AI Employees Instead
Speakers
Pedro Franceschi, Brex co-founder and CEO, talks with Peter Yang about AI-driven product work, virtual employees, agent security and costs, and his personal AI workflow.
Pedro says he spends about half his time reviewing work to set a quality bar, with teams applying shared principles and product skills before reviews. Brex separates problem alignment from solution alignment and emphasizes managing the work itself. AI helps PMs find customer and sales signals, while people still judge how proposed work fits company priorities.
Pedro presents Jim, a recruiter built on OpenClaw that connects to Greenhouse, evaluates resumes and profiles against role criteria, and helps source candidates, filter inbound applications, and analyze recruiting data. Recruiters can ask Jim questions in Slack, review candidate scores and evidence, and give feedback that improves its decisions. Pedro says Brex compares recruiter and Jim decisions to monitor their agreement.
Pedro describes CrabTrap, an open-source HTTP proxy that monitors an agent’s network requests and applies static rules or an LLM judge to decide whether requests fit policy. He also demos Magpie, an internal tool for breaking down AI usage by business area, model, employee, and caller. Pedro says Brex first gave employees broad token access, then built analytics to identify costs and opportunities to optimize.
Pedro expects more work to shift toward people using AI, and says companies may spend a larger share of payroll on tokens if the return justifies it. His Autopilot system gathers Slack, email, and meeting signals into Markdown files organized around people and programs, with tasks and interaction history. He argues AI products can charge for completed work, rather than only for access to software.
Takeaways
- Use AI to compress the path from customer evidence to a tested product decision.
- As building becomes faster, prioritize one high-value problem instead of adding features by default.
- Give virtual employees a defined role, useful context, feedback loops, and a clear human review path.
- Control agents by inspecting and governing their network requests, including actions that require contextual judgment.
- Measure AI costs by use case and business outcome before optimizing token usage.
Mentioned in this talk
Search inside this talk
What was claimed 25
- foundedPedro Franceschi founded
BrexCo-founder, CEO“Pedro Franceschi, co-founder and CEO of Brex.”
0:32inThe Most AI-Pilled CEO We KnowYouTube Verified4 sources· Talk, Web· archived
- ceo ofPedro Franceschi CEO ofBrexCEO
“Pedro Franceschi, co-founder and CEO of Brex.”
0:32inThe Most AI-Pilled CEO We KnowYouTube Verified2 sources· Talk
- role atPedro Franceschi is CEO atBrex
“Pedro, co-founder and CEO of Brex.”
YouTube Unverified1 source· Talk
- built
“We focused on that layer and built CrabTrap.”
7:18inThe Most AI-Pilled CEO We KnowYouTube Verified2 sources· Talk
- featureRiverside: Records each participant locally in 4K and provides AI tools for editing recordings and generating clips.
“It records each person locally in 4K resolution.”
YouTube Unverified1 source· Talk
- featureAI agents: Focus AI on a defined job so it behaves and feels like a coworker, with work delivered as an outcome.
“Focus on this idea of building virtual employees versus open-ended agents.”
YouTube Unverified1 source· Talk
- metricJim: A manager evaluates each candidate against role-specific criteria.
“A manager evaluates every candidate against role-specific criteria.”
YouTube Unverified1 source· Talk
- metricJim: Has been running since February.
“It's been running since February.”
YouTube Unverified1 source· Talk
- integrates with
“Fetches GitHub data to complete the full picture on the applicant.”
YouTube Unverified1 source· Talk
- integrates with
“It ingests candidates, parses resumes and rich LinkedIn profiles.”
YouTube Unverified1 source· Talk
- integrates withJim integrates with
Greenhouse“First, it syncs Greenhouse, our ATS.”
YouTube Unverified1 source· Talk
- uses
“This is all just OpenClaw running on a harness.”
YouTube Unverified1 source· Talk
- integrates with
“If you go into the Slack channel, you can see people interacting with Jim.”
YouTube Unverified1 source· Talk
- uses
“We use OpenClaw for this.”
YouTube Unverified1 source· Talk
- featureJim: Sources candidates, filters inbound applicants, and provides a recruiting analytics layer.
“It effectively does three things: sources people, filters all inbound ... and serves as an analytics layer.”
YouTube Unverified1 source· Talk
- built
“We're not in the business of building recruiting tools, but this is something we built.”
YouTube Unverified1 source· Talk
- metric
CrabTrap: Open source.“CrabTrap is open source, so any company can use it.”
YouTube Unverified1 source· Talk
- featureCrabTrap: HTTP proxy that controls agent network traffic and evaluates requests against static rules and LLM screening.
“We've built an HTTP proxy that sits around the agent and controls all its network traffic.”
YouTube Unverified1 source· Talk
- built
“We built a tool called Magpie that helps you control token costs.”
YouTube Verified2 sources· Talk
- featureMagpie: Analyzes AI token consumption by model, employee, caller, and use case to help control costs.
“We built a tool called Magpie that helps you control token costs.”
YouTube Unverified1 source· Talk
- metricMagpie: Transaction tagging costs about $0.10 per call and about $28,000 per month in the demo data.
“It costs 10 cents per call today, and we spend roughly $28,000 a month on it.”
YouTube Unverified1 source· Talk
- metricMagpie: Planned for availability to Brex customers soon.
“Very soon, every company that spends on Brex will be able to have this for its own AI spend.”
YouTube Unverified1 source· Talk
- metricBrex: Unlimited token budget for engineers, with a few caps for exceptional usage.
“We give them an unlimited token budget. We have a few caps because some people have used the process.”
YouTube Unverified1 source· Talk
- uses
“I call it Autopilot. It's an OpenClaw harness.”
YouTube Unverified1 source· Talk
- feature
Autopilot: Collects messages, email, and meeting notes, then organizes signals into programs, people, and action items stored as Markdown files.“It all starts with signals: Slack messages, emails, and meeting notes.”
YouTube Unverified1 source· Talk
