Podcast
The CEO Must Be the Chief AI Officer
Y Combinator Startup Podcast
- CEO Should Personally Own The AI Agenda
- Make the CEO the chief AI officer and own the company’s AI strategy personally.
- Pedro Franceschi argues the CEO must understand AI bounds better than anyone and focus on what only humans can do. (Time 0:00:00)
- December Models Felt Like Electricity
- Pedro compares December’s LLM progress to the invention of electricity and says that moment made coding harnesses suddenly practical.
- He recounts holiday experiments where reasoning models and tools revealed agentic product patterns and personal automations like buying a movie ticket via an agent. (Time 0:03:50)
- Secure Agents With An HTTP Proxy And LLM Judge
- Secure agent actions at the network layer by HTTP proxying requests and making traffic auditable.
- Brex built Crab Trap to record agent HTTP traffic, auto-generate policies, and use an LLM as a judge for ambiguous requests. (Time 0:05:56)
- Build Harnesses For Nontechnical Token Maxing
- Build harnesses so non-technical teams get token-maxer productivity without coding.
- Pedro describes OpenClaw-style markdowns and editable skills that let agents self-bootstrap capabilities for non-engineers. (Time 0:11:00)
- Most Teams Haven’t Gone AI First Yet
- Token-maxing lags because many treat models as too expensive or precious instead of defaulting to AI-first problem solving.
- Pedro suggests the AI pill test: if you don’t default to AI for everyday problems you miss re-wiring workflows. (Time 0:14:35)
- Start With Minimal Surface Area
- Favor minimal surface area for early products; compress intelligence into one clear interaction.
- Pedro points to Brex’s terminal API and Stripe as examples where founders nailed one interaction instead of broad UIs. (Time 0:18:38)
- Human Judgment Still Finds Out-Of-Distribution Signals
- Models lack the out-of-distribution signals founders glean from customer conversations, so human judgment remains essential.
- Pedro advises spend time on tasks only humans can do: extracting implicit signals and choosing which problems matter. (Time 0:20:50)
- Customer World Model Captures Total Information Awareness
- Brex built a customer world model by ingesting every customer touchpoint from clicks to calls.
- That model predicts what a customer needs next and surfaces issues the human team missed. (Time 0:27:27)
- Track Token Spend Like Any Other Resource
- Measure and attribute token spend across products, customers, and employees to understand ROI.
- Brex built MagPi to tag every dollar of token usage and run analytics on model-driven activity. (Time 0:30:40)
- Redesign Processes End To End With AI
- When adopting AI, redesign end-to-end processes instead of bolt-on automation.
- Pedro describes reimagining KYC to KYC leads early and changing targeting and risk orientation across the funnel. (Time 0:33:50)
- Company-Wide AI Is A Turnaround Effort
- Implementing AI across a large company is effectively a turnaround that requires CEO-level energy to refound identity and metrics.
- Pedro splits AI work into product, operational, and corporate legs and urges rethinking roles and success signals. (Time 0:39:10)
- Make Escalations Fast To Avoid Organizational Antibodies
- Make escalation and experimentation fast so internal antibodies don’t kill promising AI experiments.
- Pedro emphasizes CEOs can break glass quickly and should desensitize escalation paths to try risky changes. (Time 0:43:27)
- Turn Agent Failures Into Continuous Evals
- Bake evals into every human–agent interaction so failures become actionable bugs that improve agents.
- Brex turns flagged agent conversations into evals that trigger agents or engineers to update prompts and code. (Time 0:46:12)
- Wake Up Asking Why AI Can’t Solve Your Problem
- Start by asking every morning why your personal problem can’t be solved with AI and iterate on the 20% that fail.
- Pedro recommends measuring token use and focusing founder time on problems models can’t yet solve. (Time 0:51:38)