The CEO Must Be the Chief AI Officer
In a Nutshell
The CEO must personally lead AI adoption as chief AI officer rather than delegating it, because only the CEO has authority to redesign company architecture, processes, and self-identity from first principles instead of layering AI onto existing workflows. The core shift is treating AI as an agentic loop with tools and security boundaries like network-level proxies, then building harnesses that let both engineers and non-technical teams operate virtual employees rather than just chatbots. Companies that pause to ask how they would rebuild every function if starting today—with the same opportunity but different technological possibilities—will capture discontinuous advantages; those that don't will be left optimizing outdated systems.
These notes were generated by AI and may contain inaccuracies.
The CEO needs to be the chief AI officer rather than delegating this to an engineering or product team. A good proxy for how to spend your time is focusing on things that only you can do that the models cannot do. This includes refounding the very concept of the company's self-identity.
Pedro Franchesci is the co-founder and CEO of Brex, which started in the Winter 17 batch and became one of the most important fintech companies of the last decade. After meeting Pedro at YC for lunch, the team went down a rabbit hole of building on their own AI setups. The lunch generated significant token consumption and was the precursor of G Brain. The speaker was still working on GStack as a 2013 web 2.0 engineer who time-traveled to the AI tools of January 2026.
The craziest realization was that most people in software had been treating the LLM like a very precious and expensive thing. This led to putting the agent inside a Foxcon factory with excessive control and if statements. The alternative approach is giving agents more freedom, described as the Esalen Institute rather than Foxcon's factory. Every single good AI product is fundamentally an agent loop with tools, skills, and a model.
The first encounter with LLMs came during the pandemic when someone gave API access to GPT-3, which felt like a research project. Interest grew with ChatGPT, but things became truly interesting with reasoning models and tools. December marked the point where electricity was invented, referring to OpenAI 4.5 and open models. This was the tip of the spear where coding harnesses actually started to work. Cloud Code had existed for about a year before but wasn't valuable until this point.
During the holiday break, experimentation revealed the fundamental shift: standing 200 years into the future, five or six months after electricity was invented, most people were still playing with candles and questioning what could be done with fire. The steam engine was still 20 years away, but electricity already existed.
OpenClaw represented the next step in realizing that good AI products are agentic loops with tools. This was implemented in Brex's own product. On the personal side, markdowns proved capable of taking things very far through configuring and automating many aspects of life. An example involved buying a movie ticket entirely through OpenClaw using a Brex card provisioned through an API. The point wasn't efficiency compared to booking online in 10 seconds, but demonstrating the capability.
The initial personal setup gave read access to everything and created OAuth tokens for email, Slack, and other systems without writing permissions. The results were surprisingly capable. The next challenge was enabling write access into systems. The Google security team initially resisted due to security concerns. Four weeks were spent solving the hardest problem of security.
The insight was to focus on the network layer rather than controlling what tools the model calls. Agents have their own wills and desires and will attempt things at network boundaries. The approach was to build Crab Trap, which was open sourced about two months prior, to secure agents in production at Brex.
Crab Trap works by HTTP proxying the entire network boundary of an agent. When a request goes through, it becomes auditable. Another agent can analyze the traffic and create a policy to determine whether traffic should be allowed. Models trained on hundreds of billions of web documents understand HTTP traffic particularly well because they literally learn on the web. The ability of the model to watch a thousand requests and understand what is happening exceeded expectations.
After recording a day of agent traffic, a policy can be built that automatically approves certain requests. For uncertain cases, an LLM acts as a judge to determine whether a request should be approved based on the agent's policy. An example is a recruiting agent called Jim with its own policy. 98% of requests go through automatically, while 2% require LLM judgment.
This security approach made the team comfortable experimenting more aggressively with agents at the enterprise level. This is particularly difficult inside capital-intensive environments. While Brex does not yet do the most aggressive things with AI and does not use it extensively on customer data, there is no reason YC companies should not be at the bleeding edge.
The network-level proxy approach proved prescient, with similar patterns appearing in credential brokering solutions like agent vault. The first version of Crab Trap included credentials vault but this was removed to focus on doing one thing well. The LLM-as-a-judge capability was the determining factor for trusting the system in production. The security team, known for being rigorous, eventually accepted this approach.
Brex is not in the business of building HTTP proxies but needed to build this proxy to reach the bleeding edge of AI capabilities. The goal is for someone to build a better version that Brex can adopt.
AI adoption inside most companies follows three tiers. Tier one consists of token maxers who are engineers pushing significant amounts of code, typically living inside coding harnesses. Tier two includes average engineers building some things but achieving roughly one-tenth the productivity of token maxers. Tier three encompasses the entire rest of the company, which typically interacts with AI in Google search mode using chatbots with a few MCPs or G Suite equivalents.
The thesis was that significant value comes from the harness for token maxers, and the goal was to build an equivalent harness for non-technical teams. OpenClaw demonstrated the ability to self-bootstrap agent capabilities through editing skills and markdowns and setting up the environment around the agent.
The company-wide adoption approach rejects simply giving people a few MCPs. People want a virtual employee that exists on Slack, has an email, can be invited to meetings, can join meetings, and take notes. The infrastructure needs to support this use case. The harness for non-technical users probably looks more like OpenClaw than a coding model.
An example involved installing Aqua voice and using it through Telegram or Slack with Jared and three engineers plus someone from the events team to organize 60 dinners with 20 attendees each from Startup School, involving 21 partners and visiting partners at YC. The claw started doing the work after the conversation was captured, building markdown and performing analysis without anyone opening cloud code.
Cloud Code is not magic but rather a harness around the same models accessible via API. Claw can use Cloud Code, and it really prefers to use Codex these days. ACP helps with this preference.
Founders are often shy about burning tokens. Gary mentioned that tokens are expensive. While fortunate to afford token spending, the speaker imagines being 14 or 12 when starting to code seriously and would be token maxing in the cheapest way possible. Chinese models offer a reasonable approach, and there is a large hobbyist community building local LLMs on gaming rigs.
At Brex, token cost management and spend management are active topics. Beyond cost, more people should be complaining about max plan limits, but only about 0.1% of Twitter does so, indicating people are still early.
The AI pill test is whether you default to AI first when any problem appears in your life. At a certain point, this becomes second nature and rewires how you think. Many people have not yet intimately understood the bounds of problems using AI despite how cheap this has become. When starting a company today, the premise would be why it cannot be just you, with token consumption likely much higher than planning for three, five, or seven people.
The fundamental constraint is not AI as a cost savings mechanism. The unlock is that the fabric of the company looks different when boundaries become type systems, interfaces, and agents talking to each other rather than people. Building code with new agents is well understood, but living in a world where intelligence is on tap and the default answer is to solve problems with AI first, even if suboptimal, compounds over time.
Brex's early MVP had no web UI and was entirely terminal-based. The pattern of successful companies shows minimal surface area. Stripe's early days were literally an API. Brex had no UI, just a terminal. Airbnb's website was a simple form. DoorDash followed a similar pattern. The surface area with customers was small, and founder bandwidth was spent nailing one single interaction pattern.
The risk with AI is that agency behind choice goes away, creating a lack of discipline on what matters. Intelligence is compression. Great ideas fit on a napkin. When someone pitches an idea, it must fit on a napkin.
The hardest part of building a company is talking to customers and extracting unspoken signals from conversations. The AI lens means asking why AI cannot solve a problem. When talking to customers, they will not give answers shaped into prompts that an LLM can use to output a billion-dollar product. They give local optimum answers based on their worldviews and constraints. The job now is having the wisdom to choose what you want because execution is no longer the bottleneck.
Pre-AI, personal bandwidth allowed exploring three ideas in parallel. Now the approach would be picking a broader universe of things to do for early initial exploration, with the lens being why AI cannot solve it and which signal is not in the model. The signal typically comes from the customer.
When engaging with customers, the approach should focus on getting into the headspace of the individual rather than relying solely on synthetic customer models. While synthetic customers and customer role models are valuable once substantial knowledge about the customer exists, they are less effective when understanding is still developing. At Brex, the company initially sold to founders because the founders understood their own problems, but as the company scaled, they began selling to finance teams, which required building entirely new mental models around different value systems.
There is an intangible element where significant competitive advantage still originates. A useful proxy for allocating time is identifying tasks that only you can perform—things models cannot replicate. This represents one such area where human insight provides unique value.
Founders who successfully navigate pivots often possess a loop involving emotional connection and the ability to simulate others' thinking. The book Others in Mind addresses how people with strong emotional connections can model what others are thinking. Founders who develop this empathy can identify what customers are not verbalizing. Gary's principle of making the implicit explicit captures the subtle signals and desires that customers express indirectly.
A challenge with depending on models is that founders often do not know the right incantation or set of prompts to ask, because they lack sufficient domain knowledge to even formulate the correct questions. There exists a meta layer beyond basic prompting. Language models are not magic—they are trained on a specific corpus of information optimized for particular benchmarks and outcomes. The biggest pitfall is having no sense of how much training data the model has seen for the exact question being asked.
If models provided sampling frequency information for each answer, trust calibration would differ substantially because the distribution of training data varies dramatically across topics.
A significant blind spot for application engineers working with LLMs is understanding what lies outside the model's training distribution. Companies like Merkore and other data companies focus on identifying these blind spots. Data labeling companies attempt to understand model pitfalls, but identifying gaps requires expertise in the domain.
Founders face a particular challenge: the curse of knowledge combined with not knowing the bounds of knowledge can create false confidence in understanding something that has not actually been modeled.
After creating Gbrain, AI usage shifted toward retrieval-augmented approaches. For problems or questions where information is likely out of distribution, the process involves deep research across papers and articles, ingesting everything into a git repo, then retrieving and summarizing usable context. This approach works for any domain—for example, reading the top 500 books about running a restaurant to build a comprehensive information compendium.
Brex is building a customer world model that ingests every customer touchpoint—from dashboard button clicks to email content, phone conversations, and support calls. The system consolidates this data to predict what the customer needs next, what they should be thinking about, and issues they will face but have not yet encountered. This remains fundamentally a distribution problem.
As long as there are limits on RAM, there will be jobs. A model cannot have enough parameters to contain everything possibly needed in distribution—there are not enough atoms in the universe. World models used to train models carry inherent biases from their designers. One consideration is making LLMs work for people who look very different from the model builders, such as the average finance person in the US.
Mental models embedded in LLMs out of the box carry more bias than commonly acknowledged. When testing AI for accounting categorization, the first example generated was AI capex, reflecting that model builders primarily think about AI capex as a default category.
A recurring observation is how often crypto maxims apply to AI thinking. One relevant maxim is that the models will be the worst they will ever be. Another is telling people who hate AI coding to have fun coding at 1x speed.
The thesis around long inference suggests there will be substantially more inference than markets currently price in. Public market expectations around semi-supply chains project around 10,000x growth. Current adoption data shows 84% of the world has never used AI, 16% have used a free chatbot at least once, 0.3% pay $20 monthly for AI, and an even smaller fraction use agents. This distribution supports being long inference.
AI will become the biggest expense in companies. Even if token costs decrease by 10x, usage will increase 10x, maintaining significant cost. Brex built an internal system called Magpie to attribute every dollar of token spend to products, internal tools, or employees, enabling ROI analytics on token usage.
Data from Brex shows surprising patterns in token consumption. Within a 10-mile radius around relevant hubs (including New York), there is substantial token consumption correlating with faster revenue growth. The gap between companies in these concentrated areas and everything else is significant—even large companies with substantial budgets spend far less on tokens than would be economically rational (potentially 10x, 20x, or 100x more).
Two and a half years ago, engineering and product leaders at Brex were asked what they would do differently if starting the company in 2024. The answer was that everything would change. This realization is maddening because it reveals how outdated the fabric of the company, product building, and internal processes have become.
The approach to AI adoption involves pausing to recognize the discontinuity—not just in how problems are solved, but in the definition of the problems themselves. This requires stepping back and rethinking from first principles.
When redesigning the KYC process, the initial approach was to build an agent automating the existing workflow—80% automatable, 20% manual. Instead, the entire onboarding process was redesigned end-to-end. This revealed that deal qualification happens at the beginning of the funnel. With KYC available for free, leads rather than just customers can be qualified, introducing risk orientation earlier in the funnel and changing targeting strategy. The same principle applies to credit decisions.
Many competitors took the approach of layering AI on top of existing processes rather than redesigning from scratch. The positive discontinuities came from setting aside old approaches and asking how the process would be designed if starting the company today.
The culture difference between power users of Arc Linux versus Ubuntu parallels AI adoption patterns. Ubuntu users expect things to work out of the box, similar to ChatGPT users. Arc requires hardcore customization, maintaining skills, and markdown configuration—but when successful, enables building exceptional systems. Valve's SteamOS, which powers a Nintendo Switch-like experience, was built on Arc Linux through extensive driver customization and over-the-air updates.
Two years ago, people were working less than they are now. The question becomes where the productivity gains are. A CFO at a large public company noted increased lines of code pushed alongside rising token consumption. While ROI measurement on tokens matters, this analysis misses the historical context—standing six months after electricity was invented.
When electricity was first invented, ROI calculations would have shown poor returns. Accountants would have concluded electricity would never succeed. People adopted it not for cost savings but because of curiosity about possibilities. Working until 2 a.m. experimenting with workflows and models happens because the possibilities are visible, independent of immediate monetary return.
Tokens will likely become effectively free over a 100-year timeframe, similar to how electricity costs are not considered in daily life outside data centers.
Bringing the broader company along on the AI journey requires recognizing that the CEO needs to be the chief AI officer. This is not an engineering team responsibility or product team responsibility—it requires understanding the bounds of the technology better than anyone. Experiencing the limits of the technology daily is necessary to understand what is possible.
The CEO is the only person nobody can say no to except the board, and the board is not in the weeds. The KYC team would never think of using KYC technology to score leads—the only people who can think about reorganizing the system itself have context of the whole.
The single most important question any CEO needs to answer is: imagine transporting the current state of technology back to when the company started, with the same opportunity but entirely different possibilities for building. How would the company be built differently? Comparing this against current reality requires suffering in silence initially, then determining what to do about the gap.
This means redesigning onboarding, growth engines, customer acquisition, user communication, and data synthesis from scratch. It requires refounding the company's self-identity, how functions operate, and how success is structured.
AI as an umbrella encompasses three categories discussed internally:
- Product AI: The product shipped to customers
- Operational AI: Things directly affecting ability to serve customers at scale, including customer success, risk, onboarding, and operations
- Corporate AI: How people work internally
These three agendas matter differently depending on company timing. The guiding question is why everything cannot be solved with AI, then problem-solving from that limit. For large companies that are not AI-native, this requires a turnaround mindset.
Companies are designed not to change, similar to Foxconn factories optimized for extreme efficiency creating one thing perfectly back to back. There is a certain amount of broken glass required. It is 10x easier for the CEO to break glass than an executive, and 10x easier for an executive than an employee.
When someone attempts AI implementation and encounters resistance because a use case has not been tested, the CEO can resolve the issue in 10 seconds by understanding risks and guardrails, while an employee might spend 10 hours escalating or never succeed. The conclusion is often to build the product the old way because it works and avoids interpersonal friction.
Escalation paths need desensitization because companies build antibodies against disturbances to social cohesion. Making escalations faster enables trying new approaches with the recognition that the biggest risk is missing the opportunity to rethink problems from a starting-from-scratch perspective.
Regarding the Jack Dorsey view that every company is building its own little company AGI, domain specificity matters. A single company model with every piece of data without judgment or lens is not the right approach. The virtual employee analogy is more useful: building an agent exceptional at understanding everything that matters about a specific customer represents a well-defined problem with clear boundaries and APIs. This creates self-contained agents with defined dependencies and interactions.
Separate specialized agents for distinct functions rather than creating monolithic systems. Three distinct agents should handle different responsibilities: one that manages the product roadmap, another that interacts directly with customers, and a third that reasons about customer conversations and translates insights into product direction. Functional and domain knowledge remain essential even as AI systems advance. The structure of knowledge itself does not fundamentally change with AI adoption.
Focus on measurable impact rather than model sophistication alone. Evaluate AI systems by asking concrete questions: how many people are actually using it, is it displacing the need to hire a person inside a company, and how many hours is it saving. A model that appears impressive on paper fails if it does not deliver tangible time savings or resource reduction.
The customer world model now powers the client sales team, demonstrating practical application. The model provides total information awareness by surfacing details from support tickets, executive travel issues, and account states that would otherwise require manual research.
Cross-modal evaluations provide valuable insights when run against each other. Every human interaction within the company should function as an evolution opportunity. When an onboarding agent encounters KYC exceptions it cannot resolve, the manual review process should generate an evolution case that improves the system. An expense agent at Brex creates a bug whenever a conversation flags an issue or feels suboptimal, triggering another agent to modify the codebase and prompts accordingly.
The objective is creating a self-learning system where agents improve daily rather than remaining static after initial deployment. Companies often invest heavily in getting an agent operational but neglect mechanisms for continuous improvement, which represents the most significant unlock.
A dream cycle processes everything nightly, identifying patterns, determining what information belongs where, and surfacing recurring issues. The challenge lies in embedding this dream cycle into products, agents, and shipped features.
Building agents for personal use serves as user research. One such system now contains 350,000 markdown pages. The recognition that AI addresses the human bottleneck mirrors earlier skepticism about Neuralink that shifted upon experiencing AI's capabilities directly. Voice memos to OpenCloth represent the most used developer interface currently. Telegram's constraints forced greater agent intelligence rather than building additional client-side functionality.
Context organization remains the primary bottleneck. Once context is properly structured, models can perform sophisticated operations. LSD (Lateral Syntactic Drift) works by taking vectors and explicitly avoiding combinations within the same conceptual cone, instead requiring orthogonal or seemingly random concept pairings. The system generates hundreds of these combinations, then ranks them for coherence, producing high-quality outputs like top-ranked tweets.
Ingesting a 60 gigabyte Google Takeout archive yielded approximately 4,000 meaningful emails after filtering out noise. These emails capture consequential thinking and life moments that enhance agent understanding.
The electricity analogy frames the current moment: when electricity was invented, it initially seemed limited, yet knowing its future trajectory would have changed decisions dramatically. Founders should maintain a to-do list on their computer where every problem encountered prompts the question of why AI cannot solve it. While 80% of problems can be addressed with existing chatbots, the remaining 20% should drive building new solutions. This process develops texture and intuition for technological possibilities that cannot be gained without daily hands-on engagement.
Measure token consumption and push the limits of what a single person can accomplish. The premise should be determining why something cannot be built by one individual. Founders will encounter limitations of current models, but the key questions are identifying which problems merit solving and what limitations still require direct human intervention. Founders should spend time on choosing what matters, talking to customers, and injecting signal that models lack, while architecting around the premise that electricity is only six months old.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
