Michel Tricot is Cofounder and CEO of Airbyte.
Walk into any office in San Francisco right now, and you will see something that would have seemed impossible two years ago. Marketing teams are building AI agents to monitor campaign performance and take action without waiting for an analyst. Sales reps have agents that pull context from CRM, support tickets and Slack threads before every call. Even HR teams are delegating onboarding workflows to agents that coordinate across a dozen systems.
This is not limited to engineering teams or AI startups. In San Francisco, the shift from “using AI tools” to “deploying AI agents that do real work” has already happened.
But step outside the Bay Area, and most organizations are still in the chatbot era, asking models questions and getting answers. The agent era has not arrived for them.
The question is: Why?
The Gap Is Not About Awareness
Every executive I talk to at conferences in London, Paris or New York knows the terminology. They have seen the demos. Many have launched pilots.
The gap is not awareness. It is infrastructure.
San Francisco companies have a structural advantage that rarely gets discussed. They operate on modern software stacks. Their engineering teams are accustomed to stitching systems together. And critically, they sit in an ecosystem where the tools for building agents are being built in real time by people they can grab coffee with.
The rest of the world does not have that luxury.
Why Agents Break Outside The Bubble
When an AI agent needs to do something useful, it needs more than a language model. It needs real-time access to the systems where work actually happens, it needs to understand how data relates across those systems, and it needs permission to act. Combined, these are why most agent projects stall before reaching production.
Fragmented data access. Most enterprises have data spread across dozens of platforms, and each one has its own way of storing, organizing and granting access to that data. Even companies in San Francisco struggle with this. Connecting business systems reliably is not a solved problem anywhere. Integrations break, permissions change and data formats shift without warning. The difference is that Bay Area companies have had earlier access to emerging platforms that handle this complexity for them. For most organizations, the expectation is still to build and maintain every connection from scratch, which is why so many agent projects never get past the prototype stage.
Missing context. Access alone is not enough. An agent pulling a customer record from Salesforce, a support ticket from Zendesk and a billing event from Stripe needs to understand these are about the same customer, even when the name appears differently in each system. Without a way to connect identity and meaning across platforms, agents hallucinate, miss signals or produce outputs that look plausible but are wrong. This is what I call “context engineering,” and it is the single biggest reason agents fail in production.
Security and governance at scale. Every system an agent touches requires its own login credentials, its own permissions and its own rules about what can be accessed and by whom. In regulated industries, you also need a complete record of every action an agent takes. Building this from scratch for each business system is exactly the kind of unglamorous work that kills agent projects before they deliver value.
The Model Is Not The Bottleneck
There is a persistent misconception that AI falls short because the models are not good enough. The typical response is to wait for the next model release or feed more data into training.
Models are improving every quarter. That is not where the constraint lives.
The constraint is that most organizations have no reliable way to give agents the context they need to operate. Not just data access, but a structured, real-time understanding of how a business actually works. Who is this customer? What is their history with us? What actions are available, and which ones require approval?
Outside San Francisco, organizations need this solved at the platform level, not rebuilt from scratch by every company individually.
What The Agentic Enterprise Actually Looks Like
The companies pulling ahead, regardless of geography, share a common pattern. They have stopped treating AI agents as isolated experiments and started treating them as operational systems that need real infrastructure underneath.
Revenue teams use agents to monitor pipeline health and surface risks before they become losses. Customer teams rely on agents that coordinate across support, billing and usage data to trigger retention workflows. Operations teams deploy agents that investigate issues across systems and automate the resolution.
In each case, the agents are not just reading data. They are acting on it. Updating records, sending notifications, creating tickets, closing loops. This is what separates a useful agent from a fancy chatbot.
How To Close The Gap
The AI agent gap between San Francisco and the rest of the world is real, but it is not permanent. It is an infrastructure gap, not an intelligence gap.
For any organization serious about making agents operational, there is a simple test. Pick your most common workflow and count how many business systems an agent would need to touch. If the answer is more than three, you are looking at an infrastructure problem that no amount of model improvement will solve.
In San Francisco, people aren’t waiting for that problem to be solved. They’re building it today by architecting solutions that accumulate knowledge across business systems. The good news is that this San Francisco energy doesn’t have to stay here. Companies that embrace the attitude of forward-thinking ingenuity can close the gap.
Wherever they’re based.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
