Inside Intercom’s AI Revolution with Fergal Reid

• Fin is evolving into one of the world’s most advanced AI support agents, now handling voice, images, APIs, and real-world actions.

• Reid believes AI remains deeply underhyped, with most people underestimating its coming impact by an order of magnitude or two.

• Intercom’s breakthroughs in RAG, tooling, and task execution are paving the way for AI systems that don’t just answer questions; they perform real operations end-to-end.

Let’s dive in. No floaties needed.

Revenge of the Nerds

Fergal Reid, Chief AI Officer at Intercom

Fergal is the Chief AI Officer at Intercom. He has a PhD in the field and has worked in machine-learning roles at several fast-growing startups.

Over his career, he has built teams, developed ML products from early prototypes into revenue-generating tools, and spent the past six years leading Intercom’s machine-learning efforts.

For those unfamiliar, what does Fin do?

Fin is an AI-powered customer support agent. It answers customer questions using all the information available within a business—things like the company’s knowledge base, documentation, and other internal systems.

Today, Fin can also interact with APIs, process images, and perform fairly sophisticated actions. You can configure it to use different tools and data sources, enabling it to plan with an LLM and resolve customer issues end-to-end.

In short, it’s one of the most advanced AI customer support agents in the world. Fin now generates over $50M in annual recurring revenue and continues to grow rapidly.

Why do you believe AI is underhyped?

I still think AI is underhyped, although that’s starting to change. People often get excited about breakthrough technology, but many won’t fully believe in it until they see products that actually change their lives.

This gap between innovation and real-world impact takes time. Think of penicillin. It was discovered years before it became widely available. The same pattern happens with AI. There’s always hype at first, but most people underestimate how transformative it will be once it’s productized.

Even now, I think most people underestimate AI’s potential by an order of magnitude or two. They see ChatGPT and think, “It’s great for writing essays,” but don’t realize how much existing white-collar work could be automated or enhanced by AI-powered systems.

There’s also confusion about whether AI progress will continue to accelerate or plateau. But people often struggle to reason about what happens if it keeps improving. Until they experience it firsthand, they simply can’t picture how big the impact could be.

What do you think causes people to fear AI as it keeps getting smarter?

Some fear of AI is reasonable, but I don’t think fear is the main factor. The real issue is that people struggle to grasp just how big of a shift AI represents. When something feels too big, whether it’s good or bad, it can overwhelm them, so they tend to shut it out.

Adapting to change this quickly creates a kind of ‘future shock’. It’s emotionally hard to update your view of the world that fast. Many people working in AI have gone through that phase of realizing its transformative potential.

The key is to stay calm and open-minded. It helps to consider multiple possible futures at once, thinking through both the risks and the opportunities. That kind of balanced thinking takes effort, but it keeps you grounded and thoughtful instead of reactive.

What’s been Fin’s biggest technical or product breakthrough?

In the early days, our biggest breakthrough was learning how to actually work with large language models and build reliable retrieval-augmented generation (RAG) systems. That foundation shaped everything that came after.

Since then, there have been major advances in the models themselves. They follow instructions more reliably, handle longer prompts, and run faster and cheaper. Those improvements have made it possible to design far more complex systems.

One of the toughest challenges for us was building a ‘tasks’ interface that allowed Fin to connect with APIs and take real-world actions. It took a long time to get that right, but in the past six months we’ve crossed a threshold. Companies are now deploying Fin for live, complex workflows where it doesn’t just answer questions but actually takes action, like querying systems, verifying information, and even raising a parking garage gate once payment is confirmed.

That shift from conversation to real-world execution has been huge. It shows that both the technology and the industry’s willingness to deploy it have matured enough for AI to handle real operations safely and effectively.

What separates serious AI products from those just riding the hype?

There’s definitely a lot of hype and money flowing into AI, and everyone’s trying to get their share. But the real difference between serious AI products and those just chasing the hype comes down to one question: can you actually deliver something that works and solves a real problem?

It’s easy to build a flashy prototype. It’s much harder to create something reliable that functions in the messy complexity of the real world. That’s where technical skill, good product design, and risk management all intersect.

I often think of self-driving cars. A decade ago, they looked incredibly promising, and everyone could see the potential value. But it took years and billions of dollars to make them work in real conditions. AI builders today need to be clear about whether they’re in that kind of game.

That said, the range of what’s actually possible has expanded dramatically. So while caution is important, I think a mindset that assumes “this might be possible” rather than “this is impossible” makes a lot more sense in today’s world.

Additional Reads

• Fergal Reid’s author profile @ Intercom,

• Speaker at Turing Fest (Building products in the Age of AI)

• Podcasts @ Fin (“Closing the loop” and “Shipping reliable AI actions”)

• Alumnus In Conversation with Fergal Reid.

Quick Poll

Rate This Edition