How Intercom Bet Everything on AI—And Built Fin to 1M+ Resolutions Per Week
Paul Adams is Chief Product Officer at Intercom, leading Product Management, Product Design, Data Science, and Research. He joined when Intercom was just a 14-person company after first advising the startup, and has been on the executive team ever since. Before Intercom, Paul held leadership, product, and UX roles at Facebook (Ads, Platform) and Google (Gmail, Docs, YouTube)—he was on Google’s mobile team when the iPhone launched. He’s the author of the best-selling book Grouped on social software design and co-hosts the podcast Intercom on Product with co-founder Des Traynor.
When ChatGPT arrived in late 2022, Intercom was struggling—five quarters of declining revenue growth, a failed IPO attempt. The leadership team bet the entire company on AI within two weeks of ChatGPT’s release. That bet produced Fin, Intercom’s AI agent for customer service, which now resolves over 1 million customer problems per week with a 65% average resolution rate across 6,000+ customers.
The Top 5 Takeaways from Intercom’s AI Transformation
1. If it doesn’t feel brutal, you’re not going deep enough
Paul is blunt about this: transforming a SaaS company into a real AI company is painful. Intercom wasn’t in a great spot when ChatGPT arrived—they’d had five quarters of declining revenue growth and had abandoned an IPO process. But that pressure became an advantage.
The leadership team made the call in one to two weeks. They ripped up their strategy. Ripped up their roadmap. Told the company it was happening and it wasn’t a choice.
“If you’re a SaaS company who thinks you’re an AI company and you’ve not gone through brutal transformation, you’re not there yet.”
The mistake most companies make? They do the easy, fun stuff—building AI features, experimenting with models, talking to customers about AI—but avoid the hard, messy decisions. Like parting ways with a third of the company because they’re not fit for the new world. Like deleting the marketing calendar and rebuilding from scratch.
Paul took over two-thirds of marketing six months ago and immediately blew the entire thing up. Teams, roadmaps, calendars—gone. “The only way I knew how to build a marketing org fit for this age is to build it from scratch.”
2. The only way to know if you’ve gone far enough is to go too far
Intercom operates on a simple principle: the only way to find a boundary is to cross it.
This shows up everywhere:
Every single designer at Intercom now ships code to production. Zero did 18 months ago. The mandate was clear: this is now part of your job. If you don’t like it, find somewhere that doesn’t require it, and they’ll hire designers who love the idea.
Engineering is on a path to 2x productivity—not through incremental improvements, but by declaring it non-negotiable.
Paul constantly asks: “What would a brand new startup incorporated today do here?” Would they have separate product marketers and content marketers? Or is that the same job now? Would they have both product managers and product designers as distinct roles?
The answer usually points to consolidation, not specialization.
3. How you build software has completely changed
Intercom had principles for building great SaaS products that they’d refined over years. They’d train every new designer and engineer on these principles. They were proud of them. Des had given talks about them.
They had to ban all of it.
The old way: Pick a job to be done → Listen to customers → Design a solution → Build and ship. Execution was certain. Technology was stable. Design was the hard part.
The new way: Ask what AI makes possible → Prototype to see if you can build it reliably → Build the UX later → Ship → Learn at scale. Execution is uncertain. Design is now cheap. The hard part is the AI infrastructure—the RAG system, the custom models, the empirical evaluation.
“This AI layer, our RAG system, has been 3 years in the making by a very talented team. It’s complicated. I do not understand the depths of that RAG system at all.”
The visible UI is now the small part. The invisible AI infrastructure is where the real product lives. That’s a complete inversion of how SaaS products were built.
4. AI products compound—every tiny improvement multiplies
When you’re building workflows that chain multiple AI steps together, success rates multiply. If you’re at 99% accuracy on each of 10 steps, you’re at 90% overall. If you’re at 95% on each step, you’re at 60%.
This is why Intercom obsesses over incremental improvements at every point in Fin’s system. They’ve run hundreds and hundreds of experiments, many of which fail. Sometimes they see an improvement in one part that degrades another.
They’ve built custom models pointed at very specific customer service tasks—not general-purpose, but targeted at discrete steps in the workflow.
“Each single tiny incremental improvement in each of these steps adds up to the highest performing product, adds up to something people can trust, adds up to something people will replace their humans with.”
This compounds into what Des calls the “marketing overhang” problem: lots of companies can demo an AI product, but a demo isn’t a product that works at scale. Intercom has a rule that they won’t do a product launch that isn’t a real demonstration of something they know works at scale.
See: Apple Intelligence, announced June 2024, still waiting on delivery in spring 2026.
5. You now have two companies—and two completely different buyers
If you succeed at the transformation, you end up with a new problem: you’re running two companies.
Intercom the SaaS product: Easy product domain, predictable metrics, clear differentiators, customers who know how to talk about what they need.
Fin the AI product: New product domain, chaotic metrics (“Fin’s grown 300% year-over-year—is that bad?”), customers who don’t know what they need, everything changing so fast that nobody knows what the future looks like.
The buyer has changed too. In the old world, Intercom sold to customer service leaders. Simple.
Now the buying committee includes: the customer service leader (influential but doesn’t make the decision), a C-level executive whose job is AI transformation across the company, and an AI-fluent technical evaluator who can assess whether the product actually works.
These people live in different universes. CEOs and customer service leaders don’t go to the same events, read the same things, or operate at the same zoom level. Paul does dinners with CEOs and trade shows for customer service leaders—the Venn diagram barely overlaps.
Intercom’s Top 5 Mistakes You’ll Probably Make Too
Mistake #1: You won’t reimagine your product—you’ll just add AI to it
Adding AI features to your existing product is not transformation. Fin is a completely different product from Intercom. They have almost nothing in common. You design them totally differently, think about them totally differently.
“Fin usage is eating Intercom usage at times. They’re totally different products.”
Mistake #2: You won’t make self-harming decisions to win
You’ll protect revenue. You won’t want to upset the board. You won’t want to upset your sales leader. You’ll avoid taking a 10% revenue hit to do the right thing for the medium or long term.
But these self-harming decisions are critical. “If it doesn’t feel really painful, you’re not deep enough.”
Mistake #3: You’ll dilute the vision, delay, and convince yourself you’ve done enough
“Q1 is important, let’s do it in Q2.” “Let’s dial down the vision a little.” You’ll do the easy stuff and skip the hard stuff and tell yourself you’ve transformed.
Big company habits and slowness will creep in. You’ll listen too much to customers who say no to AI—the ones who said “we’re going to differentiate on human service forever.” (Those customers now use Fin.)
Mistake #4: You won’t refine your company to fight
You need people who are going to fight for success, fight for these outcomes. There’s tension. There’s disagreement. People fight with each other—positively fighting, not actually fighting.
“If we can do it, any SaaS company can do it. There’s nothing special about us. We just decided to do it.”
Mistake #5: You’ll make these mistakes and deny it to yourself
Intercom’s exec team has the most honest, soul-searching conversations—and they still do every day. Some version of an existential question that might determine whether they succeed greatly or not.
“You really have to look each other in the eye, build deep relationships with your peers and colleagues.”
The companies that don’t transform will slide into irrelevance slowly. New competitors will eat their lunch. Their best people will quit to work at the real AI companies. Eventually, the company will die.
Three years in, AI is inevitable. Whatever you loved about B2B, SaaS and the last decade of mobile and social—it’s gone. Will you change fast enough?
