Intercom is taking an unusual gamble for a traditional software company: building its own AI model.
The huge customer service platform based in Dublin, Ireland, founded 15 years ago End Apex 1.0 announced on Thursday, a small, purpose-built AI model that the company says outperforms leading edge models from OpenAI and Anthropic in the metrics most important to customer service.
The powers of the model Intercom’s existing Fin AI agentwhich already manages more than a million conversations with clients weekly.
According to benchmarks shared with VentureBeat, Fin Apex 1.0 achieves a 73.1% resolution rate (the percentage of customer issues resolved completely without human intervention) compared to 71.1% for GPT-5.4 and Claude Opus 4.5, and 69.6% for Claude Sonnet 4.6. That roughly 2 percentage point margin may seem modest, but it’s wider than the typical gap between successive generations of cutting-edge models.
"If you run large service operations at scale and have 10 million customers or $1 billion in revenue, a 2% or 3% delta is a really big number of customers, interactions, and revenue." Intercom CEO Eoghan McCabe told VentureBeat in a video call interview earlier this week.
The model also shows significant improvements in speed and accuracy. Fin Apex delivers responses in 3.7 seconds (0.6 seconds faster than the next fastest competitor) and demonstrates a 65% reduction in hallucinations compared to Claude Sonnet 4.6.
Perhaps most surprising for enterprise buyers: It costs about a fifth of the cost of using Frontier models directly and is included in the existing Intercom price. "by result"based pricing structure for your existing customer plans.
What is the base model? Does it even matter?
But there is a problem. When asked to specify what base model Apex was built on and its parameter size, Intercom refused.
"We won’t share the base model we used for Apex 1.0, for competitive reasons and also because we plan to change base models over time." a company spokesperson told VentureBeat. The company would only confirm that the model is "in the size of hundreds of millions of parameters."
That’s a remarkably small model. For comparison, Meta’s Llama 3.1 ranges from 8 billion to 405 billion parameters; Even efficient, open-weight models like the Mistral 7B dwarf the billion-dollar scale that Intercom describes.
It remains an open question whether Apex’s performance claims hold up in that context, or whether the benchmarks reflect optimizations possible only in limited, domain-specific applications.
Intercom says it learned from backlash AI coding start cursor faces when critics accused the coding wizard of burying the fact that its Composer 2 model was built on tight open-weight models rather than proprietary technology. But the lesson Intercom learned may not satisfy skeptics: The company is transparent that it used an open-weight basis, but not which one.
"We are very transparent that we have" used an open weights model, the spokesperson said. However, refusing to name the model while claiming it is transparent is a contradiction that will likely draw scrutiny, especially as more companies tout "property" AI that equates to post-trained open source bases.
Post-training as a new frontier
Intercom’s argument is that the base model simply doesn’t matter much anymore.
"Previous training is now a kind of commodity," McCabe said. "The frontier, so to speak, is actually in the post-workout. Post-workout is the hard part. You need proprietary data. You need proprietary sources of truth."
The company subsequently trained its chosen base using years of proprietary customer service data accumulated through Fin, which now resolves 2 million customer queries per week. That process involved more than simply feeding transcripts into a model. Intercom created reinforcement learning systems based on real resolution results, teaching the model what successful customer service really looks like: the appropriate tone, judicious decisions, conversational structure, and, crucially, how to recognize when an issue is truly resolved and when a customer is still frustrated.
"Generic models are trained with generic data from the Internet. Specific models are trained with hyper-specific domain data," McCabe explained. "Therefore, it is logical that the intelligence of generic models is generic and that the intelligence of specific models is domain-specific and therefore operates in a way that is far superior for that use case."
If McCabe is right that the magic is entirely in the post-workout, the reluctance to name the base becomes harder to justify. If the base is truly interchangeable, what competitive advantage does the secret protect?
A $100 million bet paid off
The announcement comes as Intercom’s first AI twist appears to be working. Fin is approaching $100 million in annual recurring revenue and is growing at 3.5x, making it the fastest-growing segment of the company’s $400 million ARR business. Fin is expected to account for half of Intercom’s total revenue by early next year.
That trajectory represents a notable change. When Fin was released, its solve rate was only 23%. Today, it averages 67% among customers, with some large enterprise deployments seeing rates as high as 75%.
To make this happen, Intercom grew its AI team from about 6 researchers to 60 over the past three years, a significant investment for a company that McCabe admits was "in a really bad place" ahead of its AI pivot. The average growth rate of public software companies is around 11%; Intercom expects to achieve 37% growth this year.
"We are by far the first in the category to train our own model." McCabe said. "There is no one else who is going to have this for a year or more."
The speciation and specialization of AI
McCabe’s thesis aligns with a broader trend that Andrej Karpathy, former AI leader at Tesla and OpenAI, recently described as the "speciation" of AI models: a proliferation of specialized systems optimized for specific tasks rather than general intelligence.
Customer service, McCabe maintains, is especially suited to this approach. It’s one of two or three enterprise AI use cases that have found genuine economic traction so far, along with coding assistants and potentially legal AI. This attracted more than $1 billion in venture funding for competitors like Decagon and Sierra, and opened up the space, in McCabe’s words, "ruthlessly competitive."
The question is whether domain-specific models represent a lasting advantage or a temporary arbitrage that frontier labs will eventually close. McCabe believes labs face structural limitations.
"Perhaps the future is that Anthropic will have a large offering of many different specialized models. Maybe that’s what it seems," said. "But the reality is that I don’t think generic models are going to be able to keep up with domain-specific models at this point."
Beyond efficiency to experiment
Early adoption of enterprise AI focused largely on cost reduction, replacing expensive human agents with cheaper automated agents. But McCabe sees the conversation turning to quality of experience.
"At first it was like, ‘Shit, we can do this for so much cheaper.’ And now they’re thinking, “Wait, no, we can give customers a much better experience.”" said.
The vision extends beyond simply resolving queries. McCabe envisions AI agents functioning as consultants: a shoe retailer’s robot that not only answers shipping questions but offers style advice and shows customers what different options might look like on them.
"Customer service has always been shit." McCabe said bluntly. "Even at the best brands, you are left waiting for a call, they take you through different departments. Now there is an opportunity to provide a truly seamless customer experience."
Prices and availability
For existing Fin customers, upgrading to Apex comes at no additional cost. Intercom confirmed that customer pricing remains unchanged: users continue to pay per outcome as before, at $0.99 per resolved interaction, and automatically benefit from the new model.
Apex is not available as a standalone model or through an external API. It can only be accessed through Fin, meaning companies cannot license the model independently or integrate it into their own products. That restriction may limit Intercom’s ability to monetize the model beyond its existing customer base, but it also keeps the technology proprietary in a practical sense, regardless of what the underlying base model turns out to be.
What’s next?
Intercom plans to expand Fin beyond customer service into sales and marketing, positioning it as a direct competitor to Salesforce’s Agentforce vision, which aims to provide AI agents throughout the customer lifecycle.
For the SaaS industry as a whole, Intercom’s move raises uncomfortable questions. If a 15-year-old customer service company can build a model that outperforms OpenAI and Anthropic in its domain, what does that mean for providers still relying on generic API calls? and if "post-training is the new frontier," As McCabe insists, will companies that claim to have made progress face pressure to show their work, or will they continue to hide behind competitive secrecy while touting transparency?
McCabe’s answer to the first question, featured in a recent LinkedIn postIt is clear: "If you cannot become an agent company, your CRUD application business has a diminishing future."
The answer to the second remains to be seen.
