For the past three years, the conventional wisdom in enterprise AI has been simple: use the biggest, most capable frontier model available. OpenAI, Anthropic, Google — pick your flavour and pipe it into your product. That approach is now being challenged.

On March 26, 2026, Intercom quietly announced that it had replaced the frontier model powering its Fin AI customer service agent with something it built itself. The result: Fin Apex 1.0 now resolves customer issues at a higher rate, with fewer hallucinations, faster responses, and at roughly one-fifth the cost compared to running GPT-5.4 or Claude Opus 4.5 directly.

The benchmarks are hard to ignore. Fin Apex 1.0 achieves a 73.1% resolution rate — meaning 73 out of 100 customer issues are fully resolved without a human stepping in. GPT-5.4 and Claude Opus 4.5 both sit at 71.1% on the same measure. That gap might sound small, but for a company handling over two million customer conversations every week, a 2% improvement in resolution rate is a significant operational and cost reduction.

One of Intercom’s largest customers, in the gaming industry, saw their resolution rate jump from 68% to 75% overnight after switching to Apex — a 22% reduction in unresolved conversations. According to Intercom, that kind of single-event improvement had never happened before since they launched Fin.

The model also responds 0.6 seconds faster than its closest competitor, and produces 65% fewer hallucinations. Those aren’t academic numbers — in customer service, a hallucinated answer creates a support ticket, damages trust, and costs money.

How Did They Build It?

Intercom trained Fin Apex 1.0 on an undisclosed open-weights base model, using post-training on what they describe as “billions of human and agent customer service interaction data points” accumulated since launching Fin. The company grew its AI research team from 6 to 60 people over three years to make this possible.

The key insight is straightforward in retrospect: general-purpose frontier models are optimized to be good at everything. But most real business problems don’t require a model that can write poetry and solve differential equations. They require a model that is deeply accurate in one specific domain. By narrowing the scope and training on proprietary, domain-specific data, Intercom produced a model that outperforms the generalists at their own game — in customer service, at least.

As Intercom CEO Eoghan McCabe put it: the age of vertical models is here.

What This Means for Business

This development signals something important for every business currently using or evaluating AI tools: the competitive advantage in AI is shifting away from which frontier model you have access to, and toward the quality and uniqueness of your own data.

A few things to understand:

Frontier models are becoming table stakes, not differentiators. When any company can call the same OpenAI or Anthropic API, that model is not your competitive edge — it’s the baseline. The companies building lasting AI advantages are the ones training models on their own operational data, tuning them for specific workflows, and iterating faster than a general-purpose vendor can.

Your data is your moat. Intercom’s Apex works because Fin has processed billions of real customer service interactions. That proprietary data is something no competitor can replicate quickly. The same logic applies in any industry: a healthcare company with years of patient interaction data, a financial services firm with a decade of client queries, a logistics company with millions of routing decisions — all of that data becomes a genuine competitive asset once you have the capability to train on it.

Small, specialized models beat large, general models for defined tasks. Apex is not a trillion-parameter model competing with GPT-5 on the MMLU benchmark. It’s a purpose-built model that costs far less to run and outperforms frontier models specifically on customer service resolution. That economics story is what makes vertical models commercially viable at scale.

The “plug in GPT and go” phase is ending. For businesses still treating AI as something you subscribe to and deploy out of the box, the gap between them and competitors building domain-specific capabilities will widen. The tooling to build and fine-tune custom models is maturing rapidly, and the cost to do it is dropping.

Where Enterprise DNA Fits

This is exactly the thinking that underpins what we do at Enterprise DNA. General-purpose AI tools are a starting point, not an endpoint. The businesses that will get the most from AI over the next few years are the ones investing now in understanding their data, cleaning and structuring it properly, and building AI capabilities that reflect their specific operational context.

For teams looking to go beyond off-the-shelf AI tools, our Omni Apps service focuses on custom AI application development — building tools trained and tuned for your specific workflows rather than generic use cases. And if you’re not sure where to start, Omni Advisory helps business leaders build a practical AI roadmap grounded in what’s actually deployable and valuable for your industry.

The Intercom announcement is a proof point, not an outlier. Vertical AI is not a niche — it is where this is all heading.