Intercom Built a Vertical AI Model. What Sales Teams Can Learn.
Intercom's custom AI model outperforms GPT-5.4 and Opus 4.5 at customer service. The lesson for sales: generic AI tools will lose to domain-specific ones. Here's how to tell the difference.
Illustration generated with DALL-E 3 by Revenue Velocity Lab
Intercom just did something most AI companies talk about but don't actually do. They built their own model.
Not a wrapper around GPT. Not a fine-tuned prompt chain. A custom-trained AI model called Apex, built by their 60-person AI group using billions of customer service interactions as training data. The result: it outperforms GPT-5.4 and Claude Opus 4.5 on resolution rates, speed, and cost.
One gaming customer went from 68% to 75% resolution overnight. That's a 22% reduction in unresolved conversations from swapping the model alone — no workflow changes, no new features, just a better engine underneath.
The phrase Intercom's CEO Eoghan McCabe used: "The age of vertical models is here."
He's right. And the implications go well beyond customer service.
A 12-person team built the pipeline of a 30-person team. The system discovers. Your team closes.
What vertical actually means
Most AI tools in sales today are horizontal. They take a general-purpose model — GPT, Claude, Gemini — and add a layer on top. Custom prompts, maybe some fine-tuning, a nice interface. The underlying AI knows everything about everything and nothing deep about your specific domain.
A vertical model is different. It's trained on domain-specific data. In Intercom's case: billions of real customer service conversations. What questions people ask. How agents respond. Which responses actually resolve issues and which ones don't. The model doesn't just generate plausible text. It knows what works in its domain because it's seen millions of real outcomes.
Intercom's competitive advantage isn't the interface or the features. It's the data flywheel. Every conversation Fin handles makes the model better at handling the next one. Competitors starting from scratch need 18 months to build equivalent capabilities, according to McCabe. Not because the technology is secret, but because the data doesn't exist yet.
Why this matters for sales AI
The sales AI market is still mostly horizontal. Most tools are prompt wrappers. They take GPT or Claude, feed it your prospect's LinkedIn profile and company description, and generate an email. The email is grammatically correct and sounds professional. It's also generic — because the model has no concept of what a buying signal looks like, what ICP fit means in practice, or why this company should be contacted this week instead of next month.
The gap between horizontal and vertical in sales AI will look like this:
| Horizontal (generic model + prompts) | Vertical (domain-trained model) | |
|---|---|---|
| Email quality | Grammatically correct, professional tone | Contextually relevant, signal-aware |
| Company targeting | Based on firmographic filters you set | Based on learned ICP patterns + live signals |
| Timing | Sends when you tell it to | Surfaces opportunities when signals appear |
| Learning | Same output quality in month 6 as month 1 | Gets better every week from your team's decisions |
| Explanation | "Here's a draft based on the prompt" | "This company raised Series B yesterday and posted 3 sales roles this week" |
The bottom row is the tell. If your AI tool can explain why it's recommending a company with domain-specific reasoning — not just "it matches your filters" — there's a vertical model underneath. If it can only generate text, it's horizontal.
The "good enough" trap
Here's where most teams get stuck. A horizontal AI tool that writes decent emails feels good enough. It saves time. The output is polished. Why pay more or switch to something that claims to be "vertical"?
Because the gap compounds.
In month one, the difference between horizontal and vertical is marginal. Both write emails. Both save time. But vertical learns from your team's behavior — which prospects you skip, which ones you prioritize, which outreach gets replies. Horizontal stays static.
By month six, vertical is targeting companies your team wouldn't have found manually, at moments when buying signals are fresh. Horizontal is still generating emails based on the same filters you set on day one.
Intercom saw this pattern in customer service. Fin was already resolving 2 million issues per week with generic models. Apex made it materially better — not by adding features, but by replacing the engine with one that understands the domain.
Intercom generates nearly $100M in recurring revenue from Fin. They didn't build Apex because generic models were broken. They built it because "good enough" stops compounding. The same logic applies to sales: a tool that learns from your pipeline will eventually outperform one that just generates text.
How to evaluate whether a sales AI tool is actually vertical
Three questions to ask any vendor:
"What data was the model trained on?" If the answer is "we use GPT/Claude with custom prompts and RAG," that's a wrapper. Wrappers are fine for general tasks. They won't compound over time. A vertical answer sounds like: "Our model is trained on X million sales interactions, including outreach outcomes, reply patterns, and conversion data."
"Does the system improve from my team's usage?" If your send/skip decisions, reply rates, and meeting outcomes feed back into the model, it's vertical. If the output is the same whether you've used it for one day or one year, it's horizontal.
"Can it explain why it recommended this company right now?" Generic tools can describe a company. Vertical tools can explain timing: "This company raised Series B yesterday, posted 3 sales roles this week, and matches your ICP pattern based on the last 50 companies your team converted." If the explanation relies on live signals and learned patterns rather than static filters, there's domain knowledge underneath.
What happens next
McCabe predicts competitors need 18 months to catch up to Apex. That window exists because the data flywheel takes time to build. You can't shortcut billions of domain-specific interactions.
The same dynamic will play out in sales AI. The tools that start building proprietary training data now — from real outreach outcomes, real ICP patterns, real buying signals — will have a compounding advantage. The ones still running prompts on top of GPT will feel fine today and fall behind by 2027.
For sales teams choosing tools right now, the practical advice: don't just evaluate what the tool does today. Evaluate whether it has a mechanism to get better from your usage. If it does, it's vertical. If it doesn't, it's a wrapper — and wrappers are about to become commodities.
Frequently Asked Questions
Better pipeline starts with better targeting
Most teams waste cycles on accounts that were never going to close. Enter your URL to see which companies in your market actually match your ICP.
Enter your URL → ICP-matched companies found in 30 seconds
Matches found across 50M+ companies · 50M+ company database · No login · Free