Intercom's AI Got Better. It Still Can't Help.
Key Takeaways
- Intercom launched a purpose-built customer service model and published benchmark results placing it ahead of general-purpose alternatives on customer service tasks.
- For SaaS products, that's transformative. For physical products, resolution is limited by a problem no model can solve alone: missing product context.
- A better model on the same data produces better guesses. A grounded system on serial-level product data produces correct answers.
The Fin Apex Announcement
On March 26, Intercom launched Fin Apex 1.0, a purpose-built AI model for customer service, priced at $0.99 per resolved interaction. The headline results are notable: Intercom reports a higher resolution rate than GPT-5.4 and Claude on its customer service benchmark, along with strong conversation volume and revenue growth.
Intercom built this by post-training an open-weights foundation model on proprietary customer service data: many resolved conversations across a large number of companies. The result is a vertical model tuned to how customer service works rather than a general-purpose model.
For SaaS companies, subscription businesses, and digital products, this is genuinely significant. A customer asking "how do I reset my password" or "can I downgrade my plan" has a question that can be fully resolved from documentation, account data, and conversation history. Fin Apex is built for exactly this, where AI agents can deliver real customer support at scale. But for physical products, the challenge is fundamentally different from digital product support for online services.
The Physical Product Problem
Now consider this question: "Is my boiler still under warranty?"
Fin Apex, or any model, needs to know: which boiler (the customer may own three), when it was purchased, which warranty tier applies (standard, extended, installer-registered), whether annual servicing has been completed (a condition on many extended warranties), and whether the product has been transferred to a new owner since purchase.
None of this information exists in a typical help centre, CRM, or conversation history. It exists, if it exists at all, in a serialised product record that many manufacturers don't have.
This is not a model quality problem. It is a data architecture problem. And it is a key reason why physical product support has resisted the AI shift that SaaS support is now experiencing.
Better Models, Same Blind Spots
A model upgrade can lift resolution rates when the bottleneck is intent understanding, retrieval, and reasoning. That kind of gain comes purely from model quality.
For a manufacturer of power tools or kitchen appliances, the same model upgrade tends to move the needle far less. The improvement exists, but the ceiling is fundamentally lower. Not because the model is worse, but because the model has less to work with. This is the difference between generic AI support and context-aware product agents. When customers expect instant warranty responses, the gap between generic and product-aware AI becomes a business issue.
Consider what a support AI for physical products needs to answer correctly:
- "Which filter fits my air purifier?" Requires knowing the exact model and hardware revision. Two units with the same model name, manufactured months apart, may use different filter sizes.
- "Has my warranty expired?" Requires the purchase date, warranty tier, and whether conditions (like annual servicing or product registration) have been met.
- "Can I get a replacement part for this?" Requires knowing whether the part is still in production, whether it's compatible with the customer's specific unit revision, and whether it's covered under warranty.
A generic AI, even a strong one, will either refuse to answer (low resolution) or answer with its best guess based on the product manual (risk of a wrong answer that generates a return, a complaint, or a safety issue).
The Frontier Model Is Easy to Rent
This is the line that frames the whole problem: the frontier model is easy to rent. The hard part is building the grounded system around real products.
Intercom's insight is sound: vertical post-training can produce a model better suited to a domain than a general-purpose alternative. Intercom's vertical is customer service conversations. The vertical for physical product support is the product itself: its serial number, its ownership record, its warranty state, its service history, its revision-specific parts catalogue.
That product-level data layer is what transforms a support interaction from "let me check the manual" to "your unit (serial 7NXK-4920) has a 24-month extended warranty that expires on 15 March 2027, and the HEPA filter compatible with your revision is in stock, want me to order it?" This requires the kind of serial-level verification and parts catalogue access that many manufacturers still lack. Getting spare parts inventory visible to AI support systems is a key unlock for manufacturers trying to capture aftersales revenue.
The first response is what you get from a better model. The second is what you get from a grounded system.
What Physical Product Brands Actually Need
The gap is not AI capability. The gap is the data infrastructure that makes AI capable of reasoning about specific products owned by specific customers.
Four things separate generic AI support from product-aware AI support:
Product identity at the point of contact. When a customer arrives via a QR scan on the product, the system knows the exact model, serial number, purchase date, and warranty tier before the first message.
Retrieval over serial-specific data. Not the help centre, but the product's own record: known failure modes for that revision, parts compatible with that serial number, warranty terms that apply to that purchase.
Callable tools. Warranty lookup, parts ordering, service scheduling, and ownership verification executed within the conversation, not handed off to a human with a ticket.
Accumulating context. Every interaction adds to the product's record. The AI that handles the first support call knows about it when the customer calls back months later about a different issue.
The Right Architecture, Not the Right Model
Intercom has made the case that vertical AI models can outperform general-purpose ones for customer service. A similar principle applies to physical products, but the vertical isn't conversation data. It's product data.
A manufacturer that builds a serial-level product identity layer and connects it to a competent AI model (GPT-5, Claude, Gemini, or an in-house fine-tune) gives that model something to reason with that a model without product context simply does not have. For established manufacturers, this usually means weighing building the right platform vs buying. The model matters. The data matters more.
Higher resolution rates on physical product support don't come from a magic model. They come from standard models paired with product-level context that makes more queries answerable.
What This Means for Manufacturers
Intercom's Fin Apex is a significant step for AI customer service. SaaS companies should pay attention. But if you manufacture physical products such as power tools, appliances, HVAC systems, or fitness equipment, the lesson is not "adopt Fin Apex."
The lesson is: the model layer is increasingly commoditised. The data layer is where defensible advantage lives. Build the product identity infrastructure. Connect it to whichever model serves your economics. The AI will keep getting better. The question is whether it has anything meaningful to work with when your customer asks about their product.
FAQ
Q: Can't Intercom integrate with our product database? Yes. Intercom supports custom data sources and API integrations. But the integration assumes the product data exists in a structured, queryable form. Many manufacturers don't have serialised product records, ownership data, or warranty state accessible via API. The data layer is the prerequisite.
Q: Is this only relevant for consumer products? No. B2B equipment such as industrial machinery and fleet vehicles faces the same problem at higher stakes. A wrong parts recommendation on a commercial HVAC unit is more expensive than a wrong filter suggestion for a home appliance.
Q: What resolution rate should manufacturers target? There is no single right number. The point that holds regardless of the figure: product-level context generally raises the achievable ceiling, and the limiting factor is data completeness rather than model quality.