The model isn't what makes AI work in CX. Your knowledge base is.

Ask which AI is smartest and you're asking the wrong question. The vendors selling the models will tell you so themselves.

Written by

Chris Vavra

Most teams evaluating AI for support start by comparing models. Which vendor's agent is smartest? Whose resolution rate is highest? Which one is built on the newest frontier model? It's a natural question, and it's the wrong one — because the variable that actually determines whether support AI works isn't the model. It's the knowledge base the model reads from.

You don't have to take that on faith. The companies selling these systems say it plainly, in their own documentation. Once you see the pattern, it's hard to unsee.

Why the knowledge base is the deciding factor

Modern support AI doesn't answer from what a language model happened to absorb in training. It uses retrieval-augmented generation — RAG. When a customer asks a question, the system searches your knowledge base for relevant content, then uses the model to phrase an answer grounded in what it found. The model supplies the language. Your knowledge base supplies the facts.

That division of labor is the whole story. If the right article exists and is current, a competent model will find it and answer well. If the article is missing, outdated, or contradicts another article, the model will still produce a fluent, confident response — it will just be wrong. The model's eloquence is constant. Your knowledge base is the variable. Garbage in, garbage out is not a slogan here; it's the architecture.

The vendors admit it

Here's where it gets interesting. The strongest case that knowledge beats model doesn't come from skeptics. It comes from the vendors themselves.

Salesforce. In its own engineering guidance, Salesforce warns that unstructured, unvalidated, or outdated information poisons the retrieval process and produces an agent that can't tell current information from noise — a useless agent, in their words, no matter how capable the underlying model. When Salesforce launched its packaged Agentforce Help Agent, it named poor or incomplete knowledge as the single most common reason AI agents underperform. The company rebuilt its entire data layer, Data 360, specifically to ground agents in trustworthy content — an enormous investment predicated on the idea that the data, not the model, is the constraint.

Intercom. Intercom is refreshingly direct with Fin customers: your real resolution rate depends on knowledge quality, setup, and use case. It's not marketing spin — it's how the product works. Before Intercom turned Fin on for its own support team, it audited and rewrote more than 700 help articles first. And its deployment data makes the dependency measurable: teams that spent two to four weeks cleaning up their knowledge base before launch saw resolution rates roughly 12 percentage points higher than teams that deployed on top of untouched documentation. Same model. Different knowledge base. Double-digit difference.

Zendesk. Zendesk positions knowledge as the foundation for every interaction — self-service, AI agents, and human agents alike — and now markets a tool whose entire pitch is to "fuel your AI with a knowledge base." Its guidance is blunt about the mechanics: the AI answers only what's indexed, and the quality of the answers tracks the quality and coverage of the source content. The ceiling, in other words, is set by the knowledge, not the model.

Three of the largest names in the category, each with every commercial incentive to sell you on the brilliance of their AI, all pointing at the same unglamorous prerequisite: the knowledge base.

The analysts agree. Gartner projects that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data — and that only a small fraction of enterprises consider their data AI-ready in the first place. The projects don't fail because the models aren't good enough. They fail because the ground the models stand on isn't solid.

The proof is in the spread

If the model were the deciding factor, teams running the same product would see similar results. They don't.

Independent analysis of Fin deployments finds resolution rates ranging from around 25% to over 80% across Intercom's customer base — and the difference is almost entirely a function of how well each team maintains the knowledge base Fin pulls from. The teams at the top run continuous content work tied to every product release. The teams at the bottom wrote their articles once and walked away. Identical model, a 55-point spread, explained by knowledge rather than intelligence.

The failure mode is worth understanding because it's counterintuitive. A well-built agent's hallucination rate — inventing facts from nothing — can be under 1%. But that number doesn't cover stale content. When an article is outdated, the AI cites the wrong information accurately and confidently. It isn't confused; it's faithfully repeating a knowledge base that no longer matches reality.

The Cursor incident from 2025 is the clean illustration. The company's AI support agent told users their accounts were being restricted to one device — a policy that did not exist. A session bug was logging people out, and because no article documented the real behavior, the agent filled the gap with an invention. Users began canceling; the story hit the front pages of Hacker News and Reddit within hours; the cofounder issued a public apology. No better model would have prevented it. The missing article would have.

What this means for your team

If the knowledge base is the lever, the implications are freeing — especially for teams without enterprise budgets.

It means the AI arms race is largely beside the point for you. You do not need the most expensive frontier model or the vendor with the highest benchmark to get strong results. You need accurate, well-structured, current documentation and a way to keep it that way. The most decisive input is the one entirely within your control, and it doesn't scale with your software bill.

The catch is that "keep it current" is exactly the part teams have never had the capacity for. Reading years of resolved tickets, finding where reality has drifted from the docs, and writing the articles to close the gap is real work — and it's the first thing that falls off a busy queue. That maintenance problem is itself now an AI-addressable one: the same technology that reads your knowledge base to answer customers can read your ticket history to tell you what the knowledge base is missing. But that's a capability to evaluate carefully, and a subject for another post. The point here is simpler and more foundational.

Stop asking which AI is smartest. Start asking whether your knowledge base is ready for one. The vendors have been telling you which question matters. It's worth listening.

Chris Vavra is the founder of Flexivity AI, which adds AI — including knowledge-base gap analysis and article generation — to the osTicket and Zendesk systems teams already run.