Customer support isn't AI's low-hanging fruit. Your approach to support is.
For the past two years, everyone has been saying the same thing: customer support is the “low-hanging fruit” of AI. High ticket volume, repetitive questions, a knowledge base that’s already structured. You plug an LLM into the documentation, slap a chatbot into a widget, and you’ve diverted half the tickets. Case closed—on to the next topic.
I think we’ve missed the mark.
Support isn’t the easy target. What’s easy is the solution we’ve slapped onto it: plugging a chatbot into existing documentation and calling it innovation.
The low-hanging fruit isn’t the problem we’re solving; it’s the way we’re solving it.
The Paradox of the $500 Billion Project
Before diving into the details, a useful reminder. Companies have poured hundreds of billions into AI over the past three years. AI budgets are expected to rise further in 88% of business functions by 2026 (PwC). And yet:
- The Boston Consulting Group reports that 60% of companies that have invested in AI generate no measurable value.
- The MIT Nanda State of AI in Business 2025 report cites 5% of companies that have truly transformed their operations.
- More than 85% of employees worldwide remain stuck in the “efficiency” and “improvement” stages of adoption. Less than 10% have crossed the threshold into true semi-autonomous collaboration with AI.
The problem isn't technical. The models work. The problem is that we deploy them without changing the underlying workflow. We plug a model into a CRM, we plug a model into a knowledge base, and then we're surprised that nothing changes. It's like fitting a horse with a V8 engine and hoping to invent the automobile.
What Your RAG Chatbot Actually Does
A RAG chatbot takes the user’s question, searches for relevant passages in your knowledge base, and asks the model to summarize them into a response. It’s elegant. It works for static FAQs. And it’s profoundly limited as soon as you get into onboarding or product usage.
Three limitations that keep coming up:
- RAG is fragile. It depends 100% on the quality, structure, and freshness of your knowledge base. If your documentation is outdated, poorly tagged, or incomplete, your chatbot will confidently synthesize false information. The more you scale RAG across complex domains, the more unstable the pipelines become.
- RAG doesn’t reason. It retrieves. It doesn’t anticipate, doesn’t understand underlying intent, and can’t execute a multi-step workflow. It’s a semantic search bar with natural language output. Nothing more.
- RAG doesn’t provide guidance. It gives you a paragraph to read, not a specific action to take. It’s up to the user to translate the text into clicks, within an interface they haven’t yet mastered. This is precisely the point where users were already giving up—even before AI came along.
Disruption theorists (Christensen, and how McKinsey has applied his ideas to heavy industry) have been saying the same thing for twenty years: disruption doesn’t just speed up the old process. It rewrites the entire interaction model from start to finish. Streaming didn’t “make CDs faster”; it eliminated CDs. E-commerce didn’t “make mail-order sales more efficient”; it eliminated the paper catalog.
In SaaS, the new interaction model looks like this: intention → AI-guided execution within the interface.
The user states what they want to do. The system takes control of the interface, highlights the right button, points to the right field, opens the right modal, and guides them step-by-step through its own product. No chat to open. No docs to read. No mental translation between text and clicks. AI guides through the UI instead of describing it.
