Every week I talk to a CX leader who tells me they're "exploring AI." When I ask what problem they're trying to solve, the answer is usually the same: response time, agent efficiency, ticket volume.
These are the right problems. The approach is almost always wrong.
The chatbot trap
The fastest way to implement AI in CX is to bolt a chatbot onto your existing support flow and call it transformation. The chatbot handles simple queries. Agents handle the rest. Handle time drops. Everyone celebrates.
Six months later, the chatbot is deflecting contacts that customers are immediately re-routing to a human. CSAT scores for bot-touched interactions are lower than human-only ones. Agents spend more time cleaning up after escalations. The "efficiency" was measurement fraud.
This happens because chatbots are an output. Before you choose an output, you need to understand your input.
Your data is the strategy
Before any AI deployment, the most important question is: what does your contact data actually tell you?
Not the category labels someone attached to tickets three years ago. Not the summary your outsourcer sends every month. The real signal: what are customers saying, in their own words, at each moment of the journey?
Most organizations have never done this analysis properly. When you do it — clustering contacts semantically, mapping them to product touchpoints, measuring their frequency and resolution cost — you find that the distribution is almost never what you expected.
Typically: 20–30% of your contact volume comes from 3–5 specific product gaps or communication failures. Fix those, and your contact volume drops materially — before you've touched AI at all.
Where AI actually wins
Once you understand your contact data, AI becomes a precision instrument rather than a blunt object. The applications that genuinely work:
Triage and routing. Classification models that read incoming contacts and route them — correctly, first time — to the right team or queue. This is unglamorous and enormously effective.
Agent assist. Real-time suggestions to agents during conversations — relevant knowledge base articles, resolution patterns from similar cases, escalation indicators. This accelerates experienced agents and dramatically shortens ramp time for new ones.
Summarization and post-contact work. Automating the summary and tagging of resolved contacts eliminates a significant portion of after-call work. At scale, this is hours per agent per week.
Proactive outreach. Using signal from your product data to contact customers before they contact you — when a payment fails, when an order is delayed, when a process they initiated hasn't resolved. Prevention is far cheaper than resolution.
None of these require a chatbot. All of them require clean, well-structured contact data and a clear understanding of your contact drivers.
The real investment
The organizations that succeed with AI in CX are the ones that invest first in data infrastructure. A taxonomy of contact reasons that maps to actual product events. Integration between your CRM, your product analytics, and your support platform. A culture where CX insights flow upstream to product and engineering.
That investment is harder than buying a chatbot. It also lasts longer, scales better, and actually fixes the problem instead of rerouting it.
AI in CX is a data architecture problem. Solve the architecture. The AI decisions get easy.