If you're a developer being asked to "add AI to our support," you're actually being asked to choose between three distinct architectures, and the choice changes your whole implementation.
A rule-based chatbot is just an intent classifier plus a decision tree — cheap to ship, cheap to maintain, and fundamentally incapable of handling anything outside its scripted intents. No memory, no reasoning, no tool calls.
An AI agent is a different system entirely: an LLM with function-calling access to your order, payments, and CRM APIs, plus a memory layer that carries context across turns (and ideally across a customer's full history). Implementation complexity jumps immediately once you add tool calls — you're now building permission scoping (what can the agent actually do without a human sign-off?), idempotency handling for actions like refunds, and a RAG layer if you want it grounded in your own docs instead of hallucinating policy details.
The part most teams underbuild is the escalation state machine. It's tempting to treat handoff as an exception path, but it should be a first-class state with its own triggers: confidence score below threshold, specific keyword matches, repeated unresolved turns, and a context-passing payload that gives the human agent the full transcript instead of a cold start.
On cost: outcome-billed platforms like Intercom Fin charge per resolution (~$0.99), which is easy to reason about at low volume and surprisingly expensive at scale — do the math against your own LLM API costs before assuming "buy" beats "build."
We went deep on the actual INR cost tiers — SaaS vs custom RAG build — for the Indian market in a longer piece: AI-Powered Customer Support in India 2026: Chatbot vs AI Agent vs Human Handoff.
If you're scoping this for a production system, the escalation state machine and the tool-call permissioning are the two things worth over-engineering relative to everything else.
Top comments (0)