A founder asks "should we add a chatbot or an AI agent to our support?" and gets six different answers depending on who they're talking to. The terms are used loosely, the marketing is confusing, and the actual technology differences matter a lot more than most people realize.

Here's what each term actually means, what each approach is good at, and where each one falls apart.

Chatbots: The First Generation

Chatbots have been around since the 2010s. The early ones were decision trees: scripted flows that asked yes/no questions and followed predefined paths. "Is your issue related to billing or technical support?" Click billing. "Is it about a charge or a refund?" Click charge. Eventually you reach a canned answer.

These still exist. Lots of companies use them. They work for extremely predictable scenarios where every customer follows the same path. Think: restaurant reservation bot, basic FAQ responder, appointment scheduling.

Where they break: anything that doesn't fit the script. If a customer's problem spans two categories, or they phrase their question differently than expected, or they have a follow-up that the tree doesn't anticipate, the bot either gives a wrong answer or says "I didn't understand that."

Modern chatbots use LLMs (GPT, Claude) instead of decision trees. They can handle natural language, understand context across a conversation, and generate human-like responses. This is a massive improvement over scripted bots.

But LLM chatbots have their own problems. They hallucinate (make up information with confidence). They're expensive per message ($0.01 to $0.10 per response, which adds up at scale). They can't reliably access backend systems to check order status or process refunds. And they sometimes say things that contradict your policies.

AI Agents: The New Buzzword

"AI agent" is the 2025-2026 buzzword. Everyone uses it. Salesforce calls their product Agentforce. Sierra calls their product an AI agent. Intercom calls Fin an AI agent. The term has been stretched to mean almost anything.

The original concept: an AI system that can take actions, not just generate text. A chatbot tells you "your order is on the way." An AI agent looks up your order in the system, checks the shipping status, and if it's delayed, initiates a reshipping request automatically.

The action-taking part is the key distinction. AI agents are supposed to do things: process refunds, update accounts, create tickets in external systems, trigger workflows. Chatbots just talk.

In practice, most "AI agents" in 2026 are LLM chatbots with tool access. They use function calling to interact with APIs: look up orders, check account status, maybe process simple transactions. The quality varies enormously. Some are reliable and well-constrained. Others hallucinate tool calls or take wrong actions.

The risk with AI agents is higher than with chatbots. A chatbot that gives a wrong answer is frustrating. An AI agent that processes a wrong refund or updates the wrong account is dangerous. Most companies deploying AI agents add human-in-the-loop approval for consequential actions. Which is smart, but it also means the "agent" is really a chatbot that sometimes suggests actions for a human to approve.

Classifiers: The Third Approach

A classifier doesn't generate responses or take actions. It reads the customer's message and determines what they want. That's it. Then it routes the request to the right resolution.

"I want to cancel my subscription" gets classified as intent: cancellation. The system routes it to the cancellation flow, which could be automated (show the cancel page with retention offer) or human (connect to a retention specialist).

"My order hasn't arrived" gets classified as intent: order status. The system looks up the order and provides the tracking info.

"I'm having a bug where the app crashes when I upload a photo" gets classified as intent: bug report, sub-category: upload. The system creates a ticket for the engineering team with the relevant details.

The classifier approach works because customer support requests follow patterns. There are a finite number of things customers ask about (Supp categorizes into 315 intents). For each intent, there's a known best response or routing path. The classifier connects the dots.

Advantages over chatbots: no hallucination (it's not generating text), cheaper per query ($0.20 vs $0.05 to $0.10+ for LLM responses at scale), faster (100-200ms vs 1-3 seconds for LLM generation), and more predictable (the response for each intent is pre-defined and verified).

Advantages over AI agents: lower risk (it doesn't take actions independently, it routes to the right handler), more transparent (you can see exactly why a message was classified as X and routed to Y), and easier to debug when something goes wrong.

Disadvantages: it can't handle free-form conversations. If a customer wants to have a back-and-forth discussion about their issue, a classifier provides the initial routing but doesn't maintain a dialogue. For that, you need a human or an LLM.

Which Approach Fits What

Use a scripted chatbot if: you have very simple, very predictable scenarios with low volume. Restaurant reservations, basic appointment booking, opening hours queries. Cost: low. Risk: low. Sophistication: low.

Use an LLM chatbot if: you want conversational AI for knowledge base lookups and FAQ responses, and you can tolerate occasional wrong answers. Best paired with a strong feedback loop and human escalation. Cost: medium to high. Risk: medium (hallucination). Sophistication: medium.

Use an AI agent if: you want automated action-taking for specific workflows and you can build reliable tool integrations with human approval gates. Best for high-volume, well-defined actions (order lookup, refund processing, account updates). Cost: high. Risk: medium to high (wrong actions). Sophistication: high.

Use a classifier if: you want fast, accurate routing with predictable responses and low cost. Best for teams that want to automate the simple stuff and route the complex stuff to humans. Cost: low. Risk: low. Sophistication: medium.

The Hybrid Future

Most mature support teams will end up using a combination. A classifier handles the initial intent detection and routing. Simple intents get automated responses. Complex intents go to a human, possibly assisted by an LLM that drafts a response for the agent to review and edit.

The classifier is the front door. It decides what happens to each message. Everything downstream (automated action, knowledge base lookup, LLM-drafted response, human agent) is chosen based on the classification.

This is already how Supp works. Classify the message (315 intents, 92% accuracy, under 200ms), route it to the right handler, and let each handler do what it's best at. The customer gets a fast, accurate resolution whether the handler is a bot, an API call, or a human.

The industry is moving this direction because pure chatbot and pure agent approaches both have failure modes that damage customer trust. Classification with smart routing avoids the worst failure modes of both while keeping costs low.