Every month, a manager tells me "I want an AI chatbot". In nine cases out of ten, what they need isn't a chatbot — it's an agent. And these two terms, which sound similar in 2026, cover two completely different architectures.

Here's how I draw the line, and how I choose.

The difference, in one sentence

  • A chatbot answers questions from a knowledge base.
  • An AI agent makes decisions, calls tools, and executes actions in your information system.

Put simply: a chatbot talks. An agent acts.

Decision table

NeedChatbotAI Agent
Answer product FAQs on your site✅⚠️ overkill
Qualify an inbound lead and assign to the right sales rep⚠️ insufficient✅
Refund an e-commerce customer❌ can't✅
Summarise and route 200 emails a day to the right teams❌✅
Help customers pick a size✅⚠️ overkill
Search your contracts and flag unusual clauses⚠️ limited✅
Give real-time order status⚠️ only if APIs are plugged in✅ native

Simple rule: as soon as you need to call a tool, modify data or make a conditional decision, it's an agent. Not a chatbot.

Why this confusion dominates the market

Because OpenAI, Claude, Gemini all shipped in 2024-2025 "chatbots" that can use tools — which technically makes them agents. But vendors keep calling them "chatbots" because that's the word buyers know.

Result: SMBs buy "a chatbot" at €1500/month that's actually a poorly wired agent. Or the reverse — they reject the idea of "an agent" because it sounds scary, when that's exactly what they need.

The 5 cases where I recommend an agent (not a chatbot)

1. Lead qualification and routing

The agent reads the inbound request, cross-references with your CRM, scores it against your criteria, and assigns it to the right sales rep with a pre-written brief. The rep gets everything they need in under a minute.

2. Level-1 support with action

A customer asks for the status of their order? The agent queries your logistics system and answers. They want a refund within policy? The agent triggers the refund, sends the confirmation, and closes the ticket. What a chatbot can't do.

3. Triage and synthesis of internal documents

Contracts, CVs, supplier quotes, meeting notes: the agent reads, extracts key points, compares against your reference data, and surfaces what stands out.

4. Sales copilot

For every prepared prospecting call, the agent gathers signals (LinkedIn, website, news), compiles a synthetic briefing, and suggests 3 contextual conversation angles.

5. E-reputation monitoring with action

The agent monitors mentions and reviews, detects a sentiment drift, and — depending on threshold — posts a pre-approved response, or alerts a human with full context.

When a chatbot is enough (really)

  • Stable product FAQ, volume > 500 queries/month, highly repetitive topics.
  • Navigation help ("where to find such info in your knowledge base").
  • Qualifying a contact form without complex business logic.

In these cases, a chatbot well wired to a knowledge base and an LLM is enough — and costs much less to build than a real agent.

What I build, in practice

On most of my AI agent engagements, I start with a minimal agent with narrow scope (a single use case, 50 to 200 conversations per day max), and let it run a few weeks under human supervision before expanding. Never the other way around.

The classic mistake: trying to launch an agent that "does everything". It never works in 2026 — evaluations drift, you no longer know what's good or bad, and at the first real issue, no one can debug.

How to pick for your case

Three questions are enough:

  1. Does my question require fetching data from a third-party tool? If yes, agent.
  2. Does the answer trigger an action (create, modify, delete)? If yes, agent.
  3. Does the volume and variability of requests justify the engineering cost? If no, start with a chatbot.

If you're not sure, the first conversation is free — I'll tell you what fits your case, and I never push an agent when a chatbot is enough.

To go further