"Chatbot" and "AI agent" have collapsed into the same word, and that's making it really hard to buy the right thing. Vendors slap "AI-powered" on a 2018-era decision tree and call it an agent. Other vendors ship a true autonomous system and still pitch it as a "smarter chatbot" because that's what people search for.
I build both. So here's what I tell clients on discovery calls when they're trying to figure out which one they actually need.
The line in the sand: can it take actions?
Forget the model. Forget the marketing copy. The real difference between a chatbot and an AI agent comes down to one question:
Can the thing take actions on your systems on its own — or is it just talking?
A chatbot answers questions. It might be powered by GPT-4 or Claude or a tiny intent classifier from 2019. Doesn't matter. If all it does is respond with text, it's a chatbot.
An AI agent can do all of that plus reach out and change the state of your business. Book a calendar slot. Push a row to your CRM. Issue a refund. Place an order in your POS. Update a Notion database. Send an SMS. Transfer a call. The talking is the surface; the acting is the value.
What this looks like in practice
Chatbot example
A customer asks "what are your hours?" The chatbot reads from a knowledge base and replies with the hours. Helpful. But if the customer then says "can I book a reservation for Saturday at 7?" — the chatbot says "please call us" or "click here to visit our reservations page." Dead end.
Agent example
Same conversation. The agent reads hours from your KB, then when the booking request comes in, it queries your live calendar API, sees Saturday 7pm has 4 open seats, asks for the party size and contact info, books the slot, and confirms via SMS — all in the same conversation. The customer never hits a dead end.
Both used the same LLM. Only one was wired to take actions.
The four capabilities that make something an agent
- Tool use. The system can call APIs, query databases, or trigger workflows. It has hands, not just a mouth.
- Memory. It remembers context across a conversation (and ideally across sessions) — what the customer said five minutes ago, what they bought last month.
- Planning. Given a goal, it can break it into steps. "Book a meeting" decomposes into "check calendar → propose times → wait for response → confirm → send invite."
- Decision-making. It can choose between paths. "Customer is angry → escalate to human." "Order has substitution → confirm before submitting." It's not just following a script.
If the thing you're evaluating only has #1 — but is missing memory, planning, or decision-making — it's a chatbot with API access. That's still useful for some workflows, but it's not an agent.
When a chatbot is the right answer
Chatbots aren't dead. They're just narrower. A chatbot is the right call when:
- The conversation is one question, one answer (FAQ, hours, location, status checks).
- The cost of a wrong answer is low.
- You don't need it to integrate with anything.
- You want a $20/month off-the-shelf tool, not a custom build.
For a five-page brochure site, a chatbot is fine. For a business that loses revenue when a phone rings unanswered, you want an agent.
When you need an agent
You need an actual agent when:
- The conversation has to end with a transaction. Booking, ordering, refunding, qualifying.
- The agent has to integrate with your systems. CRM, POS, calendar, ticketing, payments.
- The conversation has branches. "If they ask about X, route to A; if they ask about Y, route to B; if they're frustrated, transfer to a human."
- You want to compress headcount. Replacing a part-time receptionist or a tier-1 support seat is agent territory, not chatbot.
Why people get this wrong
Two reasons:
One: the SaaS chatbot category invented "AI" as a feature flag. Intercom Fin, Drift, every helpdesk tool now ships an "AI" mode. Some of these are real agents. Many are GPT wrappers around the existing FAQ flow with no real tool access. The marketing is identical for both.
Two: the model providers (OpenAI, Anthropic) ship demos that look like agents but are running in a sandbox. The model isn't actually plugged into your CRM in the demo — that part is the work. And the work is non-trivial. It's prompt engineering, integration plumbing, error handling, escalation logic, and tuning. That's why a real custom agent costs more than a chatbot subscription.
The buying question that cuts through
If you're evaluating an "AI" product, ask the vendor this:
"Show me the agent completing the full transaction end-to-end — from first message to the final action firing in our CRM/POS/calendar."
If they can't, or if they show you the chat half and then tell you the integration is "configurable" or "available with the Enterprise tier," you're looking at a chatbot.
What we build at Vulcani
Custom agents, not chatbots. Every Vulcani deployment ships with real tool access — your CRM, your calendar, your POS, your support stack. The agent doesn't say "please call us." It does the thing.
If you're trying to figure out which one your business actually needs, book a 30-minute strategy call. We'll look at one specific workflow you'd want automated and tell you honestly whether a chatbot will do it or whether you need a real agent.