Somewhere in 2025 every chatbot on the market was renamed an AI agent. Some of them earned the new name. Most of them are the same decision tree with a language model painted on the front, and the difference only shows up when a customer asks for something that is not on the menu.
This is a plain account of what actually separates the two, which one a given job needs, and the questions that tell you which one a vendor is really selling.
The short answer
A chatbot answers. An AI agent completes. A chatbot's job ends when it has said the right thing. An agent's job ends when the thing is done — the slot is booked, the order status was looked up from the real system, the lead is qualified and sitting with the right salesperson.
Everything else follows from that. To complete a job, software needs three things a chatbot never needed: the ability to act on other systems, the judgement to decide which action fits the conversation in front of it, and a rule for when to stop and fetch a person.
Three generations, side by side
- Rule-based chatbot — the menu
- "Reply 1 for order status, 2 for returns." Every path was drawn by hand. Predictable, cheap, and useless the moment a customer types a sentence instead of a number. This is why people type "agent" as their first message.
- LLM chatbot — the articulate FAQ
- Understands free text and answers fluently from your help pages. A real improvement, but it can only say things. Ask it to move a delivery and it will explain, beautifully, how you could move a delivery yourself.
- AI agent — the junior colleague
- Understands the request, answers from your knowledge, and calls tools: check availability, book the slot, create the record, update the stage. It works towards an outcome across several turns, and hands over when the outcome is outside what it is allowed to do.
What an agent has that a chatbot does not
1. Functions — the ability to act
Functions (also called tools) are the actions an agent may take mid-conversation: look up an order, check a calendar, create a contact, take a payment. Without them, a bot can only describe. With them, it can finish.
The agent decides when to call a function from its description, which is why "check whether a delivery slot is available on a given date" works and "getSlots" does not. How to give a bot functions.
2. Grounded knowledge — the ability to be right
A chatbot can get away with a vague answer. An agent that acts on a wrong belief does damage, so it has to answer from your documents rather than from what the model half-remembers about companies like yours. A knowledge base an agent can actually use is arranged differently from a help centre, and it matters more than the model you pick.
3. Handover conditions — the ability to stop
This is the one most "agentic" marketing leaves out. An agent that can act needs written limits on when it must not: refunds above a threshold, a customer asking twice for a person, anything medical or legal, low confidence. The moment an agent should stop talking is designed in advance, not discovered from complaints.
A bot with functions and no handover conditions can take actions it cannot explain. That is worse than a chatbot, not better.
4. Memory of the person, not just the chat
A chatbot sees one conversation. An agent that is going to act on a customer's behalf needs the contact record: the lead stage, the last call and how it ended, the open order. Otherwise it will cheerfully sell a plan to someone who cancelled last week.
Where a plain chatbot is still the right answer
Agents are not automatically better. They cost more per conversation, take longer to set up and have more ways to fail. A chatbot is enough when:
- The questions are informational and the answers rarely change — opening hours, fees, eligibility.
- Nothing needs writing back anywhere. The conversation ends with a link or a fact.
- Volume is low enough that a person can pick up everything else within the hour.
Start there if that describes you. Adding one function later — usually booking — is a smaller step than it sounds.
Where an agent pays for itself
- Booking. Appointments, site visits, demos, callbacks. The conversation has a clear finish line and a calendar that can say yes or no. What has to be true first.
- Lead qualification. Asking the four questions in an order that survives a real conversation, then moving the lead and assigning it. The questions, and the script.
- Status and tracking. "Where is my order" is one of the biggest categories in almost any e-commerce support queue, and the answer lives in a system the agent can read.
- After-hours cover. The calls and chats you miss between 7pm and 9am have a floor of zero. What to answer, and what not to.
"Agentic" washing: six questions for a vendor
If a product is called an AI agent, these separate the real thing from a renamed menu:
- What can it change, not just read? Ask for the list of actions. "It integrates with your CRM" is not an answer.
- What does it do when a function fails? Ask them to break one in the demo. The reply is what your customers will hear on a bad day.
- How does it hand over, and what does the person see? If your teammate has to scroll up and re-read, the handover is a restart.
- Where do its answers come from? Your documents, indexed — or the model's general knowledge with your name in the prompt? Five controls against invented answers.
- Is it the same agent on every channel? One set of rules for WhatsApp, Instagram, the website and the phone, or five bots that drift apart?
- What happens to the data? Nine questions to get answered in writing.
How Convarza builds them
In Convarza an AI bot is a prompt, an opening line, a knowledge base, a set of functions, a model and handover conditions — the four ingredients above, configured in one place. The same bot can answer WhatsApp, Instagram, Facebook, your web widget and email at once, or be built as a call bot that answers the phone with a voice per language.
Handover is just assignment: chats the bot is handling are the unassigned ones with an AI badge, and assigning one to a teammate takes the bot out and puts a person in, with the contact record beside the thread. When one bot is not enough, a bot flow chains a greeter that qualifies to a specialist that books.
See the AI chatbot and AI voice agents, or how a support team uses both to cut the cost per conversation.