Chatbot vs AI agent: what your business really needs
A chatbot answers your customer's question. An AI agent completes the task. That difference decides whether you end up with less admin work or just one more window on your website.
This guide walks through the three levels of these systems and how each behaves on one real conversation. It ends with four questions to ask yourself before you pay for anything.
In this guide
What is the real difference between a chatbot and an AI agent?
The difference is what happens after the system understands what the customer wants. A chatbot writes an answer and stops there. An AI agent then goes into your systems, checks the current state, performs the action and confirms it back to the customer.
Picture two people on a front desk. The first has memorised your price list and opening hours and recites them politely, and for anything else tells you to call back tomorrow. The second has the calendar open, finds a free slot while you wait, books you in and sends the confirmation before you leave the building. Both are helpful. Only one solved your problem.
Three levels, not two
In practice the market does not split neatly into two camps. There are three levels, and most of the confusion comes from all three being sold under the same name.
- Scripted chatbot: follows pre-built conversation branches. You write every question and every answer. Step outside the script and it breaks.
- Grounded assistant: reads your actual content, meaning your website, documents and price lists, and answers from that in its own words. Handles typos and messy phrasing.
- AI agent: all of the above, plus access to your systems. Checks availability, moves appointments, updates the CRM and runs multi-step processes.
One note on this split. The levels describe what the system does for you, not how it is built. On modern platforms, ours included, the same technology powers all three. Moving from answering to acting is therefore an extension of what you already have, rather than a replacement.
The three levels compared
Same distinction, seen through what an owner actually cares about. How much work setup takes, what happens when your content changes, and how much is left for your team.
| What you care about | Scripted chatbot | Grounded assistant | AI agent |
|---|---|---|---|
| Understanding questions | Matches preset keywords | Understands intent, typos and context | Understands intent and identifies the requested action |
| Source of answers | Text you typed in manually | Your website, documents, knowledge base | Content plus live system state |
| Setup | Mapping every conversation branch by hand | Connecting content you already have | Content plus integrations and defined rules |
| When your content changes | You rewrite the rules | Answers update with the content | Content and action logic both update |
| End of the conversation | Information | A correct answer from a verified source | A completed action and a confirmation |
| Work left for your team | Roughly unchanged | Fewer repeat questions | Fewer questions and less manual data entry |
A scripted chatbot is a tool for a different job. If you field ten unchanging questions and have no systems to connect, a script will do the work.
The trap is that all three levels get sold under the same label. Someone will bill you for an "AI agent" that turns out to be a script with a nicer interface. So do not ask a vendor what the system knows. Ask what it completes without a human touching it.
What this looks like in practice
It is easiest to see in a single conversation. Take a patient who messages a clinic at nine in the evening because they need to move Friday's appointment.
A scripted chatbot spots the word "appointment" and returns its prepared line: our hours are 9 to 17, please call reception. The patient calls in the morning, reception picks up, and the change gets entered by hand. Nothing was saved; the same work was just postponed.
A grounded assistant explains your rescheduling policy, because it read it in your document: changes are possible up to 24 hours before, free of charge, by contacting the clinic. Better, because the patient got an accurate answer at nine at night. The appointment is still not moved.
An AI agent recognises this as a reschedule request, opens the calendar, checks the rules you set, finds a free slot on Tuesday at ten, moves the patient and emails the confirmation. When reception starts work in the morning, there is no new request waiting to be handled.
Where a good agent deliberately stops
This part is designed in. If a patient asks to reschedule five hours before the appointment and your rule says 24 hours, the agent does not improvise. It states the rule, offers the contact route and hands the conversation to your team with the full context attached.
A proper handoff is more than a "please contact us" message. Your team gets the whole conversation, the identified intent and whatever details the customer already gave, so nobody asks the same thing twice.
The same applies to anything that is not pure admin. Questions about test results, treatment, discounts outside the price list or anything needing judgement go to a person. If a system does not draw that line, treat it as careless.
What an AI agent cannot and should not do
An AI agent does not replace a department and cannot handle work that was never written down. These are the limits worth knowing before you talk to any vendor.
- It does not know what is written nowhere. If your rescheduling policy exists only in your receptionist's head, the agent will not guess it. Rules get written down first, the system comes second.
- It should not make things up. A properly configured system is restricted to your verified content and says "I don't know that" instead of guessing.
- It does not make business decisions. Approving discounts, resolving complaints and anything carrying risk stays with people. The agent can prepare the information behind that decision.
- It does not clean up messy data. In our experience, prices cause the most trouble. With several clients, the pricing on the website was out of date and inconsistent between pages, so the assistant gave wrong answers until we reconciled the sources.
Worth knowing too: an agent takes more steps per conversation than a plain chatbot, because it checks and plans before doing anything. That makes it more expensive per conversation. It pays off wherever one completed action is worth more than a few minutes of someone's time. In healthcare, tourism and online shops, that is nearly always.
How to decide which one you need
Answer four questions and the choice becomes obvious. This is a conversation you have with yourself, before you have it with a vendor.
- How many daily enquiries are pure information? Hours, prices, location, terms. If that is the large majority, a grounded assistant solves your problem and you need nothing more.
- How many enquiries end in someone typing into a system? Moving an appointment, taking a booking, entering a new enquiry in the CRM. Every one of those is work an agent can absorb.
- Are your rules written down in one place? If not, your first step is a single short document setting out the rules. That is an afternoon's work and it pays off regardless of AI.
- What is allowed to happen without human approval? Draw that line before you deploy, not after the first awkward call with a customer.
The first two questions tell you which level you are after. If the answer to the second is "a lot," you want an agent. If it is "almost none," do not pay for integrations you will never use. The third and fourth tell you whether you are ready. Without answers to those, deployment stalls no matter who you pick.
Most of the work goes into connecting the calendar, the CRM and your rules. The system has to err on the safe side and hand off cleanly when a case gets complicated. We do that part for our clients. We map the process, put the assistant on your existing content, add the actions and integrations, and run the system afterwards. You never touch the technical side. You can see how it looks by sector on our pages for AI in healthcare and AI in tourism.
Frequently asked questions
Short answers to the questions we get asked most.
How much does an AI agent cost?
The price depends on what the system has to do, because this is not a fixed package. An assistant that only answers from your content is the cheapest option. Every integration and every action the system takes on raises both the build cost and the running cost. We agree a starting price before we build anything.
Do we need clean data and a database to use an AI agent?
You need neither a database nor spreadsheets. You need accurate, current content, meaning a website, price list or documents, and rules written down in a form that can be checked. There is more on this in the guide on preparing your content for an AI assistant.
Can an AI agent give a customer wrong information?
It can, if it is not restricted to your verified content, or if that content is itself wrong. A properly configured system answers only from approved sources and admits when it does not know something. That approach is covered in the guide on stopping AI from making things up.
Is our data safe when an agent accesses internal systems?
Your content and your customers' questions are processed by an external AI model provider, in EU regions. Your content is not used to train public models and is not shared with other customers, and access to each connected system is restricted separately. This is a question worth putting directly to any vendor, because the answers differ.
How long does deployment take?
An assistant that answers from your content can be working within minutes, because it reads your existing site. An agent that performs actions takes longer, since processes get mapped, rules get defined and integrations get built first.
Choosing between a chatbot and an AI agent comes down to where your business is losing time. If it goes on answering the same questions, an assistant that reads your content well is enough. If it goes on retyping appointments, bookings and contact details from one window into another, you want a system that takes that step itself. The expensive option is the third one, paying for an agent when nobody has written down the rules it is supposed to follow.