What is the difference between a chatbot and an AI agent?
A chatbot answers: it reads a question and replies from approved information. An AI agent acts: it uses your systems to look things up, change records and complete a task. Agents do more, but they need integrations, permissions and testing, so they cost more to build and carry different risks.
The real difference is answers versus actions
The words are used loosely, and vendors stretch both. The useful distinction is what the system is allowed to do. A chatbot produces text: it reads a question and writes a reply. An agent can also use tools: it calls your booking system, CRM, accounting software or email, and it can change something in the world as a result.
| Type | What it does |
|---|---|
| Scripted bot | Follows a decision tree of buttons and fixed replies. Predictable, cheap and frustrating once a question falls outside the tree. |
| AI chatbot | Uses a language model to understand free text and answers from approved information, such as your FAQs, prices and policies. It reads; it does not change anything. |
| AI agent | Understands the request, then uses tools to look up records and take actions: book, reschedule, update a CRM, send an invoice reminder, file a document. It works towards finishing a task, not just replying. |
Many real systems sit between the two. An assistant that answers questions but can also look up a customer’s order status is reading from a live system, which is a small step towards agent territory. What matters is being explicit about which actions it can take.
Be wary of labels in sales material. Products described as “agents” are sometimes chatbots with a booking link, and products sold as chatbots sometimes take actions. Ask a vendor for the exact list of actions the system can perform, in which systems, and with what limits. That list tells you more than the name.
One request, handled three ways
Take a message a physiotherapy clinic receives every day: “Can I move my Thursday appointment to next week?”
- The scripted bot offers buttons for “Book”, “Cancel” and “Opening hours”. The customer gives up and phones.
- The AI chatbot understands the request and replies with the rescheduling policy and a link to the booking page, or says a member of staff will be in touch. Helpful, but a person still does the work.
- The AI agent finds the customer’s booking, checks next week’s availability for the same practitioner, offers two slots, moves the appointment when the customer picks one, and sends a confirmation. Staff see the change in the calendar.
The agent saves the most time, but only if it can reach the booking system through an API and follows the clinic’s rules on notice periods and practitioners. That is the trade-off in a sentence: more value, more build.
Which does your task need?
Tick what is true of your chatbot vs ai agent work.
Tick the statements that apply and the suggestion appears here: an AI agent, ordinary automation, or a person with software support.
When is a chatbot the better choice?
- Most of your enquiries are questions with stable answers: hours, prices you publish, eligibility, directions, how a process works.
- Your systems have no usable API, or you are not ready to give software write access to them.
- A wrong action would be costly and you would rather a person makes every change for now.
- You want to learn what customers actually ask before deciding what to automate.
A well-built AI chatbot working from a short, maintained set of approved answers is a reasonable first step. It also produces something useful beyond the answers: a log of real questions that tells you which actions are worth automating next.
The main ongoing cost of a chatbot is keeping its answers current. Someone on your team should own the approved information and update it whenever prices, hours or policies change, or the chatbot will repeat old answers with confidence.
When do you need an agent?
Choose an agent when the reply is not the end of the job. If every conversation ends with a member of staff copying details into another system, the value is in the action, not the answer.
Our tutor operations agent is an example. Tutors message it on WhatsApp the way they would message a coordinator, and it logs the session, matches the payment to the right family and files it. A chatbot could only have told them where to log it. For that business, the agent cut end-of-day admin by 78%, because it did the work rather than describing it.
Agents also fit work that spans several systems: reading an email, checking a record in your CRM, updating a spreadsheet and replying. See AI agent integrations for what that connection work involves.
How do cost and risk change?
An agent costs more to build mainly because of the integrations and the testing. Each system it touches needs an account, permissions, error handling and test cases for what happens when that system is slow or returns something unexpected. Running costs rise too: a single agent task often takes several model calls, one to understand the request, more to use tools and one to write the reply. The AI agent cost guide breaks this down.
The risk changes in kind, not just size. A chatbot’s main failure is a wrong answer, which is embarrassing and sometimes costly. An agent’s failure can be a wrong action: a cancelled booking, a duplicate refund, an email to the wrong person. Agents also face prompt injection, where text the agent reads tries to give it instructions.
That test is the basis of human-in-the-loop design. Our Sosho social media agent, for example, drafts and designs a week of posts on its own but publishes only after the shop owner replies with a nod in WhatsApp.
How should you decide?
- List the twenty most common requests your team handles, and note what happens after each reply.
- Mark which ones end with a reply and which end with someone updating a system.
- For the second group, check whether those systems can be reached by API, and with what permissions.
- Start with the chatbot for the first group and an agent for one or two high-volume actions from the second.
- Measure both before widening the scope.
And consider whether you need AI at all. If a request always follows the same rules, ordinary workflow automation is cheaper and more predictable than either.
Common questions
Is ChatGPT a chatbot or an agent?
Used on its own in a browser, it is mostly a chatbot: it reads and writes text. When connected to tools such as search, files or other software, it can act more like an agent. For a business, what matters is whether the system can reach your own records and take actions within rules you control, which a general assistant does not do out of the box.
Can we start with a chatbot and add agent abilities later?
Yes, and it is often sensible. If the chatbot is built on the same foundations, adding a tool such as “look up booking” or “reschedule” is an extension rather than a rebuild. Keep the approved answers and conversation logs from the chatbot phase; they show which actions customers ask for most.
Will an AI agent take actions without anyone checking?
Only the actions you allow. A well-designed agent has a list of permitted actions, limits on each one, and a set that always needs a person’s approval. Many agents start with every action approved by staff, then move routine, reversible ones to automatic once the logs show they are reliable.
Are AI agents only worth it for large companies?
No. Small businesses often benefit most because one person handles enquiries, admin and bookings. The key is a narrow first job with enough volume to matter, such as rescheduling, payment matching or document chasing, rather than a broad assistant that tries to do everything.
Bring us the task.
Tell us what happens today, which tools are involved and where it goes wrong. On a free 30-minute review, we’ll work out whether an agent, an automation or neither is the right fix, and what a sensible first release would be.
Start a project.

Book a call or send us a message. We read everything.
tash@neu.ie
Working with teams in Ireland and worldwide.
Free 30-minute AI review.