Chatbot or AI agent?
Choose what fits.
A chatbot converses with customers via fixed routes or AI. An AI agent can also independently choose next steps within agreed boundaries and use tools to complete a task.
Knowledge hub · Sources checked on November 4, 2026
The name doesn’t tell the whole story.
Look at the room a system is given: only providing answers, following a fixed route or choosing the next step itself.
| Aspect | Fixed chatbot route | AI chatbot | AI agent |
|---|---|---|---|
| Conversation | Choices and predefined routes; sometimes intent recognition. | Natural language and generated responses within the chosen setup. | Natural language, with room to determine next steps. |
| Own knowledge | Answers you define yourself or look up via an integration. | Can use relevant information from a managed knowledge source. | Can consult knowledge while handling a task. |
| Actions in systems | Possible through preconfigured rules and workflows. | Possible; language skills alone say nothing about the available actions. | Can select permitted tools; access rights remain decisive. |
| Control | Test the paths and maintain decision rules. | Check sources, answer quality and missing knowledge. | Also check tool use, exceptions and consequences. |
| Human handover | Must be explicitly set up. | Must be explicitly set up. | Must be explicitly set up; autonomy does not replace that agreement. |
More about chatbots, workflows and AI agents
In practice, the terms overlap. A chatbot is the conversational interface; behind it, fixed rules, a language model and system integrations can work together. An AI agent can therefore be visible as a chatbot. For your choice, what matters most is what the application actually does and which decisions you leave to the system.
A useful technical distinction is that a workflow follows predefined steps, whereas with an agent, the model has more influence over the sequence and use of tools. Not every process needs that freedom. Therefore, ask a provider to demonstrate the same customer query, an exception, and an error scenario. IBM’s explanation of chatbots and Anthropic’s distinction between workflows and agents provide background on these terms.
When should you choose which approach?
Choose the simplest approach that can reliably solve the customer’s problem. More autonomy is not an end in itself.
A fixed chatbot may be enough.
Think of changing an address with fixed input fields, recognisable conditions and one follow-up step. A defined route is easy to test. Do check what happens when someone asks an unexpected question or information is missing.
AI can help with the answer.
Customers ask the same substantive question in different ways. An AI chatbot can help find relevant knowledge and formulate an answer. For this to work, it must be clear which sources take precedence and when the system should not provide an answer.
An agent may be suitable.
Some requests require retrieving information, choosing a follow-up step and checking the result. In that case, investigate whether an AI agent adds value. Limit its scope to the actions and situations you have assessed in advance.
Provide an answer or carry out a task?
Distinguish between answering questions and carrying out tasks. “How do I change my order?” calls for an explanation. “Change my order to tomorrow” may require an action with consequences. Both can use the same knowledge, but they place different demands on identification, authorisation and integrations.
Keep the employee in mind too. If colleagues mainly spend a long time looking for answers, Agent Assist may be a more logical first step than a new customer-facing bot. The employee then continues to conduct the conversation and assesses the suggested information.
Work through one customer question from start to finish.
Use a concrete contact reason in your demo. For example: a customer wants to change the delivery date of an order.
- 1.
Understand what the customer means
First ask which order and what desired change are involved. “Tomorrow doesn’t work” could mean that someone wants a different date, but it could also mean they want to cancel the order. A useful application asks follow-up questions when that distinction is still unclear.
- 2.
Check knowledge and up-to-date data
A general article can explain when changes are possible. Whether this order can still be modified is shown in the connected order system. Demonstrate how the application makes that distinction. A convincing sentence without a verified order status is insufficient.
- 3.
Only carry out an authorised change
Record which details are required and when the customer or an employee must confirm. Check after the system action whether the change succeeded. An error message must not turn into a success message. Also discuss how you prevent trying again from performing the same action twice.
- 4.
Clearly report what happened
Tell the customer whether the date has been changed, still needs to be confirmed or is unavailable. If an employee needs to take over, they should receive the relevant question, checks performed and outcome. The customer should not have to repeat the entire conversation.
What goes wrong without good agreements?
The quality of the conversation also depends on what you organise behind it.
An old answer remains an old answer.
AI does not update an outdated returns policy. Assign an owner, determine which version applies and retest changed terms. Start with the topics the application will actually handle.
A stalled conversation needs a way forward.
Determine who takes over when information is missing, an exception occurs or an integration fails. Make clear what happens outside opening hours and what information remains available to the employee.
No employee needed does not always mean resolved.
A customer can drop out without being helped. Therefore, look beyond automation at repeat contacts, the outcome of the task and customer feedback. Review a selection of conversations in detail, including the less successful examples.
How to set up a manageable pilot
Start a pilot with a limited, recognizable set of questions. Include simple questions, incomplete requests and exceptions. Define in advance what constitutes a good answer or a successful task. Agree when the pilot provides sufficient confidence to expand, and who decides to temporarily roll back an application.
Also make management concrete: who updates knowledge, who checks conversation quality and who investigates failed actions? Use the checklist for an AI-ready knowledge base to assess the content foundation before adding more contact reasons.
How does Archie fit into this?
Archie is the AI platform for customer service from VANAD Digital, part of VANAD Group.
Archie Talks
For customer-facing conversations and answers. View the AI agent for customer service and discuss which contact reasons you want to support.
Archie Acts
For actions and workflows in connected systems. On the Automation page, you can read how a conversation can lead to a concrete follow-up action.
Archie Knows
For managing the knowledge that customers and employees rely on. View Archie Knowledge for the foundation behind those answers.
From a damaged product to a next step.
The public Xenos case shows how a conversation and a configured handling process can work together.
Chatbot Xenna asks about the order and the damaged product in the described damage workflow. A suitable resolution, such as a replacement or refund, can then be initiated. This is a concrete example of conversation plus process.
The case does not provide evidence that an AI model independently chooses all follow-up steps. Therefore, use it as an example of a defined customer journey, not as a general promise about autonomous agents or results for your organization.
Your questions.
Clearly answered.
From a fixed route to more independent handling: this is what you need to clarify in advance.
Is a chatbot always based on fixed scripts?
Can an ordinary chatbot also perform an action?
Does an AI agent have access to all our systems?
When should an employee take over?
Do we need to replace our existing chatbot?
Sources and updates · checked on 4 November 2026
The definitions follow the public sources below. The decision questions and illustrative examples are editorial applications to customer service. Product capabilities depend on configuration and integrations.
- IBM · What is a chatbot?Chatbots, fixed routes and AI-supported conversations.
- Anthropic · Building effective agentsDifference between prescribed workflows and agents; published in 2024. Used for the architectural distinction.
- Microsoft Learn · Retrieval-augmented generationConnecting answers to retrieved knowledge and access management.
- Archie · AutomationPublic product explanation of connected follow-up steps.
- Archie · Xenos-casePublic description of the claims flow; no general guarantee of results.
