Some time ago I visited the website of a garage that specialises in replacing car windscreens. In one corner there was a window with a very direct message: “Ask us”.
I needed information about the car’s front glass. The problem appeared when I tried to explain it in different words. If I didn’t use exactly the term the system expected, the chatbot didn’t understand the question.
I could write “windscreen”, “windshield” or “front glass”, but the conversation never flowed naturally.
I can’t know how that system was built, but the experience as a user was obvious: it didn’t seem to understand the meaning of the question, only to look for certain words.
The feeling of talking to a machine appeared immediately. After several attempts, I lost patience and stopped using the chat.
This experience sums up one of the main differences between a traditional chatbot and an AI agent: one usually follows predefined paths; the other can interpret context, look up information and help the user move forward.
But that doesn’t mean every company needs an artificial intelligence agent.
What a traditional chatbot is and how it works
A traditional chatbot is a system designed to answer questions or guide a conversation through predefined rules.
It can work with:
- Buttons and menus.
- Keywords.
- Scripted answers.
- Decision trees.
- Closed questions.
- “Yes” or “no” checks.
It works much like a form presented as a conversation. The system asks a question, checks the answer and shows the next planned step.
This can be enough when the process is simple and has few variations.
The problem appears when the user expresses themselves in a way that wasn’t foreseen. If the chatbot expects “front windscreen” and the person types “windshield” or “front glass”, it may not find a match and return a generic answer.
That is why many conversations with chatbots end in messages such as:
- “I didn’t understand your question.”
- “Please select one of these options.”
- “Try asking it another way.”
- “These are the topics I can help you with.”
The user then has to adapt to the machine’s logic.
With a traditional chatbot, it is often not the system that understands the user: it is the user who has to figure out how to talk to the system.
What an AI agent is and what it is for
An AI agent can interpret natural language, use the available context and decide what information or tool it needs to answer.
The difference isn’t only that it writes in a more natural way.
To be genuinely useful, an agent must be able to work with three elements:
1. Context
This is the information it has about the company:
- Services.
- Products.
- Terms.
- Processes.
- Policies.
- Content.
- Frequently asked questions.
2. Tools
These are the sources or functions it can consult:
- Product catalogue.
- Calendar and availability.
- Database.
- Content search.
- Booking system.
- Authorised internal documents.
- Forms or contact systems.
3. Rules and permissions
These are the limits that determine:
- What information it can look up.
- What data it can show.
- What actions it can start.
- When it needs confirmation.
- When it must stop.
- When it has to hand the conversation over to a person.
An agent shouldn’t just find a predefined answer. It should understand what the person is trying to achieve and help them find the next step.
If someone writes “I need to change the front glass of my car”, the system should understand that they are probably asking about the windscreen.
From there, it could ask for the necessary details or show the available information.
Understanding intent saves the user from having to guess the exact words the machine expects.
Chatbot vs AI agent: main differences
| Aspect | Traditional chatbot | AI agent |
|---|---|---|
| Understanding | Looks for expected options or words | Interprets natural language and context |
| Conversation | Follows a defined path | Can adapt the path to the need |
| Information | Uses scripted answers | Can consult authorised sources |
| Actions | Usually moves from question to question | Can use tools and start permitted actions |
| Flexibility | Works best in closed cases | Handles questions phrased in different ways |
| Up-to-date information | Its answers have to be edited | Can consult up-to-date systems if connected |
| Main risk | Not understanding the user | Making things up or acting wrongly if badly configured |
| Human handover | Usually appears as a last option | Must kick in when data, permissions or certainty are missing |
Not every system fits neatly into one of these columns.
There are very well-designed traditional chatbots, and AI assistants that generate more natural answers but can’t look up information or take any action.
That is why it isn’t wise to decide based on the commercial label.
Before buying a solution, you need to check:
- What the system really understands.
- What information it can look up.
- What tools it can use.
- What actions it is allowed to start.
- What happens when it doesn’t know the answer.
When a traditional chatbot is enough
Imagine a trainer who offers free assessment sessions for a strength programme.
Before accepting a request, they need to check four or five conditions. They might ask, for example:
- Whether the person has previous experience.
- What goal they are pursuing.
- Whether they have any particular limitation.
- What times they are available.
- Whether they meet the programme’s basic requirements.
The path is closed:
- The chatbot asks a question.
- The user selects or types an answer.
- The system checks whether the answer fits.
- If it is valid, it shows the next question.
- At the end, it shows the result or the corresponding form.
In this case, a traditional chatbot can be enough.
It doesn’t need to interpret dozens of intents, consult several sources or hold an open conversation.
A chatbot can also handle these tasks well
- Answering five or six frequently asked questions.
- Showing fixed opening hours.
- Directing the user to a specific section.
- Running a simple survey.
- Collecting data through a closed path.
- Classifying requests with preset answers.
- Checking basic conditions with “yes” or “no” questions.
Adding artificial intelligence to these processes could increase cost and complexity without really improving the experience.
The most advanced solution isn’t always the most suitable. If the path is closed and predictable, a traditional chatbot can do the job perfectly well.
When an AI agent can add more value
An agent starts to make sense when questions vary, the user can express themselves in many ways or the answer depends on information that has to be looked up.
AI agent in an online store
A person can describe what they are looking for without knowing the exact name of the product:
- “I need a backpack I can take on the plane.”
- “I’m looking for trail running shoes.”
- “I want a gift for someone who is starting to paint.”
- “I need a jacket for a trip somewhere very cold.”
If the agent is connected to the catalogue, it can interpret the need, look for products that meet the conditions and check the available information.
But that ability doesn’t appear automatically.
To give information about stock, prices or real features, the agent must consult an up-to-date, authorised source.
AI agent in a service company
Someone visiting the website of a service company usually wants to know whether what they offer fits:
- Their problem.
- Their budget.
- Their availability.
- Their location.
- The time frame in which they need a solution.
An agent could:
- Explain the services available.
- Clarify what each option includes.
- Check availability if it has access to the calendar.
- Help start a booking.
- Collect the necessary details with the user’s consent.
- Send a guide or an information pack.
- Hand the question over to a person when needed.
The aim shouldn’t be to pressure the user into buying.
An agent doesn’t have to replace the salesperson. It can guide the user, answer the first questions and lead them to the start of the sales process.
Negotiation, the final quote and sensitive decisions can stay in the hands of a person.
Answering a question is not the same as meeting a need
People don’t visit a website because they want to talk to a chatbot. They visit because they need to get something done.
They might need to:
- Find a product.
- Find out whether a service fits their budget.
- Check a date.
- Request an appointment.
- Download information.
- Clear up a question before buying.
- Contact the right person.
A chatbot could answer:
“We are open Monday to Friday.”
An agent connected to the calendar could meet the whole need:
“We are open Monday to Friday. If you like, I can check the available times for a call.”
The first answer provides information.
The second understands what the user might need next and lets them move forward.
Answering a question is providing information. Meeting a need is helping the user find the next step without hiding the system’s limits.
The biggest risk of an AI agent: making up an answer
An AI agent may understand language better, but it also introduces a risk that shouldn’t be played down: it can generate a convincing answer that isn’t true.
This problem is especially serious when it speaks on behalf of a company.
The Air Canada chatbot case
In 2024, the Civil Resolution Tribunal of British Columbia examined the case of Moffatt v. Air Canada.
The airline’s chatbot had given incorrect information about the possibility of applying retroactively for a special bereavement fare.
The user relied on that information, bought the ticket and later discovered that the company’s actual policy said otherwise.
The tribunal concluded that Air Canada had not taken reasonable care to ensure that the information provided by the chatbot was accurate.
It also noted that the chatbot was part of the company’s website and that the company was responsible for the information published through that system.
See the reference to the Moffatt v. Air Canada case on CanLII
A company can’t separate its responsibility from the answers given by the agent it has placed in front of its customers.
What an agent without proper limits could get wrong
A badly configured agent could:
- Make up an offer that doesn’t exist.
- Confirm an incorrect commercial condition.
- Promise a service the company doesn’t provide.
- Show an out-of-date price.
- Accept a request a person should review.
- Look up information the user isn’t authorised to see.
- Present an assumption as if it were a confirmed fact.
The risk increases when the system tries to please the user.
If a person rephrases the same request several times, the agent shouldn’t end up giving in and confirming something it has no information about.
What an agent should do when it doesn’t know the answer
A good agent isn’t one that always answers.
It is one that can tell when it has enough information and when it must stop.
When it doesn’t know an answer, it should:
- Clearly acknowledge that it doesn’t have that information.
- Avoid filling in the gaps with assumptions.
- Take note of the question if the user wants to continue.
- Explain what will happen next.
- Hand the conversation over to the right person.
- Log the topic so the company can improve its information.
An answer like this can be perfectly valid:
“I don’t have enough information to confirm that. If you like, I can take note of your question so that someone from the team can get in touch with you.”
This answer builds more trust than a detailed but invented reply.
To achieve this, the agent needs:
- Defined information sources.
- Up-to-date documents.
- Limited permissions.
- Clear rules.
- Monitoring systems.
- A human handover procedure.
An AI agent also needs access limits
Being able to look up information doesn’t mean the agent should have access to all of the company’s data.
Some information can be shown publicly:
- Services.
- Products.
- General terms.
- Opening hours.
- Public availability.
- Articles and content.
- Frequently asked questions.
Other information requires first checking who the user is:
- Personal bookings.
- Contact details.
- Order history.
- Invoices.
- Contract information.
- Medical or financial data.
- Private documents.
In addition, certain actions should require explicit confirmation:
- Making a booking.
- Cancelling an appointment.
- Changing personal details.
- Sending information to third parties.
- Accepting terms.
- Confirming a purchase.
- Making a payment.
The more tools an agent can use, the more important permissions, validations and monitoring become.
How to choose between a chatbot and an AI agent
Before buying a tool, it is worth analysing the process it has to solve.
A chatbot may be enough if:
- There are few, predictable questions.
- The path barely varies.
- The answers can be defined in advance.
- It doesn’t need to look up changing information.
- It doesn’t have to carry out actions in other systems.
- The user can move forward with buttons or closed answers.
An AI agent may make sense if:
- Users express the same need in many ways.
- The conversation needs to keep context.
- There are several services, products or paths.
- The answer depends on up-to-date information.
- The system needs to consult authorised tools.
- It must know when to answer and when to hand the question over.
- The company can monitor its answers and correct its sources.
There is also a preliminary question many companies forget:
Is the information needed organised, up to date and ready for the agent to consult?
If the company isn’t clear about its services, prices, terms or policies, adding artificial intelligence won’t fix that lack of definition.
Checklist before installing an AI agent
Before putting an agent in front of users, the company should be able to answer these questions:
- What specific need should it meet?
- What information can it use?
- Who keeps that information up to date?
- What tools can it consult?
- What data requires identifying the user?
- What actions can it start?
- What actions need confirmation?
- What should it do when it doesn’t know an answer?
- How is the conversation handed over to a person?
- Who will review conversations and errors?
- How will we measure whether it is adding value?
- What will happen if an information source becomes unavailable?
If these questions have no answer, the project probably isn’t ready to be put in front of real customers yet.
A useful agent shouldn’t try to pass itself off as a person
The conversation can be fluent, but the user should know they are interacting with an artificial intelligence system.
Building trust isn’t about hiding that there is a machine behind it.
It is about:
- Understanding the question.
- Answering with verifiable information.
- Explaining the limits.
- Asking for permission before using data.
- Not taking actions beyond its scope.
- Making it easy to contact a person.
At FEIDERS we apply this approach when designing agents: first we define what need they should meet, what information they can use and what actions they are authorised to start.
The agent is one part of the journey, not an automatic substitute for all the people who work at the company.
The question isn’t which one looks more advanced
A traditional chatbot can handle a simple process perfectly well.
An AI agent can add more value when it needs to understand context, look up information and use tools.
Choosing the most complex option doesn’t guarantee a better experience.
The right question is:
What does the user need to achieve, and what is the simplest, safest and most useful way to help them move forward?
If the process consists of asking five closed questions, a chatbot may be enough.
If the user needs to explain their situation in their own words, compare options, look up up-to-date information and find the next step, an AI agent can offer a more complete experience.
In both cases, the technology only works when the company properly defines the information, the limits and the responsibility behind every answer.