Plenty of tools sold today as AI agents are chatbots with a new name. And plenty of companies that needed a simple chatbot end up buying something far more complex than the problem called for.
The confusion is understandable: both hold conversations, both can use AI and both show up in the same place, usually a chat window. But there’s a simple test that tells them apart, and it helps you decide which one you need.
The test: who decides the next step
In a chatbot, the next step was decided in advance by a person. Someone designed the flow: if the customer says this, reply with that; if they choose option 2, ask for their account number; if it doesn’t understand, send it to support. The chatbot follows the script.
In an AI agent, the next step is decided during the task by the agent itself. It receives a goal, looks at the situation, chooses which tool to use, reads the result and decides what to do next. Nobody wrote the exact sequence beforehand.
A chatbot follows a path someone designed. An agent chooses the path within the limits you set.
Notice that the test isn’t whether it uses AI. A chatbot can use a language model to understand what the customer wrote or to phrase a more natural reply. If the script stays fixed, it’s still a chatbot.
How each one works under the hood
A traditional chatbot is a decision tree. Each user message leads down a branch. Newer ones use AI to recognize intent (“I need a copy of my invoice”) and to look up answers in a knowledge base, but what it can do after that is designed in advance.
An agent has three pieces that change the game:
- A goal, not a script. For example: resolve the customer’s exchange request following company policy.
- Tools, which are actions in real systems: looking up the order in the ERP, checking the return window, generating a shipping label, logging the interaction in the CRM.
- A decision loop. It acts, observes the result and decides again, until it reaches the goal or concludes it needs a person.
In practice, the difference shows when a case falls outside the norm. The chatbot reaches the end of the branch and transfers. The agent tries another lookup, cross-checks information from two systems and, if it still can’t resolve it, transfers with the case already investigated.
When a chatbot is the right choice
Chatbots aren’t outdated technology. For a lot of things, they’re the best option:
- Frequently asked questions with stable answers: opening hours, address, required documents.
- Simple triage: identifying the topic and routing it to the right team.
- Collecting data in a fixed format: name, ID number, order number.
- Regulated processes where every answer has to be exactly the approved one, with no variation.
- Very high volumes of short interactions, where cost per conversation matters a lot.
They’re predictable, cheap to run and easy to audit, because the whole path is designed upfront. If your problem fits on a one-page flowchart, a chatbot will probably do the job.
The thing to watch is growth. Chatbots that work get new branches every month, and at some point nobody knows why a question lands in the wrong place. If the tree now has dozens of exceptions and the team is constantly patching flows, that’s a sign the problem has changed in nature. It may have stopped being a script and become work that takes judgment.
When you need an agent
An agent makes sense when the work requires deciding the order of things and working inside systems:
- The case depends on looking things up in more than one system and cross-checking them.
- There are so many variations that designing every branch would be impractical.
- The work ends in an action, not just an answer: issue, schedule, update, reconcile.
- Today a person does the task by following criteria, not a fixed script.
- The task doesn’t always start in a conversation. It could be an incoming email, a new file in a folder, an order entering the system.
Picture, as an illustration, an accounting firm that receives client documents by email and WhatsApp. A chatbot can ask for the document and confirm receipt. An agent can identify the type of document, check that the details match the client record, enter it into the system and alert the person responsible when something doesn’t add up.
What changes in cost, risk and control
An agent’s autonomy comes at a price. Look at three things before choosing:
- Cost. An agent makes several model calls per case, because it decides in steps. Cost per interaction is higher. It pays off when it resolves something that used to take a person’s time.
- Predictability. The same request may follow slightly different paths. That calls for testing with many cases and monitoring once it’s in production.
- Control. Because the agent acts in systems, you need to define what it can do alone and what needs human approval. Permissions, limits and a log of every action stop being optional.
A chatbot that gets it wrong gives a bad answer. An agent that gets it wrong can take a bad action. That’s why designing the limits is central to any agent project.
Often, the answer is both
In most operations, the best solution combines the two. The chatbot handles reception, solves the simple things and collects the basics. When a case needs investigation or action in a system, it passes it to an agent. And when the agent reaches a limit, it passes it to a person.
To decide, ask three questions about the process:
- Can you reasonably design every path? If so, start with a chatbot.
- Does the work end with an action in a system? If so, you probably need an agent.
- How much does a mistake cost? The more expensive, the more limits and human approval the agent needs.
And be wary of labels. When you evaluate a tool, ask to see what happens with an unusual case. That’s where it becomes clear whether the next step is decided by the system or by a script written in advance.
Chatbots and agents solve different problems. Choosing well starts with knowing which one your process really calls for.