Why the words get confusing
“Chatbot” and “agent” are used loosely, and vendors often choose whichever sounds more impressive. There is no official definition that every company follows. Still, a practical distinction holds up well enough to be useful: a chatbot mainly produces replies, while an agent is set up to carry out steps that change something in the world, such as sending a message, updating a record or booking a slot.
Understanding that difference matters less for vocabulary and more for risk. A tool that only talks can mislead you. A tool that acts can also do things you did not intend.
What a chatbot does
A chatbot takes a message and returns a message. Its job is conversation: answering a question, explaining a policy, collecting details from a customer, or walking someone through a set of options.
Chatbots come in two broad styles:
- Scripted bots. These follow menus and fixed branches. If a customer picks “track my order,” the bot shows the tracking step. They are predictable and limited, and they belong to the family of rules-based automation.
- Language-model bots. These generate replies on the fly using language models. They handle free-form questions much more flexibly, but their answers can vary and can be wrong.
Either way, the output is words. If a human reads the reply and decides what to do next, the human is still the one acting.
What an agent does
An agent is software that is given a goal and some ability to act toward it. Typically that involves a loop: look at the situation, decide on a step, use a tool or system to carry it out, check the result, and continue. The “tools” might be a calendar, an email account, a spreadsheet, a ticketing system or a web browser.
A simple illustration: you ask a chatbot “What is our refund policy?” and it replies with text. You ask an agent “Refund this customer and tell them it’s done,” and it looks up the order, issues the refund in the payment system, and sends the confirmation. The second request has consequences outside the conversation.
Many agents use a language model to interpret instructions and plan steps, but the model is only one part. The rest is the connection to other systems and the permissions that go with it.
A spectrum, not two boxes
In practice, tools sit along a spectrum:
- Answers only. The tool replies and nothing else.
- Suggests actions. It drafts an email or proposes a reply, and a person clicks send.
- Acts after approval. It prepares the action and waits for a person to confirm.
- Acts on its own within limits. It carries out certain actions automatically, such as filing routine requests.
- Acts across many steps. It chains several actions together toward a goal with little supervision.
Most reliable workplace uses sit around the middle. They save effort on preparation while keeping a person in charge of anything that is hard to undo.
Why the distinction matters for risk
When a chatbot is wrong, the damage depends on whether anyone relies on the answer. When an agent is wrong, the action may already have happened. A few points follow from that:
- Permissions. An agent should have access only to what its task needs. Broad access turns a small mistake into a large one.
- Reversibility. Sending an internal draft is easy to correct. Sending money, deleting records or emailing a customer list is not.
- Misreading instructions. Language is ambiguous, and an agent may interpret a request differently than you meant. The same weakness that produces fabricated answers in a chat can produce a wrong action in an agent.
- Hidden steps. When several actions run in sequence, it can be hard to see where something went off track unless the tool keeps a clear log.
These are reasons to build in human review at the points that matter, and to plan an escalation path for when the tool gets something wrong.
Where each fits
Chatbots tend to suit tasks where the value is in answering and guiding: common customer questions, internal questions about policies, or basic intake. Used for customer experience, a well-scoped bot can handle routine requests and hand harder ones to a person.
Agents tend to suit tasks that are routine, well defined and made of repeatable steps across several systems, where mistakes are inexpensive to detect and fix. That overlaps with older tools like robotic process automation, which also performs actions on a user’s behalf but follows fixed instructions rather than interpreting open-ended goals.
For anything involving money, legal commitments, health information or hiring decisions, the case for fully independent action is weak. Outcomes are uncertain, and a person should usually remain responsible.
Questions to ask when you hear “agent”
When a product is described as an agent, a few plain questions cut through the label:
- What can it actually do besides reply? Which systems does it connect to?
- Does it act immediately, or does it ask for approval first?
- What happens if it misunderstands? Is there a log, and can actions be undone?
- Who is responsible for what it does?
These fit naturally into the broader list of vendor questions you would ask about any AI product. If the answers are vague, the tool is probably closer to a chatbot with a good name than a dependable worker.
Keeping it simple
You do not need to memorize a taxonomy. Ask whether the tool only talks or whether it also does things. If it only talks, check its answers. If it does things, limit what it can touch, require approval for anything consequential, and keep records of what it did. Capabilities and reliability vary widely between products, so test any tool on low-stakes work before trusting it with more.