The one-sentence difference
A chatbot answers. An agent acts. A chatbot tells your customer what your refund policy is. An agent reads the order, applies the policy, issues the refund in your payment system, and writes the note in your CRM, or hands the case to a person when it is unsure.
Both are usually built on the same large language models. The difference is everything wired around the model: an agent can use tools, which means it can search, look things up in your systems and take actions in them, with permissions and logs.
What each one is good for
A chatbot is enough when the job is answering questions. Website help, first-line support from your help articles, guiding people to the right page or form. If the visitor still does the doing, a chatbot serves them well and costs less to build.
An agent earns its keep when the job is finishing a task. Resolving the ticket, not describing the solution. Processing the invoice, not explaining how invoices are processed. Booking the appointment, chasing the document, updating the record. If a person currently reads something, decides and then clicks around in software, that is agent territory.
A quick test for your own use case
Write down what happens after the conversation ends. If the answer is "the customer goes and does it", you need a chatbot. If the answer is "one of our staff goes and does it", you probably want an agent, because the conversation was never the expensive part; the doing was.
Two more questions sharpen it:
- Is there a clear right outcome? Agents need one. "Resolve this ticket correctly" is checkable; "have a nice conversation" is not.
- Is the volume worth it? An agent that saves two hours a week will not repay its build. One that handles hundreds of cases a month usually will. The arithmetic is in what an AI agent costs to run.
What they cost, relative to each other
A chatbot over your help content is the smaller project: the work is mostly retrieval quality and tone, and integrations are light. An agent adds tool connections to your real systems, permissions, confidence thresholds, a hand-off path and monitoring, which typically makes it a mid-size software project. Running costs differ less than people expect; agents use more model calls per case but handle cases that were costing staff time.
The upgrade path most companies actually take
Start with a chatbot that answers well and measures itself: which questions it handled, where it gave up, what people asked next. That data tells you exactly which task is worth automating end to end, and the chatbot's retrieval layer becomes the agent's knowledge. We build both this way; the chatbot is stage one of the agent, not a throwaway.
Read more about how we build agents on our AI agent development page, or the assistant side on LLM and generative AI apps.