Insights · AI Agents
AI Agent vs Chatbot: What Is the Difference?
The words are used interchangeably, but a chatbot and an AI agent are different tools with different risks. Choosing the wrong one wastes money or, worse, gives software more access than it should have.
A chatbot talks
A traditional chatbot follows a script or retrieves a canned answer. Many modern ones use a language model to phrase replies more naturally, but the job is the same: respond to a question with text. It does not usually know who you are, cannot look anything up in your accounts, and cannot do anything on your behalf.
That is exactly right for some jobs. An FAQ on opening hours, directions or product information does not need more. A chatbot is cheaper to build, easier to reason about and has little that can go wrong.
An agent can use tools
An AI agent is software built around a language model that can decide which of several defined tools to call to complete a task. A tool might look up an order, check a calendar, search a policy document or draft an email. The model proposes a step; the surrounding software checks it is allowed, runs it, and gives the result back so the model can continue.
The important word is defined. A well-built agent does not get open access to a database or the internet. It gets a short list of narrow capabilities, each with its own permission. “Find open invoices for this customer” is a tool. “Run any SQL” is not.
What changes in practice
Take a customer asking where their order is. A chatbot can say “please contact support”. An agent can recognise the customer, run an approved order lookup and answer with the real status and date, or say it could not find the order and hand over to a person with the details already collected.
The same pattern applies inside a business. A manager can ask “which deals have gone quiet?” and get an answer from the CRM rather than waiting for a report.
The new risks, and how they are handled
Capability brings risk, so the engineering around the model matters more than the model itself.
- Permissions: the agent acts with the rights of the person asking, never more.
- Approvals: anything that changes a record, contacts a customer or involves money waits for a person.
- Grounding: answers come from retrieved data, with sources, and the agent says when it does not know.
- Logging: every request, tool call and refusal is recorded for review.
- Prompt-injection defences: text inside an email or document is treated as data, never as an instruction.
- Cost limits: usage budgets and a kill switch.
Which do you need?
If the job is to answer common questions from a fixed body of information, a chatbot is likely enough. If the job involves looking up live information or preparing an action, and the volume justifies it, an agent is worth considering. Either way, agents still fail sometimes, so the design should expect that and keep people able to catch it.
Our guided demo shows an agent working through a request with its permission checks and approval step, using fictional data.
Talk to an engineer
If this is relevant to something you are planning, we are happy to talk it through, with no obligation.
Discuss a project