“Automation”, “chatbot” and “AI agent” are often used as if they meant the same thing. They don’t, and choosing the wrong one leads either to brittle automation that breaks every month or to an AI system doing work that a simple script would do better.
Here is what each one is, what it is good at, and how they fit together.
Chatbots: answering
A chatbot holds a conversation. Older chatbots follow scripted decision trees (“press 1 for billing”); modern ones use large language models to understand free-form questions and reply in natural language, often grounded in a knowledge base.
Good at: answering common questions, guiding people to the right page or form, collecting details before a hand-off to a person, working around the clock.
Limits: a chatbot talks, but on its own it doesn’t do anything in your systems. When the conversation needs an action, such as a refund, a booking change or an account update, it has to hand off to a person or to something else.
RPA: repeating fixed steps
Robotic process automation (RPA) software mimics what a person does on screen: clicking, copying, pasting and typing across applications, following exactly the same steps every time. It is popular where systems have no API.
Good at: high-volume, stable, rule-based work. Moving data between two systems, generating routine reports, processing standard forms.
Limits: RPA follows rules, not intent. It breaks when a screen layout changes, struggles with unstructured inputs such as free-text emails or varied PDFs, and cannot handle cases its rules didn’t anticipate. Where an API exists, a direct integration is usually more robust than screen automation.
AI agents: deciding and acting
An AI agent uses a language model to work towards a goal. Given a task, it decides which steps are needed, calls tools (APIs, databases, search, business applications) to carry them out, looks at the results, and continues until the task is done or it needs help.
Good at: work that varies from case to case and needs some judgement. Triaging and resolving support requests, researching and drafting responses from several systems, reconciling mismatched records, handling exceptions that would stop RPA.
Limits: agents are probabilistic. They can misread a situation or take a wrong step, they cost more per task than a script, and they need careful design to be trustworthy.
Side by side
| Chatbot | RPA | AI agent | |
|---|---|---|---|
| Main job | Answer and guide | Repeat fixed steps | Decide and act towards a goal |
| Handles unstructured input | Yes | Poorly | Yes |
| Takes actions in systems | Not on its own | Yes, fixed steps | Yes, chosen per case |
| Predictability | Medium | High | Medium; needs guardrails |
| Breaks when | Questions fall outside its knowledge | Screens or rules change | Permissions or guardrails are weak |
| Best for | Front-line questions | Stable, high-volume tasks | Variable tasks needing judgement |
How to choose
Map the process step by step and ask of each step:
- Is it the same every time, with structured inputs? Use a direct API integration, or RPA if there is no API. Don’t pay for intelligence you don’t need.
- Is it a question someone needs answered? A chatbot or knowledge assistant grounded in your content.
- Does it vary, need judgement, or involve unstructured information, and lead to an action? That is where an agent earns its place.
Most real processes mix all three. A customer email might be read and classified by an agent, routine cases resolved through fixed automation, and unusual ones drafted by the agent for a person to approve.
Making agents safe in production
Agents act on your systems, so the controls matter as much as the model:
- Least privilege. Give each agent only the tools and data it needs, with read-only access wherever possible.
- Approval for consequential actions. Payments, deletions, customer-facing messages and anything hard to undo should pause for human approval until the agent has a proven record.
- Full logs. Record every decision, tool call and result, so you can audit what happened and why.
- Clear hand-offs. When confidence is low or a case is unusual, the agent should stop and route to a person with the context it gathered.
- Evaluation on real cases. Test against a set of real, anonymised cases, including the awkward ones, before launch and after every change.
Start with one well-defined process, measure it against how the work is done today, and expand once the agent has earned trust.