"AI agent" has become the phrase of the year. Vendors attach it to everything from a website chat bubble to a fully automated back office, which makes it hard to know what you're actually being sold.
The distinction is simpler than the marketing suggests, and it matters, because the two solve different problems.
A chatbot talks. An agent does.
A chatbot is a conversation. Someone asks a question, the chatbot answers, usually from your own information: services, policies, opening hours, FAQs. A good one hands over to a person when it isn't sure. Its job ends when the conversation ends.
An agent is given a goal and works towards it in steps. It can read inputs (an email, a PDF, a form), look things up, make decisions against rules you've set, and take actions in other systems: create a record, draft a quote, send a message, flag something for review. It may never talk to a customer at all.
| Chatbot | Agent | |
|---|---|---|
| Starts with | A question from a person | A task, a trigger or a schedule |
| Main output | An answer | A finished piece of work |
| Works across systems | Rarely | That's the point |
| Best for | Repetitive questions, enquiries, triage | Repetitive multi-step jobs |
When a chatbot is the right answer
If your team answers the same twenty questions every week, by phone, email and DM, a chatbot trained on your own material is usually the quickest win. It can also qualify enquiries and book appointments, then pass warm leads to a person with the details already collected.
The important part is what it won't do: invent prices, make promises, or answer outside the information you've approved. We cover that in detail in our guide to keeping chatbot answers accurate.
When you actually need an agent
Agents earn their keep on jobs that follow a pattern but take a person's time:
- Reading construction plans and measuring quantities for a draft estimate, as we did for Logic Building Co.
- Watching competitors' products and marketing, then writing tailored reports, as we built for Dynamic Frontend.
- Turning a topic and a student's interests into a complete lesson, the engine behind Sprout Lessons.
None of these is a conversation. Each is a chain of steps that used to sit on someone's desk.
Keep a person in the loop where it matters
The agents that work well in real businesses aren't fully autonomous. They do the slow, repetitive part, then stop at a checkpoint: an estimate waits for review, a report waits for approval, an email sits in drafts. Anything involving money, legal commitments or a customer's trust gets a human sign-off.
That design choice is the difference between an agent your team relies on and one they quietly stop using after a bad week.
How to decide
Start from the problem, not the technology:
- Where does the time go? List the jobs that eat the most hours each week.
- Is it questions or work? Questions point to a chatbot; multi-step work points to an agent.
- What can go wrong? The higher the stakes, the more checkpoints you design in.
- What does it connect to? Agents are only as useful as the systems they can read and update.
Sometimes the honest answer is neither, and a better form or a simple automation does the job. That's worth knowing before you spend anything.