What is the difference between an AI agent, a chatbot and an automation?
The three terms get used interchangeably in sales decks, but they describe different things with different costs and risks. The useful question is who decides what happens next: a rule you wrote, a person in a conversation, or a model.
| Workflow automation | Chatbot / assistant | AI agent | |
|---|---|---|---|
| Who decides the next step | Rules you define | The user, turn by turn | The model, towards a goal |
| Typical job | Move data, send reminders, create records | Answer questions, qualify a lead | Research, triage, multi-step tasks across tools |
| Predictability | Very high | High, if grounded in your content | Lower; needs guardrails and review |
| Running cost | Mostly fixed (hosting, API calls) | Per conversation (model usage) | Per task, often several model calls |
| Failure mode | Breaks loudly when an input changes | Gives a wrong or vague answer | Takes a wrong action |
| Good first project | Yes | Yes, for repeated questions | Rarely |
When is plain workflow automation enough?
If you can write the process down as "when this happens, do that", you need automation, not AI. Lead capture from a website form into a CRM, WhatsApp reminders before an appointment, a daily sales report, an invoice copied into accounting — these are rules. Tools such as n8n, Make or a small custom service run them reliably and cheaply, and they are easy to audit.
AI earns a place inside an automation only at the step that needs reading or judgement: pulling fields out of a scanned invoice, summarising a long enquiry, classifying a message by intent. The rest of the workflow stays deterministic.
When does a chatbot make sense?
A chatbot pays off when customers ask the same questions many times a day and the answers already exist in your documents: prices, timings, eligibility, how to book. Ground it in your own content, let it hand over to a person, and log what it could not answer — that log tells you what to write next.
- Good fit: clinics answering timing and booking questions, real-estate enquiries asking for budget and location, course admissions questions.
- Poor fit: anything where a wrong answer has legal, medical or financial consequences without a person checking it.
When do you actually need an AI agent?
Agents make sense when a task needs several steps across tools and the steps change case by case — for example, reading an enquiry, checking stock and past orders, drafting a quote and routing it for approval. Even then, start with the agent proposing actions and a person approving them. Give it permission to act on its own only for actions that are cheap to undo.
In what order should a small business adopt them?
- Automate the repetitive, rule-based work first — it has the clearest payback and teaches you where your data lives.
- Add a grounded chatbot or AI summary where people spend time answering or reading the same things.
- Pilot an agent on one narrow, reversible task with human approval, and measure it against the manual process before widening it.
Where this fits in our work
This guide relates to our ai automation service. For a worked example, see Raybotix Workspace, agency operations.