What is an AI agent?
Less magic, more job description.
“AI agent” is being used for everything from a website chat window to software that claims to run a department. For a business owner deciding whether to pay for one, the useful question is narrower: what job does it do, what is it allowed to touch, and what happens when it is unsure. This guide answers those in plain English.
The examples below are the agents we build and sell, described the way they are set up. Where something is not live yet, we say so.
The short answer.
An AI agent is software built on a language model that is given a goal, the tools it may use and the limits it must keep, and then decides its own next step until the goal is met or it needs a person. A chatbot answers the question in front of it. An automation follows fixed rules. An agent sits between them: it can read a messy situation, choose what to do next and act through connected tools, which is exactly why it needs clearer limits than either.
Chatbot, automation, agent: side by side.
| What starts it | What it decides | Main risk | |
|---|---|---|---|
| Chatbot | A person asks a question | The wording of one answer | A confident wrong answer |
| Automation | A fixed trigger, such as a form or a date | Nothing; it follows the rule | Breaks on input the rule did not expect |
| AI agent | A goal and a situation to work on | Its next step, within set limits | Acting where it should have asked |
What an agent is made of.
Four parts. When an agent misbehaves, one of them was set up badly.
A goal
One job, stated plainly: get a reply on this quote, confirm this booking. An agent with a vague goal makes vague decisions.
Facts it may use
Your approved business information: services, hours, policies, standard answers. It should answer from these and nothing else.
Tools it may touch
The inbox, the calendar, the job system. Each connection is a permission, so it gets only the ones the job needs.
Limits and a hand-off
When to stop, when to ask, and who it hands over to. This is the part that makes an agent safe to leave running.
A worked example: the quote that went quiet.
How our Lead Follow-Up Agent is built to work, step by step, once our SMS service is live.
It notices the wait
A quote or enquiry has had no reply for the time agreed with you at setup.
It prepares the message
A follow-up written from the quote's own details and your approved wording, on the channel agreed with you.
A person approves, until trusted
By default anything going out in your name is drafted for your approval first.
It stops when they reply
The moment the customer answers, the agent steps back and the conversation is yours.
This agent is built to work by text message, and our SMS service is not live yet, so it will not text customers until it is. It includes 250 text message segments a month. Extra segments cost 15 cents each, GST included, billed monthly in arrears, and unused segments do not carry over. A message over 160 characters counts as more than one segment. Your written quote states what runs from day one.
When an agent is the wrong answer.
When the step needs a judgement call you could not write down for a new staff member. When a mistake cannot be undone, such as a payment, a contract or a refund. When it happens too rarely to matter.
And when the information it would need lives in one person's head. Writing that down comes first, whatever happens next.
In those cases a checklist or a plain automation is cheaper and safer, and we will say so rather than sell an agent for it.
Agents are what AI means in practice for most businesses today. The bigger words, AGI and superintelligence, describe something nobody has built yet; our superintelligence explainer covers what they mean and what is worth doing about them.
Give an agent one job.
Tell us which job in your business waits on someone remembering it. We will tell you whether an agent fits, which one, what it costs and where a person stays in charge.