"AI agent" is one of the most used phrases in technology right now. Vendors attach it to almost everything, from chat windows on websites to software that promises to run your business while you sleep.
If you own a small business, you have probably heard the term and wondered whether you are missing out. This article explains what an agent is in plain language, where it can genuinely help, where it can go wrong, and how to try one without taking on much risk.
What an AI agent actually is
An AI agent is software that is given a goal, works out the steps, and uses tools to carry them out with limited supervision.
The tools are what make it an agent. A chatbot can only talk. An agent can also do things: read an inbox, look up a booking, search the web, fill in a spreadsheet, draft a reply or update a record.
A useful way to picture it is a capable new starter in their first week. You give them a clear job, show them which systems they can use, tell them what they must check with you first, and review their work until you trust it.
Chatbot, automation or agent?
These three get mixed up constantly. The difference matters, because each one suits different work.
- A chatbot answers questions. You ask, it responds. It does not act on anything unless you copy its answer somewhere yourself.
- An automation follows fixed rules: "When a form comes in, add it to the spreadsheet and send the standard reply." It is reliable and predictable, but it cannot cope with anything outside its rules.
- An agent chooses its own steps towards a goal: "Go through this morning's enquiries, sort them by urgency, and draft replies to the ones about bookings." It can handle variety, but because it makes choices, it can also make mistakes.
Anthropic, one of the major AI developers, draws the same line in its guide to building agents. Workflows follow "predefined code paths", while agents "dynamically direct their own processes and tool usage". The same guide recommends "finding the simplest solution possible, and only increasing complexity when needed". (Anthropic)
That is good advice for business owners too. If a simple automation does the job, you do not need an agent.
What agents can do in a small business today
The most useful agents right now are narrow. They do one job well, inside clear limits. Some realistic examples:
- Inbox triage. Reading new emails each morning, flagging what is urgent and drafting replies for a person to check.
- Enquiry handling. Answering common questions about hours, prices or availability, and passing anything unusual to a person.
- Reporting. Pulling numbers from a few systems each week and writing a short summary of what changed.
- Research. Checking competitors, suppliers or tender sites on a schedule and reporting anything new.
- Admin follow-up. Chasing overdue paperwork, reminding customers about appointments or keeping a job list up to date.
Notice what these have in common. The work is repetitive, the rules are mostly clear, and a mistake can be caught before it reaches a customer.
What it looks like in practice
SI1 runs on agents itself, so we can describe this from experience rather than theory.
Every morning, an agent checks the SI1 contact inbox, sorts what came in and sends a short report. If a reply is needed, it writes a draft for a person to approve. It is set up not to send, delete or move anything on its own.
Another agent posts social content for Domestic Robot, our home robotics publication, following a written brief. In the Melbourne venue group where SI1's founder leads marketing, separate agents handle inbox triage, marketing and reporting, and a person approves recurring reports before they go to leadership.
None of this replaces judgement. What it does is take the first draft, the sorting and the checking off people's plates, so they spend their time deciding rather than assembling.
Where agents go wrong
The honest version of the agent story includes the failures.
They can be confidently wrong. AI can produce answers that sound right and are not. An agent that acts on a wrong answer turns a bad sentence into a bad action.
You are responsible for what they say. In 2024, a Canadian tribunal ordered Air Canada to compensate a customer after the chatbot on its website gave wrong information about bereavement fares. The airline argued the chatbot was responsible for its own actions. The tribunal called that "a remarkable submission" and found the airline responsible for all the information on its website. (CBC News) The same logic applies to an agent replying to your customers.
Access is risk. An agent with access to your email, files or accounts can make mistakes with them. Give it the least access it needs, and never hand it your passwords.
Privacy still applies. The Office of the Australian Information Commissioner says privacy obligations apply to personal information you put into an AI system and to what the system produces. It recommends against entering personal information, especially sensitive information, into publicly available AI tools. (OAIC)
How to start without the hype
You do not need a big project. A sensible first agent looks like this:
- Pick one annoying, repetitive job. Something that takes a few hours a week and follows a pattern, such as sorting enquiries or writing a weekly summary.
- Write the job down as if you were briefing a new staff member. What it should do, what good looks like, what it must never do and who it should ask when it is unsure.
- Start with drafts, not actions. Let the agent prepare the work and a person approve it. Only let it act on its own once it has earned that over several weeks.
- Limit what it can touch. Read-only access where possible. No payments, no deleting and no sending to customers until you are confident.
- Measure the before and after. Note how long the job takes today. If the agent is not saving real time after a month, change it or stop.
The Australian Government's National AI Centre publishes free Guidance for AI Adoption, including a foundations version for organisations that are just starting with AI. Its six practices include deciding who is accountable, testing and monitoring, and maintaining human control. It is a useful checklist even for a business of one.
For a wider view of where AI could help across your business, The AI Opportunity Map works through it function by function, and AI Strategy for Australian Businesses covers how to decide what comes first. If you would rather have agents designed for you, see AI Agents on our How we help page.
The bottom line
An AI agent is not magic, and it is not a digital employee you can leave alone. It is a capable tool that can take real work off your plate when it has a narrow job, clear rules and a person checking its output.
Start small, keep a person in the loop, and judge it on the time it saves rather than how impressive it sounds.