What Is an AI Agent? A Plain-English Guide

No, although they are built on the same underlying technology. ChatGPT is a chat interface: it waits for you to type something, generates a response, and stops. An AI agent wraps that same language-model capability in a goal, a set of tools, and guardrails, then runs without anyone typing prompts. When a supplier invoice lands in your inbox at 6am, ChatGPT does nothing because nobody asked it anything. An agent connected to that inbox reads the invoice, extracts the details, checks them against your supplier records, and creates a draft bill in Xero before your bookkeeper starts work. The practical difference is initiative: you drive ChatGPT, whereas an agent works its queue and only involves you when something needs a human decision.

Agentic AI describes systems that do things rather than just say things. A standard AI model generates text, images or answers when asked. An agentic system takes a goal, breaks it into steps, decides which step comes next, calls tools to perform each step, checks the result, and adjusts. The word gets used loosely in marketing, so apply a simple test: does the software decide its own next action and execute it in a real system, such as creating a record in MYOB or sending a reply from your support inbox? If a human still performs every action and the AI only suggests wording, it is an assistant, not an agent. If the steps are fixed in advance regardless of what the data says, it is a workflow, not an agent.

Yes, they make mistakes, and any provider claiming otherwise should worry you. Well-built agents are engineered around that fact. Every output carries a confidence score, and anything below the threshold is routed to a person instead of being committed. On invoice data extraction, for example, first-pass accuracy of around 94% improves toward 98%+ as the agent learns your suppliers, and the uncertain fields are exactly the ones flagged for human verification, so the effective accuracy of data actually entering your accounting system is higher still. Add approval rules for anything that moves money, a complete audit trail of every action, and a weekly review of the exception queue, and errors are caught before they cost anything.

Only the access required for its specific job, granted on a least-privilege basis. An invoice-processing agent needs to read a nominated inbox and create draft bills through the Xero or MYOB API using a scoped OAuth connection, not an administrator login. A support agent needs read access to order or booking data and write access to a ticketing queue. Good practice for Australian businesses is to start read-only, review what the agent would have done for a week or two, then grant write access to the narrowest scope that works. Every credential should be auditable and revocable, and because agents often touch personal information, access decisions should be documented against your obligations under the Privacy Act 1988 and Australian Privacy Principle 11 on security of personal information.

For most Australian small businesses, ongoing costs sit around $500-$1,500 per month for agents covering two to five workflows such as email triage, invoice processing and scheduling, with one-time implementation typically between $2,000 and $15,000 depending on how many systems need integrating. The sensible way to evaluate the spend is against the labour it replaces: manual invoice processing alone costs $30-$40 per invoice once error correction and delays are counted, so a business handling a few hundred invoices a month often covers the entire agent cost from that single workflow. Most implementations pay for themselves within three to six months, which is why we recommend starting with one high-volume workflow and expanding only once the first is measurably saving money.

AI agents are arguably more valuable to small businesses than to enterprises, because a five-person business feels every hour of administration directly. You do not need an IT department: agents run as a managed service, connect to the cloud software you already use, such as Xero, MYOB, Gmail, Outlook and Shopify, and are maintained by the provider. The realistic constraints are volume and process clarity rather than company size. If a task happens only twice a month, automation rarely justifies itself; if nobody can describe how the task is done today, the agent has nothing to learn from. A small business with one high-volume, reasonably consistent workflow, such as inbound email or supplier invoices, is an excellent candidate for a first agent.