AI agents for Australian small business: when to build one

What is an AI agent?

An AI agent is a program that uses an AI model such as Claude to complete a task on its own: it starts on a schedule or when something happens (a new email, a new order), reads from and writes to your tools, works through several steps, and hands back a result. The useful ones are narrow, do one job well, and have a person approving anything that leaves the business.

The word gets used loosely, so it helps to see the ladder from plain chat to a full agent:

Level	What it is	Who starts it	Example
Chat	You ask, AI answers	You	Drafting a reply in Claude or ChatGPT
Project or custom assistant	Chat with your business context loaded	You	A Claude Project with your price list and tone
One-off script	A program that runs a multi-step AI task when you press go	You	”Categorise this month’s Xero export”
Scheduled agent	The same task, run automatically on a timer or trigger	A schedule or event	Every morning at 6am, triage the overnight inbox
Customer-facing agent	Responds to customers in real time	Your customers	Website chat that answers questions and books
Multi-agent system	Several agents coordinating	Other agents	Rarely worth it at small business scale

Most of the value for a small business sits at the scheduled-agent level: single-purpose, running in the background, with output a person reviews.

Does my small business need an AI agent?

Probably not yet if you have not used AI by hand for the same job. Build an agent when you have done the task with AI chat at least several times a week for a month, the inputs look roughly the same each time, the output can be checked before it acts, the time saved clearly beats the running cost, and someone will review its work daily for the first month.

Run this five-question test before you spend anything:

1. Have you done it by hand with AI, repeatedly? Doing the job in chat first gives you the working prompt and shows you the edge cases.
2. Are the inputs predictable? “New enquiry from our website form” is predictable. “Whatever customers email us as attachments” is not.
3. Can a person check the output before it acts? Drafts for approval are low-risk. Sending, posting, paying or changing records without review is not, at first.
4. Does the maths work? Rough estimate: if an agent saves 30 minutes a day and that time is worth $60 an hour to you, it is worth about $650 a month. Compare that with build and running costs.
5. Will someone watch it for the first month? Agents drift when inputs change. Daily review in month one, then weekly.

If you are new to AI chat itself, start with our beginner’s guide to Claude for small business.

What can AI agents do for an Australian small business?

The agents that pay off for small businesses are unglamorous: overnight inbox and enquiry triage, stock and listing checks for online stores, first replies to new leads, Xero transaction categorisation for your bookkeeper, and a weekly summary of what happened. Each does one job, uses tools you already run, and leaves the final decision with a person.

Agent	What it does	Typical tools	On Autopilot package
Enquiry and booking triage	Replies to website chat, SMS and email enquiries, books appointments, escalates the rest	Cliniko, Nookal, ServiceM8, Cal.com, Google Calendar	AI Front Desk, $1,500 setup + $199/month
Stock and catalogue check	A daily check, before the team starts, of stockouts, low stock, slow movers and listing errors	Shopify, Xero, Slack or email	AI Inventory Watch, $497 setup + $99/month
Lead qualification and reply	Scores each new enquiry, drafts a reply in your voice, offers a booking link, updates the CRM	HubSpot, Pipedrive, Calendly	AI Lead Engine, $2,000 setup + $499/month
Bookkeeping helper	Suggests Xero categories daily, preps month-end, runs BAS-ready checks for your bookkeeper or BAS agent	Xero, Slack	AI Bookkeeping Assist, $1,500 setup + $399/month
Weekly briefing	Summarises bookings, sales, enquiries and overdue tasks every Monday	Calendar, inbox, Shopify or Xero	Usually part of managed AI

We run agents like these in our own businesses. Our skincare brand Hey Bud runs an inventory system that syncs warehouse stock and Shopify sales at 6am every day, raises stock and reorder alerts, drafts purchase orders for a person to approve and puts a Claude-written daily brief on its dashboard. Our VA agency DotVA runs every website enquiry through an automated pipeline that scores spam and job applicants before a person sees it. AI for Australian Shopify stores walks through the inventory system in more detail.

How much does an AI agent cost to build and run in Australia?

Building costs anywhere from your own time (no-code tools) to a fixed-price productised build to a bespoke project. Running costs are usually small for scheduled agents: the AI usage behind our productised agents typically comes to $3–15 a month. Customer-facing agents handling many conversations cost more, and every agent needs someone to maintain it as connected tools change.

• Do it yourself: no-code tools such as n8n, Make or Zapier charge a monthly subscription; the AI usage is billed by Anthropic or OpenAI on top. Your time is the main cost.
• Productised build: On Autopilot’s agents range from $497 setup plus $99 a month (Inventory Watch) to $2,000 setup plus $499 a month (Lead Engine), AI usage included. See pricing.
• Run and maintained for you: managed AI from $1,500 a month for up to two live agents.
• Learn to build your own: a Claude Code Setup Day is $1,500 fixed and leaves you with a working agent on your machine.

How do you build an AI agent safely?

Build it narrow, give it the least access it needs, keep a person approving anything that leaves the business, cap its spending, and make it tell you when it fails. These five habits prevent most agent problems, and they are how we build our own: the Hey Bud inventory system, for example, drafts purchase orders for a person to approve instead of placing them, skips every step that cannot be undone when the warehouse data looks incomplete, and emails critical problems.

1. Prove the job in chat first. Keep the prompt that worked.
2. Narrow the scope. One job, clear inputs, a defined output.
3. Start read-only. Let it read your Shopify, Xero or inbox before it is allowed to change anything.
4. Keep a person in the loop. Drafts for approval until it has earned trust over weeks, not days.
5. Set a budget cap on your Anthropic or OpenAI account so a loop cannot run up a bill.
6. Alert on failure. A silent failure is worse than a loud one.
7. Log what it does and review the log weekly for the first month.

What can go wrong with AI agents?

The common failures are cost blow-outs from agents with no spending cap, agents given too many tools and too much access, agents acting without approval and getting something wrong in public, no fallback when the agent is unsure, and silent failures nobody notices for weeks. Agents that read emails or web pages can also be manipulated by instructions hidden in that content, called prompt injection.

Your business is responsible for what its agents say and do, as Air Canada found when a Canadian tribunal held it liable for refund advice its website chatbot got wrong (Moffatt v. Air Canada, 2024 BCCRT 149). In Australia the same logic runs through the Consumer Law and the Privacy Act. Our AI privacy guide covers what data an agent should and should not see.

How On Autopilot can help

On Autopilot builds the agents in the table above as fixed-price packages: AI Inventory Watch ($497 setup + $99/month), AI Front Desk ($1,500 + $199/month), AI Bookkeeping Assist ($1,500 + $399/month) and AI Lead Engine ($2,000 + $499/month). If you want several agents built, run and improved for you, managed AI starts at $1,500 a month, month to month, and you own the code, prompts and settings. A free 30-minute AI audit with Jenn Yang will tell you whether an agent is worth building yet, or whether chat is enough for now.