AI agents for business in Australia: a practical guide
What an AI agent actually is, when a plain workflow is the better answer, what one costs to run in Australian dollars, and the controls to build before it goes near a customer.
An AI agent is software that uses a language model to choose its next step and then take it in your systems: reading a record, drafting a reply or updating a field. For most Australian businesses the first useful agent does one named job, reads before it writes, and sends anything uncertain to a person.
This page is the map. It covers what an agent is and how it differs from a chatbot or a HubSpot workflow, when a job needs one, what it costs in Australian dollars, what goes wrong, and what Australian guidance asks you to build before you switch one on. Each section links to the longer guide we have written on that part.
What is the difference between an AI agent, a chatbot and a workflow?
The three get mixed up because they can all sit on the same screen. The difference is who decides what happens next, and what it costs each time it runs.
| HubSpot workflow | Chatbot | AI agent | |
|---|---|---|---|
| Who decides the next step | The rules you wrote | A script, or a model answering one question at a time | A model, working towards a goal you set |
| Can it take actions? | Yes, the ones you built in | Usually only answers | Yes, within the access you give it |
| Copes with variation | Badly. Anything outside the rules falls through | Within its script or knowledge | Well, which is also why it needs checking |
| Cost per run | Nothing extra inside a plan you already pay for | Depends on the tool | Per action or per outcome. A resolved HubSpot Customer Agent conversation is 50 US cents, about 77 cents AUD before GST |
| Good first job | Routing, reminders, field updates | Answering the same few questions on a website | Triage, research, drafting from several sources |
An agentic workflow is the combination: a workflow that hands one step to an agent and takes the result back, so the rules still run the process and the model only does the part that needs judgement. HubSpot's own version is covered in running the Data Agent inside your HubSpot workflows.
When does a job need an agent, and when will a workflow do?
Use a workflow when you can write the rule down. Use an agent when every case is a bit different and a person currently reads something before deciding. Support questions phrased a hundred ways, prospect research, and pulling an answer out of several documents are agent jobs. Sending a reminder 30 days before a balance is due is a workflow job, and paying per run for it is money spent for nothing.
The test, with the AUD numbers, is in do you need an AI agent, or will a HubSpot workflow do it? If the job lives in HubSpot, start with what HubSpot's Agent Hub and Agent Builder launched with before you build anything custom.
How does an agentic workflow run, and where does a person step in?
Every agent we build follows the same shape. Something triggers it, it reads only what its job needs, it decides, and then the risk of the action decides whether it runs on its own, runs and tells someone, or waits for a person.
The Green, Amber and Red tiers are the action policy we run on our own agents, published in our AI Security and Governance Overview. A senior person reviews the Red queue before anything in it goes ahead.
Two parts of that picture carry most of the safety. Reading only what the job needs means the agent gets its own scoped access, never a person's login. And the risk tiers mean the decision about what can run unattended is made once, in writing, before launch. Our post on the handoff nobody's job is to catch covers the moment between the agent and the person, which is where most builds go quiet.
What can AI agents do for an Australian business?
The agents that last do a job a team can name. These are three we built on Claude and run in production, each with a review step and a log of every run:
- Sales briefings for Plungie. Written from its own HubSpot and sales data, for teams in Australia and the US, so a salesperson has a short brief on each buyer before the call.
- A business briefing for NHFA. Written regularly from a data warehouse we built and run for them.
- A fact-find for Flint. It reads the documents a borrower uploads, fills the loan record and flags mismatches for a broker to check.
Inside HubSpot, the Breeze agents cover a lot of first jobs without a custom build. We have written up the Prospecting Agent, the Closing Agent and the checks it needs, and the number to get before you buy an AI support agent. For outbound sales, read whether an AI agent can replace an SDR in Australia before you let one send anything, because the Spam Act still applies to the business.
What do AI agents cost to run in Australia?
Two numbers matter: what it costs to set up, and what it costs every month after, because an agent spends money each time it runs.
| Route | What you pay | Watch for |
|---|---|---|
| HubSpot Breeze agents | HubSpot credits, from the pool that comes with your subscription tier, plus per-outcome pricing on the Customer Agent. HubSpot usually bills Australian accounts in US dollars | Credits do not roll over, and every team draws from one shared pool |
| Your own API key | The model provider's usage price, counted in tokens | Volume. A build that handles 200 records a month and one that handles 200,000 can cost the same to make and a thousand times more to run |
| A custom build with us | As a guide from our signed engagements, ex GST: a prototype $15,000 to $25,000, a build $20,000 to $50,000, running it $8,000 to $15,000 a month, with model usage at cost | Data clean-up is part of the build, and it is the line most quotes leave out |
The custom build ranges are the ones on our AI consulting page, which sets out what each stage includes. The full breakdowns are in what HubSpot's AI agents cost to run in AUD, Breeze credits or your own API key at 150 people and what a custom AI build costs an Australian business.
What goes wrong with AI agents?
The failures are rarely the model being clever in the wrong way. They are ordinary gaps that an agent makes faster and harder to see.
| What goes wrong | What it looks like | What we build in |
|---|---|---|
| Someone hijacks it by email | An email the agent reads contains instructions, and the agent follows them. Australia's cyber agency uses an email-monitoring agent as its own example | Untrusted text never sits beside the power to act. Scoped access, and a person before anything leaves the building. More on prompt injection |
| It acts on bad data | Four records for one company, and the agent picks one and answers confidently | Five data checks before launch, starting with duplicates on ABN. The five checks |
| It does something you cannot undo | An email is sent, a deal stage changes, two contacts merge | Irreversible actions sit in the Red tier, with versioning and rollback planned before launch. Can you undo what an agent did? |
| It is confidently wrong in front of a customer | A fluent, specific answer that is false | A fixed test set, scored by someone other than the builder, with a pass bar agreed first. How we test an agent |
| It stops being useful and nobody notices | The workflow ticks over and the dashboard stays green while quality slides for weeks | Alerts on silence as well as errors, and a weekly look at real outputs. The failure nobody builds for |
| Nobody can fix it | The builder left with the prompts and the logins | Prompts and keys in your accounts, under version control, with a runbook. Why AI bolted onto HubSpot breaks |
The customer-facing case is worth remembering. In February 2024 a Canadian tribunal ordered Air Canada to pay $812.02 after its website chatbot gave a customer the wrong bereavement fare rule. The airline argued the chatbot was responsible for its own words. The tribunal did not accept that, and nobody else will either.
What does Australian guidance say about AI agents?
There is no AI Act in force in Australia, but the guidance on agents is specific. On 1 May 2026 the Australian Signals Directorate's Australian Cyber Security Centre published Careful adoption of agentic AI services with its US, UK, Canadian and New Zealand counterparts. Its starting point is to assume an agent will sometimes behave in ways you did not plan for, and design for that: least privilege, logging by default, rollback, and a named person accountable for what the system does.
The National AI Centre's Guidance for AI Adoption sets out six practices, and the last of them is to keep human control: a person can intervene, override and stop it. And from 10 December 2026 the Privacy Act requires your privacy policy to describe automated decisions that could significantly affect someone, which can include an agent's. All of it is pulled together in AI governance in Australia.
How do you start with AI agents?
Small, and in this order. It is the same order we run a blueprint and prototype in, where the prototype runs in shadow mode for 4 to 8 weeks: it does the work alongside your team and changes nothing until its output has been checked.
- Name one job. One a person does every week by reading something and deciding. Write down what good looks like.
- Check the data it will read. Duplicates, stale fields and who is allowed to see what. Fix those first.
- Give it its own access, reading first. A scoped key or integration user, never a person's login, and no writing until its reading has been checked.
- Test it before a customer sees it. Thirty or more real cases with known answers, scored by someone other than the builder.
- Name the owner and switch on the log. One person who is alerted when it fails or goes quiet, and who can stop it.
The National AI Centre's six practices map onto these steps: accountability, impact, risk, information, testing and human control.
If HubSpot is where the work lives, the data checklist to run before you switch on a Breeze agent is the place to begin. If you want a second opinion on the job you have in mind, our AI consulting team in Brisbane scopes it with you, and our AI engineers build and run it on Claude. Claude for business in Australia covers the plans and where your data goes.
Frequently asked questions
What is an AI agent? An AI agent is software that uses a language model to decide its next step and then take it in your systems, such as reading a CRM record, drafting a reply or updating a field. It works towards a goal you set, within the access you give it.
Is an AI agent the same as a chatbot? No. A chatbot answers questions. An agent takes actions as well, which is why it needs its own scoped access, a log of what it did and a person who approves anything risky or irreversible.
How much does an AI agent cost in Australia? HubSpot's Breeze agents run on credits, and a resolved Customer Agent conversation is 50 US cents, about 77 cents AUD before GST. A custom build with us is $20,000 to $50,000 ex GST as a guide, with running costs of $8,000 to $15,000 a month and model usage at cost.
Is it safe to connect an AI agent to our CRM? It can be, if the agent has its own scoped access, reads before it writes, logs every action and waits for a person on anything irreversible. Check your data first, because an agent believes whatever your CRM tells it.
Do AI agents have to follow Australian law? The business does. The Privacy Act, the Spam Act, consumer law and workplace law all apply to what an agent does on your behalf. From 10 December 2026 your privacy policy must also describe automated decisions that could significantly affect people.
Read next
- Do you need an AI agent, or will a workflow do?
- The five data checks we run before an agent
- The test every AI agent has to pass first
- Can you undo what an AI agent just did?
- Can someone hijack your AI by emailing you?
- What HubSpot's AI agents cost to run in AUD
- Who owns the AI system after we leave?
- AI governance in Australia: the 2026 guide
Sources
- Careful adoption of agentic AI services, Australian Signals Directorate and partners, 1 May 2026.
- Guidance for AI Adoption, National AI Centre, 21 October 2025.
- Guidance for transparency in automated decision making, OAIC.
- HubSpot pricing, for Breeze credits and agent pricing. Prices change, so check on the day you budget.
- AI Security and Governance Overview, Neighbourhood, for the Green, Amber and Red action policy.
Neighbourhood builds AI agents on Claude for Australian businesses, and sets up HubSpot's own agents where HubSpot already does the job.
Which job in your business suits an agent?
Book 15 minutes and bring one job your team does every week by reading something and deciding. We will tell you whether it needs an agent or a workflow, roughly what it would cost to run, and what your data needs first.