Insight
AI agents for business in Australia: what they are, what agentic means, and where to start
By Brett Raven, APAC Lead & Board Member, AIREPublished
An AI agent is software that takes a goal and works through a task on your behalf, using tools within set limits. Agentic AI describes the system: several agents and people handing work to each other toward an outcome. Start with one narrow agent on one high-volume task, with a person reviewing its work, before you connect agents into a wider system.
What changes
Searches for AI agents in Australia have risen every quarter this year. Most of what comes back is vendor glossaries. Here's the plain version, with what Australian businesses are actually running and where to start.
- An agent does one job. Agentic describes the system the agents work in.
- Westpac runs 5 narrow agents in its lending process, with bankers reviewing their work (iTnews, 2026).
- Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 (Gartner, 2025).
- From 10 December 2026, Australia's privacy law asks you to disclose decisions that programs make or inform.
What an AI agent is
An AI agent is software that works through a task for you, deciding its own next step. OpenAI puts it plainly: "Agents are systems that independently accomplish tasks on your behalf." It also says what isn't one: simple chatbots, single-turn LLMs and sentiment classifiers "are not agents" (OpenAI, 2025).
The difference is control. A chatbot answers. An agent decides what to do next, uses tools such as your systems or documents to do it, and keeps going until the task is done or it hits a limit you set.
Anthropic draws a useful line between workflows, where a model follows "predefined code paths", and agents, which "dynamically direct their own processes and tool usage". Its advice is to find "the simplest solution possible" and add complexity only when needed (Anthropic, 2024). For most business tasks, that means a well-defined workflow beats a free-roaming agent.
Agent or agentic: why the difference matters
We use "agent" for the building block and "agentic" for the system. Anthropic's vocabulary differs slightly: it calls the whole family "agentic systems" and reserves "agent" for the most self-directing kind. The practical point is the same.
| AI agent | Agentic system | |
|---|---|---|
| What it is | One piece of software with a model, a goal, tools and limits, doing one job. | Several agents and people handing work to each other until an outcome is reached. |
| Example | An agent that reads a payslip and checks it against lending policy. | A home loan application moving from document checks to assessment to a banker's decision. |
| Who owns it | The team whose task it does. | The owner of the end-to-end process. |
| How to govern it | Its scope, its tools and what it may never do. | The handoffs, the human decision points and the fallback when a step fails. |
| How to measure it | Accuracy and time saved on its one task. | The outcome: cycle time, error rate and customer result for the whole process. |
Mixing them up causes two mistakes. Teams govern a system like a tool, approving one agent and missing the handoffs where things break. And they measure a tool like a system, expecting one agent to move a business outcome on its own. Gartner's own forecast follows the same order: task-specific agents in 40% of enterprise applications by the end of 2026, then agents collaborating within applications by 2027, and networks of agents across applications by 2028 (Gartner, 2025).
Agentic AI and generative AI
Generative AI produces content: a draft, a summary, an image. Agentic AI uses that capability to act: it plans steps, calls tools and hands work on. Most agents run on a generative model. The difference is that they're allowed to do things, not just write things.
What Australian businesses are running today
The clearest Australian example is Westpac. Westpac's Andrew McMullan said in August 2026 that 5 specialist agents built on Amazon Bedrock AgentCore are "running in [Westpac's] production systems". They classify payslips, extract the data, run the calculations and check them against policy, before a banker reviews the output. He credited them with "more than 150,000 hours" saved, and said the time to get an agent into production fell from six months to "about four-to-six weeks" (iTnews, 2026).
Note the shape. Each agent does one narrow job. The banker stays in the loop. In McMullan's words, "specialist agents work together around the banker." NAB is taking the same care: it plans to test the security and operational guardrails of an agentic platform before wider pilots (Bloomberg, 2026).
Leaders are ahead of their teams. In Microsoft's 2025 Work Trend Index, 71% of Australian business leaders said they were familiar with AI agents, against 31% of employees (Microsoft, 2025).
Familiar with AI agents, Australia, 2025
- Business leaders
- 71%
- Employees
- 31%
Why agent projects fail
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, "due to escalating costs, unclear business value or inadequate risk controls". It also warns of "agent washing": existing chatbots and automation rebranded as agents. It estimates only about 130 of the thousands of agentic AI vendors are real (Gartner, 2025).
The three causes map to three questions to answer before you build:
- Cost. What does each run cost, and what happens to the bill when volume doubles?
- Value. Which number moves if this works: hours, cycle time, errors or revenue?
- Risk. What may the agent never do, and who reviews its output before it reaches a customer?
Where to start
- Pick one high-volume, rules-heavy task. Document checks, data extraction and first-pass triage are good candidates. Avoid anything where a wrong answer reaches a customer unchecked.
- Write the agent a job description. Its goal, the tools and data it may use, what it must never do, when it hands to a person, and the measure it's judged on.
- Keep a person reviewing its output. Westpac's bankers review every agent result before the application moves on.
- Measure for 30 days against a baseline. Time per task, error rate and how often the agent escalates.
- Only then connect agents into a system. Design the handoffs and the fallback when a step fails, and give the whole process one owner.
We design and build these with you: see AI automation and agent builds. Your people need to know how to work alongside them too, which our AI training for teams covers.
The rules that apply in Australia
Australia has no agent-specific law. Existing obligations apply, and one new one starts soon.
- Privacy, from 10 December 2026. If a program makes or informs a decision that could significantly affect someone, using their personal information, your privacy policy must say so. The privacy regulator says this applies even where the output is "subject to human review" (OAIC, 2026).
- The Guidance for AI Adoption. The National AI Centre's six practices include "test and monitor" and "maintain human control", which is exactly what an agent needs (National AI Centre, 2025).
Put agents in your AI policy before the first one goes live. Our AI policy template is a starting point.
What to do next
3 actions, in order:
- This week: list the 3 highest-volume, rules-heavy tasks in your business.
- This month: write a one-page job description for an agent on the best of them.
- Within 90 days: run it with a person reviewing every output, and measure it against the baseline.
Not sure which task to start with? Our advisory work starts with your workflows, not the tools.
Thinking about AI agents? We'll help you pick the first task, design the guardrails and build it with you.
Frequently asked questions
What is an AI agent?
Software that works through a task on your behalf, deciding its own next steps and using tools such as your systems or documents, within limits you set. OpenAI describes agents as systems that independently accomplish tasks on your behalf, and says simple chatbots are not agents.
What's the difference between an AI agent and agentic AI?
An AI agent is the building block: one piece of software doing one job. Agentic AI describes the system those agents work in, where several agents and people hand work to each other toward an outcome. Govern and measure each at the right level: the agent on its task, the system on the outcome.
What's the difference between agentic AI and generative AI?
Generative AI produces content, such as a draft or a summary. Agentic AI uses that capability to act: it plans steps, calls tools and passes work on. Most agents run on a generative model; the difference is that they're allowed to do things, not just write things.
Are AI agents safe to use in a business?
They can be, if each one has a narrow job, clear limits on what it may do and a person reviewing its output. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, mostly for cost, unclear value or weak risk controls, so start small and measure.
Do Australian privacy rules apply to AI agents?
Yes. The Privacy Act already applies to personal information an agent handles. From 10 December 2026, if a program makes or informs a decision that could significantly affect someone using their personal information, your privacy policy must disclose it, even when a person reviews the output.
Sources
We read each source on the date shown beside it. Where a figure is a company's own claim, we say who made it. Check the original before you quote a figure.
- iTnews, Westpac plugs five AWS AI agents into core lending processes (4 August 2026) (Westpac CIO Andrew McMullan, as reported). Read 21 October 2026.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025). Read 21 October 2026.
- Gartner, Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025 (26 August 2025). Read 21 October 2026.
- OpenAI, A practical guide to building agents. Read 21 October 2026.
- Anthropic, Building effective agents (19 December 2024). Read 21 October 2026.
- Microsoft, Digital team members will be common in Australian workplaces within five years: Microsoft report (29 April 2025) (2025 Work Trend Index, Australian results). Read 21 October 2026.
- Bloomberg (via The Star), NAB tests AI agents in plan to deploy tech to banking customers (4 August 2026). Read 21 October 2026.
- Office of the Australian Information Commissioner, APP 1.7-1.9 Transparency Obligation Fact Sheet (September 2026). Read 21 October 2026.
- National AI Centre, Guidance for AI adoption: implementation guidance. Read 21 October 2026.