Most “AI agents” are a chat window with a new badge. Here is the boring version — what the tiers actually are, what a worker does step by step, and what Australian privacy law expects of you before you switch one on.
Search “AI agent Australia” today and you’ll get a hundred pages that all say the same thing: agents are the future, they work while you sleep, book a demo. Almost none of them tell you what the thing actually does between the prompt and the outcome.
That vagueness is not an accident. Gartner has a name for it — “agent washing”: vendors rebranding existing chatbots and automation scripts as agentic AI without shipping genuine autonomy. Of the thousands of vendors making agentic claims, Gartner estimated only around 130 offered real agentic capability, and predicted that over 40% of agentic AI projects will be cancelled by the end of 2027 on cost, unclear value, or inadequate risk controls.
If a vendor’s “agent” finishes by showing you a paragraph describing what it would do, that is a chat window wearing an agent badge. A worker finishes with an artifact you can open, edit and send — and with a record of how it got there.
The word “agent” gets applied to all three. They are not the same product, and the difference is easiest to see in what each one leaves behind.
| Chat | Assistant with tools | Autonomous worker | |
|---|---|---|---|
| Who decides the next step | You | You approve each tool call | It plans and executes the sequence |
| What you give it | A question | A question plus permissions | A goal and a finish line |
| What it produces | Text in a window | Text, plus a lookup or an action | Finished files — a doc, a sheet, a deck, a schedule |
| What it leaves behind | A transcript | A transcript | Real artifacts you can open, edit and send |
| Runs while you’re away | No | No | Yes — scheduled or event-triggered |
| Good for | Thinking out loud | Research with live data | Recurring work with a defined output |
Here is the honest sequence inside Oppermind, with no mystique attached.
“Every Monday, pull last week’s numbers from the finance sheet, build the board update deck, and leave it in my workspace” is a goal. It has a trigger, inputs, an output format, and a place the output lands.
The worker breaks the goal into steps before it runs any of them. In Oppermind this is a shipped mode — Workspace Agent 1.0 plans and uses tools; Workspace Agent 2.0 goes further and delivers finished documents, sheets, decks and designs straight into their editors.
Dry-run previews let you see the intended sequence before anything executes. This is the single most useful safety feature in agentic software, and the one most often missing.
Human approval gates hold the run at the steps you nominate — before it sends, before it spends, before it writes to a system of record.
Not a log line saying “generated report” — the actual report, in an editor, versioned, editable, exportable. That artifact trail is also your evidence later, when someone asks what the system did and why.
Oppermind is an MCP client: you bring your own connectors and servers, from a library of 40 templates, and the worker uses what you have authorised. There is no mystery integration layer — you can see which connectors are attached and revoke them.
The desktop web workspace and mobile web are home base. Beyond that: a Chrome extension that puts the assistant in a side panel on any page, with that page’s context, and can drive the tab; a browser operator that watches the page, acts, and shows you what it did (when available); a Windows desktop co-pilot that reads your screen, moves the cursor, types and runs jobs on your PC (a separate install, Windows only); task panes for Word, Excel and PowerPoint that read your selection and edit the open document (downloadable, manual install); an Android app; and a developer API. The iOS app is coming to the App Store.
If a step needs a small capability that does not exist yet, the platform can build a custom tool for it, then use it.
Australian businesses are adopting this faster than the coverage suggests. The ABS found 12% of businesses used AI in 2024–25, up from 1% in 2021–22 — 35% of large businesses, 22% of medium, and around 11% of small and micro businesses. That is a lot of first deployments happening without a governance conversation.
Three things are worth knowing before an agent touches customer data.
The OAIC has published specific guidance on privacy and the use of commercially available AI products. Its blunt best-practice position: given the privacy risks, organisations should not enter personal information — particularly sensitive information — into AI chatbots that are not configured to protect it. Agents raise the stakes, because an agent does not just receive data, it moves data between systems.
Under Australian Privacy Principle 8, before personal information is disclosed to an overseas recipient you must take reasonable steps to ensure that recipient does not breach the APPs. And under s 16C of the Privacy Act, if the overseas recipient breaches the APPs, you are treated as if you had breached them yourself. Every connector you attach to an agent is a potential disclosure. Map them before you switch it on.
The Privacy and Other Legislation Amendment Act 2024 inserts APP 1.7, requiring APP entities to disclose in their privacy policy when personal information is used in automated decisions that significantly affect people, and what kinds of information and decisions are involved. The OAIC has signalled a broad reading and has been running compliance sweeps of privacy policies. If you are planning an agent deployment this financial year, your privacy policy is part of the deployment.
None of this is an argument against agents. It is an argument for the ones that show you their plan, gate their own actions, and leave an artifact trail you can hand to a compliance officer.
One subscription, one login, one bill — a full suite of AI tools and editors with the autonomous worker included, rather than a separate agent product bolted onto a chat licence. Unlimited documents, spreadsheets, decks and designs on every plan.
| Plan | Price (AUD / month) |
|---|---|
| Free | A$0 — no card |
| Starter | A$9.95 |
| Pro | A$29.99 |
| Pro Plus | A$59.99 |
For the comparison against a typical multi-tool AI stack — the A$130+/month version most teams are already paying — see One AI subscription replaces 8.
The fastest way to tell a worker from a chatbot is still the same test: give it a goal with a deliverable, walk away, and see whether there is a file when you come back.
A chatbot answers in a window and leaves you a transcript. An autonomous worker takes a goal, plans the steps, uses the tools you connected, and leaves behind finished files — a document, a spreadsheet, a deck, a schedule. The simplest test is whether there is a file at the end of the run.
Agent washing is Gartner’s term for vendors rebranding existing chatbots and automation scripts as agentic AI without shipping genuine autonomy. Gartner estimated that of the thousands of vendors making agentic claims, only around 130 offered real agentic capability, and predicted over 40% of agentic AI projects will be cancelled by the end of 2027.
Under Australian Privacy Principle 8, before personal information is disclosed to an overseas recipient you must take reasonable steps to ensure that recipient does not breach the APPs, and under s 16C of the Privacy Act a breach by that recipient is treated as your own breach. From 10 December 2026, APP 1.7 also requires APP entities to disclose in their privacy policy when personal information is used in automated decisions that significantly affect people.
In Oppermind the autonomous worker is included in every plan rather than sold as a separate agent product. Plans are Free with no card, Starter at A$9.95 per month, Pro at A$29.99 per month, and Pro Plus at A$59.99 per month, billed in AUD.
A full suite of AI tools and editors, with the autonomous worker included in every plan.