When Your AI Agent Gets It Wrong, You're the One Liable

Six in ten Australian workers have caught AI getting it wrong

Six in ten Australian workers have already caught an AI tool making a mistake at work. That's from research the automation firm Appian released on 8 October. The figure that should hold a business owner's attention isn't the error rate, though — it's the timing. AI is shifting from a tool that drafts and answers to one that acts: placing orders, booking jobs, approving transactions, routing work. When an AI that only drafts gets it wrong, someone catches it before it leaves the building. When an AI that acts gets it wrong, the first you hear of it is the consequence.

The same survey — 500 Australian workers, conducted by Zoho Research — found only 14 per cent fully trust what AI produces, and 94 per cent think AI should not make decisions affecting employees or customers without human oversight. Appian sells automation software, so read the enthusiasm with that in mind. But the unease is the part worth keeping: workers want a human between the machine and the decision. On this, the law agrees with them.

In Australia, you own what your AI agent does

Australia has no AI Act, and that is exactly why the exposure sits with you. There is no special regime that quarantines an AI's mistakes from the business running it. Australian governance and legal commentators — among them the Governance Institute and law firm Mallesons — make the point plainly: the law treats an AI agent as part of your IT systems, not as a separate actor you can hold at arm's length the way you might a rogue employee. And it reaches the top of the business. Directors and officers can be personally exposed under the Corporations Act duty to act with reasonable care and diligence, and ASIC's stated view is that the duty now requires officers to know what AI their business uses and what it can get wrong.

"The algorithm did it" is not a defence, and the Federal Court has already said so. In ACCC v Trivago, it ordered the company to pay $44.7 million for conduct driven by its ranking algorithm — a system that quietly favoured the booking sites paying the highest fees over the cheapest rooms the consumer was promised. The dollar figure belongs to a big company; the principle belongs to every business. If your AI misquotes a customer, leaks their data, or makes a biased call in hiring, it is the Australian Consumer Law, the Privacy Act and anti-discrimination law that reach you — not the vendor whose logo is on the tool. Overseas the point is already concrete: a Canadian tribunal rejected Air Canada's argument that its own chatbot was responsible for the fare advice it gave a customer.

The fix is deciding where a human stays on the loop

None of this is an argument against AI agents. It's an argument for deploying them on purpose. The useful question isn't "do we trust the AI" — it's "which decisions can it make alone, and which need a name attached before they happen". Draw the line at consequence and reversibility. An agent drafting a quote, triaging the overnight inbox, reconciling the ledger or summarising a long file is low-risk and easily checked — let it run, with a person approving what goes out. An agent that sends money, commits you to a client, changes a customer's data or makes a call about a person is a different category. That one needs a human signing off, every time.

This is where most businesses are exposed without knowing it. In global research sponsored by Appian, 92 per cent agreed AI agents need rules-based guardrails — and only about half had actually defined them. The gap between buying the capability and governing it is where the liability lives. If you can't say which of your workflows an AI already touches, map where it sits before you hand over another decision. And before you buy the next "agent", check it's a real agent and not a chatbot in costume — you can't govern a capability the vendor can't actually deliver.

The pattern that works is the one Australia's Fair Work Commission used when it let AI draft and a human sign off: the machine does the grunt work, the person keeps the judgement and the accountability. You get the productivity. You don't get the 2am surprise. Write the rule down — which decisions are the agent's, which are a human's — and you've done more for your risk position than any vendor warranty will.