Australia’s AI Inquiry Has a Question Nobody Can Answer: Who Pays When Your AI Agent Spends More Than You Meant?

AI agents have started making payments on Australian cards. That simple fact has opened a question that, according to a new report, nobody in the country's financial system can yet answer: if your AI agent spends money within the permission you gave it, but not in the way you meant, who carries the loss?

The question sits at the centre of a 106-page report published on 5 November 2026 by the Sydney-based Humanware Institute. Titled “AI on Australia's Terms”, it analyses the full public record of the federal Joint Select Committee on Artificial Intelligence, led by principal author Kate Cooper, the institute's founder.

The evidence is vast: all 564 public submissions, spanning 6,158 pages, ten hearing days totalling 53.4 hours, and 66 answers to questions on notice. Together, that is 575 distinct contributors.

Most people said yes to AI, then set terms

The report's central finding is not opposition to AI. Of the 575 contributors, 367 supported AI, and 290 of those attached conditions.

What people wanted, in short, was terms: skills to use AI well, a say before it changes their work, a fair hearing when it decides about them, control over what is built from their work and voices, and someone who answers when it acts on their behalf.

“The people and organisations who wrote to Parliament mostly said yes to AI. Then they set terms,” Cooper said in a statement accompanying the report. “Skills, a say, a fair hearing, and someone who answers when AI acts for them. The committee's job on 30 November is to turn those terms into conditions people can rely on.”

The committee's final report is due by 30 November.

Skills topped the list of asks

AI skills and literacy was the most common request, raised by 138 contributors, about one in four. The gap between demand and supply looks wide: MEGT, which works with about 44,000 employers, told the committee that only 3.7 per cent of employers providing AI training use nationally recognised training. Westpac told the committee the greatest constraint on adoption is no longer access to technology but workforce capability.

Eighty-two contributors asked for notice, reasons and a human review that can change an automated decision, and 61 asked for clear liability across the AI value chain. The human review point came with a sharp example: aged-care assessors could see an automated funding decision getting their clients wrong and were barred from overriding it. “A person checking a decision needs the power to correct it,” Cooper said.

The payment question nobody answered

The report's most concrete finding concerns money. Nine financial-system institutions gave evidence: three major banks, their peak body, four regulators and Treasury. Asked who answers when an agent acts for a consumer, the banks pointed to the companies that build agents, the consumer regulator pointed to Treasury, the markets regulator pointed to the law as it stands, and the Productivity Commission pointed to a review not yet done. None said who carries the loss.

Mastercard reported Australia's first authenticated agent transactions in January 2026, in trials on two major banks' cards. The Reserve Bank's Review of Payments System Regulation, which covers agentic commerce, does not announce its conclusions until 2028.

The report draws a careful line here. Existing rules for unauthorised transactions address a payment the account holder did not authorise. The new case is different: an agent acting within its permission but beyond the person's intention. The report finds the application of existing law to that case contested or untested, and its comparison of eight other governments found no settled allocation of loss for it either. “Agents are already booking, cancelling and paying on people's behalf,” Cooper said. “That can be settled now, before it becomes a dispute, and payments is the place to start.”

The institute recommends settling responsibility for AI agents in payments first, because the consequences are tangible and rules for unauthorised transactions already exist.

Workers, creators and control

The evidence also shows a sharp divide over who gets a say when AI changes work. Seventeen of 18 unions asked for consultation before AI changes work, compared with one of 35 banks and other businesses.

Creators had their own stories. A voice actor told the committee he found his cloned voice earning money on a major video platform and spent 12 months and more than 100 emails trying to enforce his rights, without success. A voice is not a copyright work, the report notes, so copyright offers no remedy.

Then there is control versus location. Of the 195 submissions that addressed the question, about 190 said Australian hosting alone does not guarantee public benefit. Two consultancies told the committee what happened on 13 June 2026, when a US Department of Commerce order suspended access to two Anthropic models for users outside the United States. Access returned on 1 July. “Keeping data in Australia does not keep control in Australia,” Cooper said. “Control is who can switch the service off, who gets the capacity, and whether you can leave.”

Pace, not just rules

The report also turns to delivery. It maps relevant responsibilities across at least 19 Commonwealth ministers and agencies, alongside Parliament and the states, and calls for one named point of accountability for coordination, pace and reporting, not a new regulator. The pace problem is documented in the timeline: Australia's largest AI commitments have waited two to three years, and 17 of the 20 major model releases came after the government committed to a law on automated decisions in November 2023. That law still has no date. “Australia has plenty under way. What it lacks is pace, and someone to set it,” Cooper said.

The committee's final report lands by 30 November.

Frequently Asked Questions

What is Australia's AI inquiry?

It is the federal Joint Select Committee on Artificial Intelligence, which gathered public submissions and held hearings between 18 September and 22 October 2026. Its final report is due by 30 November 2026.

Who wrote the Humanware Institute report?

The 106-page report, “AI on Australia's Terms”, was published on 5 November 2026 by the Sydney-based Humanware Institute, with Kate Cooper as principal author. It analyses all 564 public submissions, ten hearing days and 66 answers to questions on notice.

Who pays if my AI agent overspends?

Right now, nobody can say for certain. Nine financial-system institutions gave evidence and none identified who carries the loss when an agent acts within its permission but beyond the person's intention. Existing unauthorised-transaction rules cover payments you did not authorise, which is a different case.