Australian CFOs Are Deploying AI Agents They Can't Explain

Seventy-five per cent of Australian CFOs surveyed by Avalara in June said no one on their finance team can explain how their AI agents actually work. The sample was 250 leaders at companies with more than $10 million in revenue — mid-market firms with resources most Australian SMEs don't have. The next answer in the survey is worse: 88 per cent feel personal career pressure to prove those agents are delivering ROI, and only 12 per cent said governance was the priority in how they were being deployed.

The gap between how fast AI agents are moving into Australian finance functions and how well anyone can explain them to an auditor is now the biggest single risk in the stack.

The "Agents of Change" report, published 22 July, surveyed 1,500+ CFOs and senior finance leaders across the US, UK, India and Australia — 250 of them here — every one at a company that had deployed, piloted or evaluated an AI agent inside a financial process in the past year. Fieldwork was done by Censuswide from 15 to 22 June 2026, so the numbers are two months old at most.

The Australian cut is unusual. Only 32 per cent of Australian respondents called the pressure to ship agents 'significant' — versus 57 per cent in the US, the lowest of the four markets. That looks like a good thing until you read the follow-up: 59 per cent of the same Australian respondents said they were only 'somewhat confident' they could explain what an AI agent had done if an auditor or regulator asked. Less deployment urgency, still can't defend the work.

If your AI agent misreads a supplier invoice, over-claims a GST credit, or lets a customer refund through when it shouldn't, the legal question in Australia is not 'was that the software's fault?' It is whether the director or officer who approved the system exercised the care and diligence a reasonable person in their position would — section 180 of the Corporations Act 2001. In April, the Federal Court in *ASIC v Bekier* [2026] FCA 196 endorsed the view that boards must actively govern AI use, not tolerate it informally. Justice Lee treated the s180 duty as personal and non-delegable — AI can help a director discharge the duty; AI cannot replace their judgement.

Avalara found only 16 per cent of the executives it surveyed globally think the person who approved the AI investment would be held accountable if the agent errored. In Australia that is a misread of the risk. The executive who signed off is exactly the person s180 is written about. The same shift is happening with insurance: professional indemnity policies are moving toward 'affirmative' AI cover, which means non-affirmative policies may not respond to an AI-caused claim at all. Two independent lines — the courts and the underwriters — converging on the same person.

If your name is on the AI sign-off and you can't get a clean answer from your finance team about what's already deployed, book a 30-minute governance review before the next agent goes live.

Want to know what this means for your business? We'll map it in one call.

Avalara asked its 1,500+ CFOs what capabilities would most increase their confidence in deploying agents. The top four asks are a useful checklist: agents that operate within the rules, permissions and controls of existing systems of record (27%), evidence that outputs are tested against known compliance requirements (25%), outputs grounded in verified data (25%), and contractual accuracy and accountability commitments from the vendor (24%). None of those is 'the model is impressive.'

Three moves worth making this quarter, in order. First, map every AI agent already inside your finance stack — the ones your bookkeeping software vendor switched on last quarter, the AP tools offering them by default, the Xero and MYOB add-ons that now include one. Almost a third of Avalara's respondents haven't updated their internal control framework in the past year, which is roughly the same window in which most agents were deployed. Second, before granting any new agent access, work through the ACSC and AI Safety Institute checklist for agent access: start on low-risk tasks, add privileges slowly, keep a log. Third, ask the vendor two questions in writing. Where does this agent's answer come from? And what does your contract say if the answer is wrong? If the vendor is vague on either, your defence is vague when the mistake gets found.

Australia's lower urgency is a rare advantage — a few months to install governance before market pressure catches up with the US level. The firms that come out of this cycle in a defensible position won't be the ones that shipped agents fastest. They will be the ones that can hand an auditor, an insurer, or an ASIC officer a clean answer to a simple question: what does this agent do, who approved it, and how do you know it's right?

Key takeaways

Common questions

Related notes

Want a governance check before your next AI agent goes live?

We map what's already deployed inside your finance stack, what each agent can reach, and what your existing controls actually cover — one honest conversation about where the exposure sits.