The ROI Looks Great. The Governance Is Not Ready.

The numbers look good. That is the problem.

New research from Avalara has put a familiar tension in front of Australian finance leaders. The report, called "Agents of Change," surveyed CFOs and senior finance leaders here who have deployed, piloted or actively evaluated AI agents in financial processes over the past year, and the headline number reads like a win. (Source: Avalara, 'Agents of Change: How the Race to Deploy AI Agents is Outrunning Financial Governance', July 2026)

Ninety percent say their AI agent initiatives are delivering at least some measurable ROI to date. For anyone who has sat through two years of AI pilots that went nowhere, that is a result worth noticing. (Source: Avalara, 'Agents of Change', July 2026)

Then you read the next line. Eighty-eight percent of respondents feel moderate or significant career pressure to demonstrate that AI agent investments are paying off, with half calling that pressure significant. And only 12 percent say their organisation prioritises governance over deployment speed. (Source: Avalara, 'Agents of Change', July 2026)

That is the real shape of the risk. The value signal is arriving before the controls that make it defensible, and the people under the most pressure to show returns are the same people who will wear it when an agent gets something wrong.

Where the gaps actually sit

The survey points to three soft spots, and they are worth naming plainly.

First, explainability. Fifty-nine percent of Australian finance leaders are only somewhat confident they could explain an AI agent's specific actions to an auditor or regulator. In tax and compliance, where a decision has to withstand scrutiny after the fact, "somewhat confident" is not a position you want to be defending. (Source: Avalara, 'Agents of Change', July 2026)

Second, accountability. Nearly one in five, 18 percent, say responsibility for a significant AI agent error would be unclear or sit with no one, while another 18 percent believe the executive who approved the investment would be held personally accountable. (Source: Avalara, 'Agents of Change', July 2026)

Third, expertise. Around 75 percent lack dedicated in-house expertise to understand how their agents actually function, leaning instead on vendors or IT. (Source: Avalara, 'Agents of Change', July 2026)

This is not a finance-only story, and the wider Australian picture confirms it. Deloitte found that only 22 percent of Australian companies report having a highly advanced model for agent governance. (Source: Deloitte, 'State of AI in the Enterprise: The Untapped Edge', February 2026) On the technical side, Google Cloud's infrastructure research found that 83 percent of organisations globally need infrastructure upgrades before agentic AI can run in production rather than in a pilot, and local reporting puts 58 percent of Australian enterprises on architectures too rigid for it. (Source: eWeek, 'Can Australia's Legacy IT Handle the Agentic AI Rush', July 2026) The oversight gap runs to the top too: Microsoft's 2026 Work Trend Index found only 28 percent of Australian users said their leadership was clearly and consistently aligned on AI strategy and policies. (Source: Microsoft, '2026 Work Trend Index, Australian findings', June 2026)

So the pressure is real, the value is real, and the plumbing and the accountability lines are still being built underneath it.

A rollout sequence you can defend

The answer here is not to slow down for the sake of it. It is to sequence the rollout so that speed and defensibility move together. A few practical moves work for most finance functions.

Start where the blast radius is small. The organisations getting the most out of agents are starting with lower-risk use cases, building governance capability, and scaling deliberately from there. (Source: Deloitte, 'State of AI in the Enterprise: The Untapped Edge', February 2026) A reconciliation assistant is a better first agent than one that files a return.

Name an owner for every agent decision before it goes live. Not the vendor, not "IT," a named human who can say what the agent can see, what it can do, and where a person has to sign off. Build the audit log and the human approval step into the design, rather than adding them after something breaks.

Write down the incident plan and then test it. If an agent produces a wrong tax position, who catches it, how fast, and who talks to the regulator. A plan that has never been run is a hope, not a control.

Watch the regulatory horizon, because it is moving. On 15 July 2026 the Government announced plans to legislate an Australian Standards for AI framework, with a new Office of AI inside PM&C and standards expected to be legislated in early 2027. (Source: SafeAIAus, 'Guidance for AI Adoption (AI6)', July 2026) Governance built now is governance you are not scrambling to retrofit later.

The leaders who come out of this well will be the ones who treated the 90 percent ROI figure as a reason to get the controls right, not a reason to skip them.

The right next steps for you depend on where you are right now. Our AI Maturity Assessment is an easy place to start. After that, let's have a chat about your next steps.