Surprisingly Useful AI Article Enhancements
Nine in 10 finance leaders feel career pressure to prove the return on their AI investments, but only a fraction say their organisations are putting governance first.
That is the central tension in new research from tax and compliance platform provider Avalara, which surveyed 250 CFOs and senior finance leaders in Australia as part of a broader global study of more than 1,500 respondents across the US, UK, India and Australia.
The report, titled “Agents of Change: How the Race to Deploy AI Agents is Outrunning Financial Governance,” paints a picture of finance teams caught between the push to adopt agentic AI and the slower work of building the controls needed to manage it.
The pressure is real
Among Australian respondents, 88 per cent reported feeling moderate or significant career pressure to demonstrate that their AI agent investments are delivering ROI. Half described that pressure as significant.
Globally, the figure climbs to 92 per cent, with half of all respondents worldwide calling the pressure significant. Regional differences are stark: 57 per cent of US-based respondents described the pressure as significant, compared with 32 per cent in Australia.
The pressure is also heavily weighted toward speed. In Australia, 59 per cent of respondents indicated the push to deploy agents is focused primarily on getting them into production quickly. Only 12 per cent said their organisation prioritises governance over deployment speed.
Globally, 71 per cent of respondents indicated the pressure they feel relates entirely or mostly to the speed of deployment. Just 7 per cent reported that their business is putting governance first in its deployment strategy.
“Australian finance leaders are right to move quickly to capitalise on agentic AI opportunities, but speed without accountability creates new forms of risk, and speed without rethinking workflows limits ROI,” Hugo Sarrazin, Chief Executive Officer at Avalara, commented.
“The organisations that realise the greatest value from AI won’t simply deploy more agents. They’ll leverage agents with trusted data, governed workflows and clear controls that enable automation with confidence.”
ROI remains limited
While 90 per cent of Australian respondents reported that their AI agent initiatives are delivering at least some measurable ROI, the global picture is more nuanced. Across all markets, 88 per cent reported at least some return, but half of those described the returns as limited. Twelve per cent reported no clear return at all, or indicated it was too early to tell.
That figure rises to 24 per cent among finance leaders based in India, suggesting some markets are lagging in their ability to realise returns from agentic AI investments.
A governance gap
The report highlights what it describes as a growing gap between the pace of AI agent deployment and the maturity of the governance frameworks meant to oversee it.
Globally, 76 per cent of the finance leaders surveyed indicated they lack dedicated in-house expertise to understand how their AI agents work. Instead, they lean on IT departments (20 per cent) or their AI vendor (20 per cent) for support. A fifth of respondents are actively recruiting for internal expertise, while 16 per cent have nobody responsible for understanding their agents at all.
Thirty per cent of respondents have not updated their internal control framework in the past year to account for AI agents taking or recommending actions. Close to half (46 per cent) have AI incident response plans that are either untested or still in development.
“Finance leaders are being asked to move quickly with AI, but governing agents requires a new combination of domain, AI, IT and data governance expertise,” Frank Cirone, VP of Commercial Strategy at Snowflake, observed.
Related: Best Business Laptops for work & school
Related: Best Gaming Laptops
Related: Best Portable Laptop
“As AI agents gain access to financial and compliance workflows, organisations need to know what those agents can see, what they can do and when human approval is required. That kind of control has to be built into the architecture, not added after the fact.”
Who is accountable when agents fail?
One of the more striking findings in the report concerns accountability. When asked who would be held responsible for a significant AI agent error, nearly a quarter (23 per cent) of global respondents indicated accountability would be unclear or sit with no one. In India, that figure rises to 27 per cent.
Other responses were fragmented: 20 per cent pointed to the individual who deployed or manages the agent, 19 per cent to the team responsible, 22 per cent to the vendor where contractually specified and 16 per cent to the executive who approved the AI investment.
Among Australian respondents, 18 per cent indicated accountability for a significant error would be unclear or sit with no one, while the same proportion believed the executive who approved the investment would be held personally accountable.
The report also examined vendor accountability. The most common contractual arrangement (42 per cent) is a shared liability model between vendor and customer. Thirty-eight per cent of contracts make vendors fully liable for agent errors, while in 17 per cent of cases the vendor bears only minimal accountability.
One Australian finance leader quoted in the report argued that contractual standards should place greater accountability on vendors for AI-related issues.
Confidence in explaining AI to regulators
In Australia, 59 per cent of respondents described themselves as only “somewhat confident” they could explain an AI agent’s actions to an auditor or regulator.
Globally, just over half (53 per cent) expressed strong confidence in their ability to do so, while 44 per cent were only somewhat confident. Ten per cent reported low confidence, with the figure particularly pronounced in India, where one in 10 respondents indicated they were not confident at all.
“Organisations are increasingly exploring how to use AI to do reconciliations, validations and much more,” Cirone added. “The risk comes from the delegation of authority that happens when those things start to be infiltrated by AI. You may be indirectly giving an agent access to data that you do not want them to have.”
How vendors are vetted
The governance gaps extend to how organisations evaluate the vendors supplying their AI agents. A third of those surveyed reported that no safeguards are formally required before a vendor’s AI agent is allowed to operate. Just 31 per cent require documented human review or escalation thresholds, and only 28 per cent require documented audit logs showing how the agent reaches its decisions.
Close to a third (29 per cent) rely primarily on the vendor’s reputation and existing contract terms, with no formal technical or compliance requirements in place.
Where CFOs draw the line
Despite the rush to deploy, finance leaders are attempting to set boundaries. When asked about the role of general-purpose large language models in regulated workflows, opinion was divided: 23 per cent considered them inappropriate for decision-making in regulated workflows, while another 23 per cent viewed them as inappropriate for finance, tax or compliance workflows altogether.
Twenty-nine per cent believed LLMs could support regulated workflows, but only if every output is reviewed by a human. A further 24 per cent indicated LLMs should be restricted to low-risk tasks such as summarising, transcribing or drafting.
James Erwin, Head of Indirect Tax at Zayo, stressed the importance of collaboration between finance and technology teams: “You can’t give a tax team an AI tool and expect transformation, just as you can’t ask AI engineers to solve complex tax problems on their own. The strongest results come when tax professionals and AI specialists work together to build solutions that are accurate, practical and scalable.”
What would build confidence
When asked what capabilities would increase their confidence in expanding AI agent use, respondents pointed to measures that reinforce control and accountability. Twenty-seven per cent identified agents operating within the rules, permissions and controls of existing systems of record. Twenty-five per cent wanted outputs grounded in verified tax, compliance or financial data, and the same proportion sought evidence that outputs are tested against known compliance requirements.
Audit-ready documentation for every AI-driven action was identified as the most valuable capability for financial operations, selected by 37 per cent of Australian respondents.
“AI agents are now moving into business processes that require trust, transparency and governance by design,” Jim Lundy, Founder, CEO and Lead Analyst at Aragon Research, remarked. “As enterprises scale agentic AI, the question becomes less about whether the technology can act and more about whether organisations can understand, control and explain those actions.”
Mark Gaeto, CFO of Anderson Process, echoed the concern: “As AI agents begin taking actions instead of simply generating recommendations, organisations need new governance models that clearly define what an AI agent can do, who approves its actions and how those actions are monitored. AI tools are evolving faster than corporate policies and control frameworks.”
Five steps to close the gap
The report concludes with five recommendations for finance leaders. First, it urges organisations to build AI agent accountability into their organisational structure, noting that the lack of dedicated expertise is the most actionable gap. Second, it calls for insistence on verified compliance data rather than reliance on general-purpose AI. Third, it recommends keeping humans in control of high-stakes decisions. Fourth, it presses organisations to test their incident response plans before they are needed. Fifth, it argues that governance should be built into the architecture from the start rather than added after deployment.
Jayme Fishman, EVP and Chief Strategy and Product Officer at Avalara, framed the findings as a set of standards finance leaders are right to demand: “The pressure to adopt agentic AI is real, the desire to get it right is genuine, and the gap between the two is wider than most organisations expected.”
The research was conducted by Censuswide among CFOs and senior finance leaders aged 30 and over, at companies with revenue of approximately $14 million or more (US$10 million or more). Respondents had deployed, piloted or actively evaluated AI agents within their organisations in the 12 months prior to the survey. Industries covered included financial services and insurance, healthcare, manufacturing and distribution, professional services, retail and eCommerce, and tech and SaaS. Fieldwork was conducted between 15 and 22 June 2026.
Last Updated on July 26, 2026 by Nick Ross
