Avalara published research on July 21, 2026, showing that finance leaders are being pushed to deploy AI agents quickly and to prove those agents are paying for themselves, while the governance and internal controls that would support such a claim have fallen behind. The report, "Agents of Change: How the Race to Deploy AI Agents is Outrunning Financial Governance," draws on a survey of more than 1,500 CFOs and senior finance leaders in the United States, the United Kingdom, India, and Australia whose organizations have deployed, piloted, or seriously evaluated AI agents in financial processes over the past year.
The central finding is a mismatch between what finance is being asked to demonstrate and what it is equipped to demonstrate. Ninety-two percent of respondents report moderate or significant career pressure to show that their AI agent investments are delivering a return, with half describing that pressure as significant, and half saying those initiatives have produced only limited measurable return to date. Seventy-one percent say the pressure to deploy is focused mainly on speed, and only 7 percent say their organization puts governance ahead of it.
"Finance leaders are right to move quickly to capitalize on agentic AI opportunities, but speed without accountability creates new forms of risk, and speed without rethinking workflows limits ROI. The organizations that realize 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."
Hugo Sarrazin, Chief Executive Officer, Avalara
What makes the finding useful for anyone working on AI return measurement is where the agents have landed. Avalara reports that agentic AI has moved into tax, compliance, financial close, accounts payable, and invoicing, which are processes that get audited and are difficult to unwind once something has gone wrong. A return figure produced from those workflows has to survive a question from an auditor, and a large share of the finance leaders surveyed say they could not answer that question with confidence today.
Conditions Driving This Change
Agentic AI has moved out of pilots and into tax, compliance, financial close, accounts payable, and invoicing, which places it inside workflows that are audited and hard to reverse when an error surfaces.
Boards and senior executives have started asking the finance function to account for AI spending, and 92 percent of the leaders surveyed report moderate or significant career pressure to show those investments are returning something.
The same executives are asking for speed at the same time, since 71 percent say deployment pressure is focused primarily on how fast agents ship while only 7 percent say their organization puts governance first.
Half of the finance leaders surveyed say their agent initiatives have delivered only limited measurable return so far, which tends to increase the pressure to show results rather than relieve it.
Most finance teams lack anyone internally who understands the systems they are being asked to account for, with 76 percent reporting no dedicated in-house finance expertise on how their AI agents actually work.
Internal control frameworks were written for software that behaves the same way every time it runs, and 30 percent of respondents have not updated theirs within the past year to reflect agents taking or recommending actions.
Auditors and regulators are entitled to an explanation of a financial action regardless of what produced it, and 44 percent of respondents say they are only somewhat confident they could provide one for an agent.
What AI ROI Looked Like Before This
Finance has measured the return on software for decades using a method that worked because the software held still. A company counted what it paid in licenses and implementation, counted the hours or headcount the tool saved, and compared the two over a defined period. The arithmetic was rarely glamorous, and it was defensible, because the system under review behaved the same way in December as it had in June.
Accountability was equally settled. A piece of accounting software executed the instructions a person gave it, so when a number came out wrong, the trail led back to whoever entered it or whoever configured the system. The control frameworks that governed those processes were built on that assumption, and they had been refined across many years of audits, which is why nobody spent much time asking who was responsible for the software itself.
Agents unsettle both halves of that arrangement at once. A system that recommends or executes an action in a tax filing or a payment run is making a decision that used to belong to a person, and the return it produces depends on how well those decisions hold up rather than on hours saved. The Avalara data suggests most finance functions are still measuring the first thing while being held responsible for the second.
What It Looks Like Now
The sharpest numbers in the report concern ownership. When Avalara asked who would be accountable for a significant AI agent error, no single answer commanded anything close to a majority, and the largest share of responses pointed nowhere in particular.
Who would be accountable for a significant AI agent error | Share |
|---|---|
Accountability would be unclear, or would sit with no one | 23% |
The person who deployed or manages the agent | 20% |
The team managing the agent | 19% |
The executive who approved the AI investment | 16% |
Figures as reported by Avalara and summarized by ERP Today. The four responses above account for 78 percent of respondents, with the balance distributed across other answers.
The surrounding controls are in similar condition. Seventy-six percent of respondents say they have no dedicated in-house finance expertise for understanding how their agents work, with some relying on IT, some relying on the vendor, and 16 percent saying nobody currently holds that responsibility. Thirty percent have not revised their internal control framework within the past year, and 46 percent have AI incident response plans that are either untested or still being written.
The India results, released separately on July 22, show the same pattern with sharper edges. Eighty-five percent of Indian finance leaders report moderate or significant pressure to prove a return, only 8 percent say their organization prioritizes governance over speed, and more than one in four say accountability for a serious agent error would be unclear or would sit with nobody, which Avalara notes is higher than the other markets surveyed. Twenty-eight percent say their agent controls have been reviewed or tested by IT or cybersecurity teams, and 34 percent report a review by risk or compliance.
"While Indian enterprises are moving fast to automate, their internal rulebooks are being left behind."
Dulles Krishnan, Vice President and General Manager, Avalara India
Avalara also asked what would raise confidence enough to expand agent deployments, and the answers describe a shopping list built around evidence. Finance leaders pointed to agents operating within the permissions and controls of existing systems of record, outputs grounded in verified tax and compliance data, proof that those outputs have been tested against known compliance requirements, vendor commitments on accuracy and accountability, and documented audit trails showing what an agent did and why it did it.
Our Take
AI ROI Take
The easy reading of this survey is that governance is lagging deployment, which is true and which nearly every write-up of the report will say. The more useful reading is that these finance leaders are being asked to produce a number their own governance makes impossible to calculate honestly. A return figure has two sides, and the cost side of an AI agent includes what its failures cost and what supervising it costs. When 23 percent of respondents cannot name who owns an agent error, and 46 percent have incident response plans nobody has tested, the losses that belong in that calculation have no owner and therefore no reliable measurement.
Our AI ROI metrics framework calls that layer risk-adjusted return, meaning the value an AI system created minus what its failures and oversight actually cost, and it is the layer almost nobody measures. Avalara has now supplied the reason it stays empty. A CFO working under a 92 percent pressure rate can still produce a productivity figure, since hours saved is the number these tools generate without argument, and that figure describes the upside of a system whose downside remains unassigned. Any organization serious about proving an AI return should treat naming an accountable owner for each agent as a measurement task rather than a compliance chore, because the owner is the person who can eventually tell you what the failures cost.
One caveat belongs on the research itself. Avalara sells tax and compliance automation, and the report concludes that finance leaders want agents operating inside existing systems of record with verified compliance data and documented audit trails, which is a fair description of what Avalara sells. The sample of more than 1,500 leaders across four countries is substantial and the questions look reasonable, so the findings deserve to be taken seriously with the sponsor's interest kept in view. Finance teams working out how to measure a defensible return can start with the four layers in our AI ROI metrics guide and compare the platforms that own each measurement layer in the AI ROI category at GetAIGovernance.net.
