The 21% Problem: Finance Is Funding AI Faster Than It Is Banking the Returns

Finance has rarely had an easier time getting an AI budget approved, and rarely a harder time proving what the money bought. A survey of roughly 300 finance and accounting professionals published on September 16 found that 66.5% of organizations are increasing AI investment, 24.2% now treat it as a top budget priority, and a vanishing 0.7% are cutting spend. The same dataset found that only 21.0% report meaningful, measurable results, while 64.8% describe their outcomes as mixed or outright unsuccessful. Enthusiasm is not the constraint. Evidence is.

Two Surveys, One Number

The figure would be easy to dismiss as one vendor's framing if a very different study had not produced the same result. Deloitte's CFO Signals survey polled 200 finance chiefs at North American companies with at least $1 billion in annual revenue between November 14 and December 7, 2025. Among that group, 87% expect AI to be extremely or very important to finance operations in 2026, only 2% say it will not matter, and 54% name integrating AI agents as their single biggest transformation priority, ahead of the 52% who picked strengthening data quality, access and usability.

Then comes the other half of the same survey. As CFO Dive reported, 63% of those CFOs say they have fully deployed AI solutions, but only 21% believe those investments have delivered tangible value, and just 14% have fully integrated AI agents into the finance function. Two studies with different populations, different sponsors and different question sets arrived at 21%. Billion-dollar CFOs and mid-market controllers are describing the same shortfall.

"The willingness to fund AI is clear. What is still missing is consistent proof that it can carry real finance work from start to finish." – Rohit Gupta, CEO and co-founder, Auditoria.AI

Speed Is Real. The Cash Cycle Is Not.

The uncomfortable part is that AI is clearly working somewhere. Consero's 2026 CFO survey of 102 private equity and venture-backed finance leaders, fielded with Cascade Insights in the first quarter, found 97% now using or testing AI in finance, up from 74% in 2024. Fully embedded AI nearly doubled in a year, from 22% to 42%. Most tellingly, 65% of those teams now close the month in under 10 days, against 8% two years earlier.

That is a real operational gain, and it is concentrated in work finance controls by itself. Reconciliation, variance commentary, document extraction and first-draft reporting all compress well because nobody outside the department has to cooperate. The cash cycle is different, and it has barely moved. In the September survey, 44.5% of teams still spend 11 or more hours a week on follow-up and 13.4% spend 30 hours or more. Only 16.1% answer shared inbox requests in under two hours, and 43.8% manage it within 24. Asked to name their single biggest daily pain, 22.1% pointed at inaccurate vendor and customer information and 21.7% at shared inbox volume.

The pattern is consistent enough to be a planning assumption: AI is compressing the tasks a finance team can finish alone and leaving untouched the ones that end with waiting for someone else to reply. A faster close does not collect a receivable, and a model that drafts a dunning note does not resolve the dispute underneath it.

Where the Value Leaks: Integration, Data and Permission

Respondents were fairly precise about what stops the work. Integration leads at 36.7%, followed by high initial investment cost at 36.3%, security and privacy concerns at 34.5%, and difficulty finding suitable tools at 33.8%. Where results fell short, 29.2% said the technology simply underperformed. The data story is corroborated from the CFO seat: 52% of Deloitte's respondents ranked data quality, access and usability a transformation priority, and 33% of Consero's finance leaders named data readiness the number one blocker to AI ROI.

The quieter constraint is authority. Only 2.5% of finance organizations describe their operations as fully autonomous, 13.5% as governed autonomy and 26.3% as supervised. A third, 33.1%, still require a person to approve every AI action, and 48.8% keep execution entirely in human hands against 39.9% who let AI execute work. The average team now runs AI in 2.43 functions, a 36% increase year over year, so the footprint is widening even where the leash stays short.

None of that is irrational. Finance is the control function, and 75.4% of respondents said they are confident their reporting is accurate, a position nobody wants to trade for speed. But an agent that cannot act without a human clicking approve is an expensive autocomplete, and that is a plausible explanation for the 21%. The organizations seeing returns are not the ones that bought more; they are the ones that decided, function by function, what an agent is allowed to finish on its own.

"Finance does not need autonomy for autonomy's sake. The goal is to give AI exactly the authority it needs to get the work done, with the policies, permissions, controls, and accountability finance already expects." – Rohit Gupta, Auditoria.AI

The Playbook for the Next Budget Cycle

• Measure cash, not minutes: If the business case was written in hours saved, rewrite it against DSO, disputed balances and unapplied cash. Close speed is already improving without AI's help and makes a weak proof point.
• Fix the data before buying the agent: With 22.1% naming inaccurate vendor and customer records their biggest daily pain and a third calling data readiness the top blocker, master data work is the prerequisite, not the follow-up.
• Write the authority boundary down: Decide per process what an agent may complete, what needs supervision and what always needs approval. Blanket human sign-off on every action is what turns 63% deployment into 21% value.
• Budget for integration, not licences: Integration is the leading barrier at 36.7%. A pilot that never touches the ERP or the AR platform will demo well and change nothing.
• Target the wait states: Follow-up consumes 11 or more hours a week for 44.5% of teams. The best candidates for automation are the queues that depend on external parties, not the spreadsheets finance already controls.
• Set a kill date: With 64.8% of programs reporting mixed or unsuccessful outcomes, a pilot with no defined end and no exit criterion is how a department accumulates six of them.