AI may be doing more work inside finance departments, but it has also created a sizeable new task: checking what it produces. According to Datarails and Global Surveyz Research, US CFOs and finance leaders spend an average of 26% of their working day verifying or correcting finance-specific AI output. The report, ‘2026 CFO Sentiments: How AI Is Changing Finance Departments’, found almost all of those surveyed (96%) devote at least a tenth of their day to that work, while 40% spend more than a quarter of it.

AI is already routine across the companies surveyed. Every respondent reported using it within finance software, while 76% said they were under high or very high pressure to implement the technology fully.

Yet only 7% felt ready to introduce AI across every workflow and just 4% had a single source of truth for finance data. Finance has moved beyond the pilot phase, but human reviewers remain close to the numbers, particularly around forecasting, cash management and financial reporting.

AI reaches finance but not final sign-off

Microsoft Copilot was the leading general-purpose AI product, used outside Excel by 93% of respondents. ChatGPT followed at 65%, narrowly ahead of Claude at 64%, while Google Gemini reached 20% and Perplexity 13%.

Part of Copilot’s lead comes from its inclusion within Microsoft 365. Yet ChatGPT and Claude were almost level, suggesting finance teams are looking beyond the model already distributed through workplace software, particularly for variance analysis and driver-based forecasting.

Finance teams were using 2.5 large language models on average. For some, that may mean using one model to check another, adding yet another step to the verification work already falling on finance professionals.

AI has also moved inside the tools finance teams already use. AI features within ERP systems were used by 89% of respondents and Excel with Microsoft Copilot by 88%. Just over half, 53%, used AI within FP&A technology, while 44% had adopted it in expense or accounts payable tools.

CFOs were far more willing to leave administrative work to AI than the financial judgements for which they remained accountable. Some 89% would accept AI’s final output without human review for drafting emails and 84% would do so for meeting summaries.

That confidence fell to 46% for creating PowerPoint presentations, 45% for consolidating data from multiple sources and 44% for expense categorisation and coding. Once the work became more financially consequential, the numbers dropped sharply.

Only 23% trusted an unreviewed AI output for revenue forecasting. The proportion fell to 13% for variance analysis, 11% apiece for cash flow management and spend control, 5% for board-ready financial reports and 4% for month-end close. A further 3% would not accept AI’s final output for any financial task without review.

The time absorbed by those checks varied considerably. Most respondents, 56%, spent between 10% and 25% of their day verifying or correcting AI-generated finance work. Another 32% devoted between 26% and 50%, while 8% spent more than half their working day on it. Only 4% kept verification below 10%.

Finance teams had plenty of reason to check. Nearly two-thirds, 65%, had received a confident AI answer built on the wrong data during the previous year. More than half, 56%, had seen materially different outputs produced from the same prompt and data.

Another 51% had been unable to explain or trace how AI reached a number when challenged. In more serious cases, 22% later discovered that an AI-generated financial figure was fabricated, while 16% said a report containing an error had reached senior leadership or the board before the problem was caught.

Vendor claims had also disappointed 40% of respondents, who said AI capabilities they had been sold failed to deliver in practice. Respondents could select more than one problem, so the figures describe overlapping experiences rather than distinct groups.

Traceability sat at the heart of the trust problem. Three-quarters cited an inability to audit or explain how AI reached its answer, while 71% were concerned about accuracy and hallucinations. Regulatory or compliance concerns followed at 54%, with data privacy at 49%.

Fragmented financial data was identified by 38%, while 26% pointed to insufficient AI skills within finance. Only 12% cited a lack of executive support, underscoring how far the mandate to use AI has run ahead of confidence in the controls surrounding it.

Bad data turns speed into rework

Verification took longest where finance data remained fragmented. Only 4% said at least 91% of their core data was centralised and consistent, meeting the survey’s definition of a single source of truth.

Nearly three-quarters (73%) described their data as mostly centralised, meaning 71% to 90% was integrated but minor gaps remained. Almost a quarter, 23%, operated with data spread across multiple disconnected systems and needed frequent manual reconciliation. One respondent reported that less than 40% of core data was centralised.

Checking time rose sharply where systems were disconnected. Some 76% of those finance teams devoted more than a quarter of their day to verification, compared with 31% of respondents whose data was mostly centralised.

The same pattern appeared in output quality. Among CFOs who named manual reporting and data consolidation as their greatest operational challenge, 86% had received an AI answer built on the wrong data. The survey cannot prove that fragmentation caused those errors, but it shows how often the two travelled together.

Manual reporting and consolidation was itself the most commonly identified operational problem, selected by 32%. An inability to produce ad hoc analysis quickly enough for senior leaders followed at 19%.

Managing spending and expenses and keeping pace with AI and wider technology change each ranked first for 13% of respondents. A lack of real-time cash visibility was the main problem for 11%, while 8% chose an overly long month-end close and 4% cited difficulty recruiting and retaining finance talent.

Manual reporting, slow analysis and poor cash visibility are old finance problems. AI can accelerate the work, but incomplete or inconsistent inputs can also produce wrong answers faster. Finance then gives back much of the time saved through reconciliation, investigation and review.

Most teams sat in an uneasy middle. Some 71% called themselves mostly ready while acknowledging gaps in governance and auditability. Another 22% were only partly ready, with significant shortcomings in governance, auditability or skills.

Implementation pressure leaves little time to close those gaps. Some 67% said pressure to deploy AI fully was high and another 9% called it very high. A further 23% described the pressure as moderate and only 1% as low. Finance leaders are being asked to extend AI through their workflows while most still lack either the data foundation, the control framework or both.

Budget overruns fail to cool spending

CFOs are carrying much of the financial responsibility for that transition. Finance owned the organisation-wide AI budget at 84% of respondents’ companies, compared with just 3% where the CTO or IT department held it. Another 13% had no clearly identified budget owner.

The wider technology bill was already substantial. Finance departments spent an average $361,000 on finance and accounting software during 2025. Two-thirds spent more than $250,000, including 18% that paid at least $500,000 and 9% whose expenditure exceeded $750,000.

The most common spending band was $250,000 to $350,000, reported by 32%. Another 22% spent between $150,000 and $250,000, while 17% fell between $350,000 and $500,000. Only 1% spent less than $100,000, underlining the scale of the existing software estate into which AI is being introduced.

AI costs have not always followed plan. Almost a third of organisations, 32%, exceeded their allocated AI budget by at least 10% during the previous 12 months. Most of that group, 30% of the full sample, ran between 10% and 25% over budget, while 2% exceeded the allocation by between 26% and 50%.

A 57% majority kept expenditure within 10% of its budget and 11% spent at least 10% less than allocated. Those overruns have done little to weaken demand: 53% of finance leaders planned to expand AI licences during the coming year, 40% intended to maintain their current level and only 7% expected to cut or consolidate them.

The familiar finance wish list still led buying plans. Planning and FP&A software topped the next-purchase list at 42%, reflecting continued demand for better forecasting and budgeting. Financial close and reporting tools attracted 14%, while 12% prioritised a business intelligence or visualisation layer over existing data.

Second place went to what the survey calls a finance operating system, selected by 32% and defined as a governed data layer producing consistent, auditable and AI-compatible figures. Datarails, which launched its FinanceOS product in March 2026, uses the term for its own offering as well as the category discussed in the research.

The finance operating system category appealed most to teams already feeling the strain. Half of the CFOs struggling to keep pace with AI and technology change prioritised such a system, as did 49% of those whose main problem was manual reporting and consolidation. The proportions fell to 19% among those constrained by ad hoc analysis and 11% among those focused on managing spend and expenses.

AI changes roles faster than headcount

Headcount produced the clearest departure from earlier expectations. Some 60% said they were moving staff whose routine work had been absorbed by AI into higher-value roles.

Only 3% were actively reducing headcount where AI had replaced work. Elsewhere, 9% had frozen recruitment without cutting existing roles and another 9% said it was too soon to make employment decisions. A further 7% were adding staff because of AI, while the same proportion was growing for unrelated reasons. Another 5% planned no AI-related change.

So far, the clearest change is in job content and where people are deployed. Recruitment freezes point to continuing staffing pressure, while outright AI-related cuts remain rare.

That marks a change from the previous edition of the CFO Sentiments survey, in which most respondents expected AI to produce layoffs within their departments. The latest findings capture reported actions rather than earlier expectations, although recruitment freezes show that AI can still affect staffing without generating redundancies.

Global Surveyz Research administered the online study in July 2026 for Datarails. It covered 270 US CFOs and finance leaders working at organisations with at least 1,000 employees and annual revenue of $100m or more, spanning 16 industries.

Directors represented 36% of respondents, vice-presidents 35% and C-suite executives 29%. Some 71% worked for businesses employing between 1,000 and 2,000 people, with the remaining 29% at larger organisations. Segment-level findings were based on groups ranging from 35 to 197 respondents.

The findings are based on self-reported experiences among large US companies, making them a focused view of that market rather than a measure of every finance department’s adoption.

Within this large-company sample, AI is already a CFO-owned operating cost. Licences are expanding and the technology sits inside core finance systems, yet very few departments will let it handle forecasts, cash or board reporting without supervision.

Access is no longer the main constraint. Finance leaders now face the harder work of giving models consistent data, tracing how figures were produced and matching human review to the risk of each task. Until those controls catch up, a sizeable part of AI’s promised productivity gain will continue to be spent checking AI itself.