AI Auditability in Finance: Datarails 26% Survey

Finance teams at large US companies spend an average of 26% of their workweek verifying or correcting AI output, and 75% name AI auditability in finance as their biggest obstacle to trusting the technology, according to a Datarails survey of 270 finance leaders.

Lack of auditability ranked ahead of accuracy and hallucinations, cited by 71%, and regulatory or compliance concerns, cited by 54%. Respondents were asked what makes them hesitate to use AI for mission-critical tasks.

The finding matters to any firm choosing AI tools for the close, the forecast or the audit file. If the main objection is that a result cannot be traced, a more accurate model does not fix it. Controllers need to see inputs, logic and sign-off, and a vendor that cannot show them is adding review work, not removing it.

Datarails sells AI-enabled FP&A software, so it has a commercial interest in how the problem is framed. The survey is vendor-commissioned, and the figures below should be read with that in mind.

What the Datarails survey found on AI auditability in finance

Datarails published the research in September 2026 as its 2026 CFO Sentiments Survey. Data was collected in July 2026 by Global Surveyz Research, an independent firm, through a business-to-business research panel contacted by email.

All 270 respondents work in the United States, at companies with at least 1,000 employees and $100 million or more in annual revenue. By seniority, 29% are C-suite, 36% are directors and 35% are vice presidents. They span 16 industries.

Of the respondents, 96% spend at least 10% of their working time checking AI outputs, and 32% spend between 26% and 50%. Another 8% spend more than half their time on it.

The 26% figure is an average reported by Datarails. The methodology page describes the question as the share of the workday spent “verifying or correcting finance-specific AI outputs”, which means the number is self-estimated rather than timed. Breakdowns by segment rest on base sizes as small as 35 respondents.

Other findings point to organisations under strain. Some 76% report high or very high pressure to implement AI, and 32% exceeded their AI budgets by at least 10% in the past year. Only 4% say they have a single source of truth for data, and 65% say they have received confident AI answers based on incorrect data.

“What this survey shows is that although AI adoption is now standard for CFOs, caution hasn’t disappeared,” said Didi Gurfinkel, chief executive of Datarails.

Even so, 53% plan to expand AI licences over the next 12 months. Spending is rising while trust lags.

What auditability requires in practice

The summary of the survey does not define auditability. Finance teams generally mean four things when they use the word, and each maps to a control that already exists in the close.

An audit trail that links each output to the source data and the transaction it came from.
Logging of the model, prompt and version used, so a result can be examined later.
Reproducibility, meaning the same inputs produce the same answer on a re-run.
Named human review and sign-off before anything reaches a ledger or a board pack.

A tool that lacks any of the four forces the reviewer to rebuild the work by hand. That reconstruction is the verification time the survey records. Teams drafting controls for agents can start from Accountio’s guide to AI agent governance for finance teams.

KPMG has made a similar point from the audit side. Matt Johnson, KPMG US AI audit and assurance leader, said the risks created by sophisticated AI “demand a new playbook”, and that a cyber threat to an AI system is now a direct threat to the accuracy of financial information.

How other surveys compare on AI trust in finance

The closest comparison is research from Sage, which published a study with IDC on July 6, 2026. It surveyed 2,275 senior finance decision-makers, 63% in North America and 37% in Europe, the Middle East and Africa, at companies with 20 to 1,999 employees.

That study found finance professionals spend an average of 13 hours a week reconstructing, validating and defending AI outputs. It also found that 26% of expected AI time savings are absorbed by explanation work. The figure matches Datarails’ headline number by coincidence of definition, not method: Sage measures lost productivity gains, while Datarails measures the share of the working week.

Sage reported that 71% of finance leaders would reject an AI tool that could not explain its decisions, even at 99% accuracy. Some 54% would pay more for greater visibility into how outputs are produced. Aaron Harris, chief technology officer at Sage, said “in finance, almost right has always been wrong”.

Sage also sells accounting software with AI features, so its findings carry the same caveat as Datarails’. Both surveys point the same way, which is more persuasive than either alone, but both were commissioned by vendors with products to sell.

KPMG’s survey of 1,013 senior finance leaders in 20 countries, titled “AI in Finance: The Decision Advantage”, found 48% worried about the accuracy of AI-generated output. KPMG ranks accuracy as a concern but not the primary one in the summary reviewed. Its sample includes 163 respondents from the US.

Gartner’s 2025 AI in Finance survey of 183 CFOs and senior finance leaders found that 59% of finance functions were using AI, with adoption levelling off. It measures use and optimism, not auditability, so it is not a direct test of the Datarails result.

Why verification time matters to finance budgets

The 26% figure undercuts the usual business case for AI in finance. If a team saves hours on drafting variance commentary but spends a quarter of its week checking results, the net gain is smaller than the licence pitch suggests.

The effect depends on where the tool sits. A forecast produced from consolidated, governed data is easier to verify than one drawn from spreadsheets. Datarails’ own finding that only 4% of respondents have a single source of truth shows why: reviewers cannot trace an output to data that does not agree with itself.

For budget holders, the survey suggests asking vendors for evidence before signing. That means a demonstration of the audit trail, the log of model versions and the sign-off workflow, not only an accuracy percentage. Accountio’s finance software selection kit includes a scoring sheet where controls can be weighted alongside price and features.

Teams comparing FP&A products can use Accountio’s list of the best software for FP&A teams and its vendor comparison pages. Teams starting with the close can read how CFOs prioritise month-end close automation.

What to check before relying on the numbers

Several details limit how far the survey can be read. The sample is US-only, so it says little about UK or EU finance teams, who work under different audit and reporting regimes. It is limited to large companies, and the panel was recruited by email, which tends to favour people who respond to surveys.

Datarails has not published how respondents estimated verification time, and the survey did not compare it with time saved. The 26% figure comes from reporting of the survey; the methodology page the company published gives distribution bands. The average is described as a share of the workweek, while the methodology page refers to the workday.

A vendor survey also asks questions that suit its product. A respondent asked whether lack of auditability is a barrier is likelier to say yes than one asked to rank barriers freely. The full questionnaire would show which was used.

Finance leaders should treat the direction of the findings as credible, given that Sage and KPMG point the same way, and the precise numbers as indicative.

The next test is vendor behaviour. Sage has already put “glass box” design at the centre of its pitch, and Datarails is likely to follow with auditability features of its own. Buyers will see how many can show a reviewer the trail behind an answer before the year-end close.