The Verification Tax: AI Use Is Routine Now, But So Is Checking Its Work
Finance teams at large U.S. companies now spend an average of 26% of their workweek verifying or correcting AI-generated output, according to Datarails' 2026 CFO Sentiments Survey of 270 finance leaders at firms with 1,000-plus employees and at least $100 million in revenue, published today. Ninety-six percent of CFOs say they spend at least a tenth of their time on this review work, and 8% say it now eats up more than half their week. Datarails CEO Didi Gurfinkel put it simply: "Although AI adoption is now standard for CFOs, caution hasn't disappeared."
That finding would be notable on its own. What makes it part of a bigger story is that it landed the same week as two other large, independently conducted surveys describing the identical pattern in different functions: AI use inside enterprises has stopped being a pilot-stage curiosity and become default behavior, and the labor of checking that AI's output is correct is becoming a second, under-budgeted job of its own.
The same bottleneck shows up in engineering
Bain & Company's 2026 Global Technology Report, based on a survey of nearly 300 senior technology leaders and covered by CIO Dive this week, found that developers using AI coding tools are completing about 21% more tasks and juggling 47% more concurrent workstreams than before. But code review time has risen 91% over the same period, and Bain estimates organizations are currently capturing only 20% to 27% of the productivity gains their leaders expected. Technology leaders surveyed still project a 148% improvement in release-cycle speed within one to two years — but only if the review bottleneck gets solved first.
"The bottleneck has moved from writing code to trusting it." — Purna Doddapaneni, Partner, Bain & Company
That line could just as easily describe what's happening in finance. In both domains, the constraint isn't whether AI can generate an answer, a forecast, or a pull request — it clearly can, fast. The constraint is how much human judgment it takes before anyone is willing to act on what it produced. Code has to be read, tested, and security-reviewed; a finance forecast or reconciliation has to be checked against source data before it goes into a board deck or a regulatory filing. That verification work doesn't disappear just because the first draft got faster.
CEOs see the gap too — and can't always measure it
A third data point, from EY-Parthenon's CEO Outlook Survey of 1,200 chief executives across 21 countries conducted in August and September, adds the view from the top: half of CEOs name AI as the single biggest driver of measurable productivity gains over the past year, but 23% say they're struggling to convert those gains into financial results, and only 16% say they have clear, real-time visibility into AI's return on investment. EY-Parthenon global vice chair Andrea Guerzoni framed the stakes directly: "AI is creating real productivity gains, but productivity alone is not a strategy."
None of this means adoption is stalling — the opposite. A separate KPMG survey of more than 1,000 finance leaders across 20 countries, also published this week, found active AI use in finance functions has jumped to 75% in 2026 from under a third in 2024, with 71% of leaders saying AI is meeting or exceeding ROI expectations. Enterprises aren't backing away from AI; they're running into a different problem than the one most of them planned for. The scarce resource was never going to be AI's output — it was always going to be the organization's capacity to trust that output quickly enough to act on it.
That has a quieter workforce implication than the layoff headlines of the past year suggest. Datarails found that 60% of finance organizations are redeploying staff toward higher-value work as AI absorbs routine tasks, while only 3% report cutting headcount directly because of AI. The people whose jobs are changing most are increasingly the ones doing the checking: reading the model's reasoning, spot-auditing its numbers, deciding when a second opinion is worth the delay. Three unrelated surveys, three functions, one week — all pointing to the same next problem enterprises will have to solve, which isn't getting AI to do more work, but building the verification capacity to keep up with it.