There's a quiet problem sitting inside a lot of finance departments right now, and most leaders hadn't put a number on it until recently. New research from IDC, commissioned by Sage, did. The finding is uncomfortable: finance leaders are spending roughly 13 hours every week verifying the outputs of generic AI systems. That's not 13 hours of using AI. It's 13 hours of double-checking it.
If you run a finance team, that number probably lands somewhere between "not surprised" and "painfully accurate." You brought in tools that promised to compress the close, automate reconciliations, and free your people from grunt work. In a lot of ways, they delivered. But then a second cost showed up that nobody put in the business case. Someone still has to confirm the AI got it right. Someone has to be able to explain, when the auditor or the board asks, why the number is the number.
The productivity is real; the problem is how much of it survives contact with reality. The research found that about 26% of the time savings AI generates get clawed back by verification, explanation, and reconstruction work. So a quarter of your gains evaporate the moment your team has to trace assumptions, validate the logic, and rebuild the reasoning the system didn't show its work on.
This is the part that rarely makes it into the ROI slide. When a tool gets bought, the pitch is about hours saved. What doesn't get modeled is the new category of labor the tool creates on the back end. The work didn't disappear. It moved. It went from doing the analysis to defending the analysis, and that second job often lands on your most experienced, most expensive people, because they're the only ones who can actually tell whether the output holds up.
For a stretched team, this compounds fast. Every AI output that goes unchecked becomes a small pocket of risk. Every output that does get checked burns senior capacity that should be going toward forecasting, strategy, or partnering with the business. The 13-hour figure isn't just a productivity story. It's a capacity story, and capacity is the thing most finance teams are already short on.
The single most revealing data point in the study is this: 71% of finance leaders said they would veto an AI system that was 99% accurate if it couldn't produce a reasoning trace. Given a tool that is right almost all of the time, seven in ten would still shut it off if they couldn't see how it got there.
At first glance that looks irrational. Why turn off something that's almost always correct? But it's actually finance doing exactly what finance is supposed to do. In an audit-driven, compliance-bound function, an answer you can't defend is worthless no matter how correct it is. You can't take "the model said so" to an auditor. You can't put it in front of the board. You can't file it. If the reasoning isn't visible and traceable, the accuracy is beside the point.
This reframes the whole AI-in-finance debate. The bottleneck was never intelligence. The models are already good enough to be right most of the time. The bottleneck is trust, and trust in finance is built on explainability, judgment, and accountability. Those are human properties. A machine can generate a journal entry; it takes a person who understands ASC 606 to know whether that entry survives scrutiny. A model can flag an anomaly; it takes an experienced accountant to decide whether it's an error, a timing difference, or something that needs to be escalated.
The study also shows where autonomy is actually taking hold, and the pattern is instructive. Around 38% of organizations expect to deploy transactional AI agents within two years, but adoption is concentrated in narrow, reviewable, rule-based workflows: accounts payable, reconciliations, recurring close tasks.
In other words, AI is moving fastest exactly where the work is repetitive and the output is simple to verify. These are high-volume, low-ambiguity processes where a human can confirm correctness at a glance, so the verification tax stays low and the automation actually pays off.
The flip side matters just as much. The higher-stakes, judgment-heavy work – technical accounting decisions, revenue recognition calls, anything that requires interpretation rather than rule-following – stays firmly in human hands. That's the work that defines your finance function, and it's precisely the work AI can't yet be trusted to own. The map that emerges is clear: machines take the transactional layer, people keep the judgment layer, and the value of that judgment layer is climbing.
Put these three findings together and a strategy falls out of them. The winning setup over the next two years isn't "AI instead of people." It's capable people who can supervise, interpret, and stand behind AI-assisted work.
The 13-hour verification tax is really a talent problem in disguise. If your team is thin, or leaning on staff who can't confidently trace a system's logic, that burden compounds and automation stalls. But a strong, experienced team flips the equation. They verify faster, catch what matters, and let you push automation deeper precisely because someone competent is standing behind every output. As AI absorbs more of the transactional load, you need fewer people doing manual data entry and more people who can think, explain, and own the result.
That's the real headline buried in the data. Skilled finance professionals don't get less valuable as AI improves. They get more valuable, because they're the ones who make AI safe to rely on. The teams that scale automation successfully won't be the ones with the most tools. They'll be the ones with the judgment to use them well.
The catch is that this profile – technically strong, audit-fluent, able to supervise automated work – is expensive and slow to hire through traditional channels in the US market. Which is why a growing number of finance leaders are rethinking where that talent comes from. Building a bench of pre-vetted, globally sourced finance professionals lets you put the human judgment layer in place without the cost or the timeline of a conventional hire. In an environment where the constraint on AI is trust, the smartest investment is still the people who can be trusted to verify it.
Generic AI systems often don't show their reasoning, so finance professionals have to manually trace the logic, confirm data sources, and validate compliance before they can trust an output. In a function built on defensibility and audit trails, an answer without a visible reasoning path can't simply be accepted, which is why the IDC research found leaders spending around 13 hours a week on this work.
No. The productivity gains are real, and most of the time savings are retained. The point is that roughly a quarter gets consumed by verification, and the teams that minimize that loss are the ones with skilled people who can review AI work quickly and confidently rather than rebuilding it from scratch.
Adoption is strongest in narrow, rule-based, easily reviewable workflows such as accounts payable, reconciliations, and recurring close tasks. These are high-volume processes where outcomes are simple to check, which makes them a natural first place for finance autonomy to scale.
Because accuracy alone doesn't make an output defensible. The study found 71% of leaders would veto a highly accurate system that couldn't produce a reasoning trace, since finance requires the ability to explain and justify every number to auditors, boards, and regulators. Explainability, not just correctness, drives trust.
As AI takes over routine transactional work, demand shifts toward professionals who can supervise, interpret, and stand behind AI-assisted output. Building a team of experienced, audit-fluent finance talent becomes the way to capture AI's benefits while managing the verification burden, rather than treating automation as a replacement for skilled people.