Time saved by AI partially canceled out by time spent checking AI

The time saved by using AI is being eroded by the time spent verifying its outputs. 

This is according to recent data from accounting ERP provider Sage, which found that 48% of finance professionals across various industries spend 15 or more hours per week on verification activities. And those are the lucky ones, as 19% devote more than 30 hours a week to such activities. The poll also found that figuring out exactly how an AI model came to its conclusions so they can be explained to others also takes a lot of time: 26% of the respondents said this eats up more than a quarter of the expected productivity gains from using AI in the first place, while 22% said it takes up more than half. Sage called this effect "the verification tax." 

"AI adoption is introducing measurable operational overhead associated with validation, debugging, explanation recovery, repeatability testing, exception handling, traceability and governance review," said the report. 

Possibly as a result, organizations report implementing AI first in environments where outputs can be reviewed quickly, validated consistently, traced clearly and governed operationally. Overall, few intend to cut out human review entirely: Only 9% of the survey respondents intend to pursue broad, "lights-out" autonomy for transactional finance; in contrast, 29% said they planned to automate selected activities with defined scope and controls, reflecting a clear preference for bounded autonomy.

When presented with the hypothetical scenario of a fully reliable "continuous close" with a human-readable reasoning trace, only 23% of the respondents said they would shift their monthly close to a review-only model, compared to 42% who would adopt a hybrid close, 28% who would make no change at all and 7% who would shift to a reduced-frequency formal close (quarterly or annual). 

And because so many intend to keep the human at the forefront of their finance operations, the ability for someone to understand and explain what an AI is doing and why it is doing it has grown in importance for organizations. While AI is typically thought of as a black box, many leaders do not think this should be the case: 70% of finance leaders said they would  intentionally limit AI autonomy if they could not obtain real-time visibility into the agent's logic or a robust post-hoc audit trail, with 71% saying they would veto a 99%-accurate AI tool that could not produce a human-readable reasoning trace for every decision. And 71% agreed that the lack of AI transparency fundamentally undermines their ability to fulfill their fiduciary duty to the board. 

So suspicious are finance leaders of black box AI, 71% would veto a 99%-accurate AI tool that lacked a human-readable reasoning trace. Relatedly, 74% said "glass box" AI is either a mandatory requirement or a context-dependent requirement for high-risk financial decisions. In contrast, only 4% said accuracy of the final output is the only metric that matters. 

"The defining challenge of finance AI is no longer generating outputs," said the report. "It is operationalizing trust. Demand for the explainability layer is no longer aspirational. It is becoming a near-universal organizational requirement."

The Sage report is but the latest in a series of papers indicating that verification of AI outputs is becoming a major time sink that at least partially cancels out purported efficiency gains. The numbers may vary but recent research has been aligning on the observation that saving time also costs time. 

For instance, corporate accounting solutions provider Insightsoftware, in a recent poll of its own, found that while most respondents report time savings of between one to five hours per week, 20% say they frequently experience extra work in checking AI outputs, while 63% say it "sometimes" creates extra work. 

Another study from document solutions provider Foxit was even grimmer. It found that executives believe AI saves them an average of 4.6 hours per week, yet they spend four hours and 20 minutes per week validating AI-generated outputs. End users report saving 3.6 hours per week, but spend three hours and 50 minutes reviewing those outputs. The pattern holds across both markets. U.S. respondents experienced a net time loss of 10 minutes per week, while U.K. respondents saw a marginal gain of two minutes. 

But it's not just verification that is eating up time. Correcting, refining and redoing poor AI outputs — both one's own as well as those of coworkers — is also eating away at ostensible efficiencies. Workday, for example, concluded in its own research that productivity gains are not translating into better outcomes for most organizations, as an estimated 37% of the time saved by AI is offset by time spent correcting, clarifying or rewriting low-quality outputs. For every 10 hours of efficiency gained through AI, nearly four hours are lost to rework. For highly engaged employees who frequently use AI, Workday believes this translates into about 1.5 weeks per year devoted to correcting bad AI outputs. And so while 77% of employees say they are more productive due to AI use over the past 12 months, and 85% of employees believe they personally save between one and seven hours per week on their tasks, this does not necessarily translate into better overall performance.  

This points to a phenomenon colloquially called "workslop," which is defined as AI-generated work content that masquerades as good work, but lacks the substance to advance a given task meaningfully. Examples include reports that look polished and read well but make no sense, computer code missing vital context, or a slide deck that looks fine until you realize half the information is outright wrong.  Of 1,150 U.S.-based full-time employees across industries polled, 40% report having received such content in the previous month, according to research from BetterUp and the Stanford Social Media Lab in September 2025. 

And much of this comes from people who should have spent the time (perhaps 15 or more hours per week, as per Sage) verifying information and correcting errors themselves. In this respect, AI workslop pushes the verification burden onto someone else. This happens between peers, at 40%, but not entirely. The study found 18% was from direct reports to managers and 16% was from managers to their team members or even higher up in the chain than that. 

A recent MIT study found that verifying AI outputs is no longer just a compliance function but a core part of how value is created from AI. However, the study also said this may not be sustainable in the long run as AI systems are producing more outputs than any human being can reasonably check, and the difference between the two is growing. 

Further, the paper noted that verification skills depend on experience. As AI is beginning to take over more entry-level work, it is starting to erode the training ground through which workers build that experience, so it will limit the degree to which they can even properly evaluate what the AI is spitting out. This, in turn, will limit the value people can get from AI. 

"As AI systems are becoming more capable, it's getting harder to verify everything they produce," said the paper. "This will put a cap on how fully the benefits of AGI can be realized in the economy: AI makes it cheap to produce work, but not to judge whether that work is any good."

While one might think of using AI to check AI, the paper warned against this "tempting shortcut." While both systems share the same assumptions, they can reinforce the same errors, creating what the paper described as a false sense of confidence and not a real solution. 

Overall, the Sage report recommended that organizations prioritize reviewability over autonomy, optimize for time and trust, treat validation as a core operational workflow, prioritize bounded autonomy before full autonomy, and treat explainability as an economic multiplier.