Faster Work, Flat Profits: Why the Time AI Saved Isn't Hitting the Bottom Line

In 1930, economist John Maynard Keynes imagined a bold future in his short essay, "Economic Possibilities for our Grandchildren." He forecasted that if technological advancement and capital accumulation continued, the labor required to make a living could drop to just three hours a day, or a 15-hour workweek, a century later.

The year 2030 he envisioned is now just around the corner, yet a 15-hour workweek remains a distant dream. This is not because the technological advancements of the past century failed to boost productivity. On the contrary, labor productivity and living standards have improved to a level incomparable to Keynes’s era.

What missed the mark in Keynes’s prediction was not the premise that "technology will save time," but rather the assumption about where that saved time would go. When productivity increases, humans might choose to work fewer hours, but they might also choose to produce more goods and services within the same timeframe to enjoy a higher standard of living. Just because technology frees up time does not mean that time automatically turns into leisure.

The same applies to businesses today. The mere fact that AI reduces working hours does not automatically generate economic outcomes. What matters is where that saved time was redirected and what it produced there.

Saving Time is Not the Same as Saving Money

Recently, the Bank of Korea analyzed the effect of generative AI on working hours among domestic workers. The results showed that employees utilizing generative AI spent an average of 3.8% less time to complete the same tasks. Based on a 40-hour workweek, this equates to saving about 1.5 hours of processing time every week.

A figure of an hour and a half per week for a single employee might seem negligible. However, if a company has 10,000 employees, the story changes entirely. That creates 15,000 hours of free time every week, or roughly 780,000 hours annually. Multiply this by the average hourly wage, and it quickly snowballs into a figure worth tens of millions of dollars.

This is precisely where corporate ROI calculations for AI begin to go astray.

Suppose an employee earning a $100,000 salary uses AI to reduce their task processing time by 10%. This does not mean the company’s labor costs have decreased by $10,000. Unless the employee’s salary is cut or the headcount is reduced, the actual cash the company spends remains exactly the same.

What the company has gained here is not cash, but time. More accurately, it has secured surplus capacity that can be utilized for other purposes. Conflating the two leads to flashy reports claiming "AI saved $10 million," only to result in a bizarre situation where that $10 million is nowhere to be found on the income statement. Multiplying saved time by the average wage rate may indicate economic potential, but it is not a realized cost reduction in itself.

More importantly, time savings do not automatically translate into increased output. The Bank of Korea analysis revealed that the correlation between the time saved through generative AI and the actual increase in task throughput was virtually zero. AI accelerated the processing speed of individual tasks, but the time secured did not lead to additional output.

So, where does the time saved by AI actually flow? In a corporate context, we can categorize the destination of secured time into five distinct paths:

Actual Cost Elimination: This occurs when overtime pay decreases, previously outsourced work is handled internally, or departing employees do not need to be replaced. In these instances, actual cash outlays drop, translating time savings into tangible cost reductions on the ledger.
Capacity Increase and Additional Revenue: If a customer service center is backlogged with inquiries or a production line has pending orders, the same workforce can absorb more demand. In this scenario, revenue and contribution margins increase without additional hiring. However, this absolutely requires the existence of additional demand. If there are no incoming orders, simply increasing processing capacity by 1.5 times creates zero economic value.
Reallocation to High-Value Tasks: An employee who used to spend three hours writing a report finishes it in one hour using AI, and redirects the remaining two hours to prospecting new clients, solving complex problems, developing products, or strengthening internal controls. The labor cost invested remains the same, but the possibility arises for the same workforce to generate significantly higher value.
Securing Organizational Slack: Because work finishes early, employees can use the time to rest, assist colleagues, or learn new knowledge. While this "slack" is difficult to measure as immediate revenue or cost savings, it enhances organizational resilience and the ability to respond to unexpected demand. It can also reduce employee burnout and contribute to long-term skill development. However, this must not be calculated as an immediate labor cost reduction.
Evaporation (Missing in Action): The organization knows working hours have decreased, but has no idea what was done with the remaining time. Costs remain the same, output hasn't increased, and there is no evidence of new capabilities being accumulated. The time created by AI simply vanishes without converting into any economic value.

Will Employees Voluntarily Do More Valuable Work?

A recent study highlights how saved time can be reallocated in real-world business settings. A 2026 paper published in the Journal of Accounting Research by Devin Shanthikumar and Il Sun Yoo analyzed how AI investments at brokerage firms altered the way their analysts process and acquire information.

One of the core tasks for a sell-side analyst is to rapidly read and analyze massive volumes of corporate disclosures and financial data to update earnings forecasts. The researchers used the ratio of AI-related job postings out of total job postings as a proxy for a firm’s AI investment level. They then compared how analyst behavior differed between firms with high and low AI investments.

The analysis revealed that analysts at firms with high AI investment levels tended to update their earnings forecasts much faster following a company's 10-K filing. This relationship was particularly pronounced after the introduction of iXBRL, which enhanced the machine-readability of disclosures. This is consistent with the explanation that AI assists in extracting and organizing massive amounts of public information, thereby boosting analysts' processing speed.

However, the more fascinating aspect of this study is not the processing speed itself, but the behavioral changes that followed. Analysts at high AI-investment firms expanded their coverage of newly analyzed companies and industries. They participated more actively in earnings conference calls. Their earnings forecasts became more accurate and bolder—a difference that was especially stark in situations where predicting outcomes solely with public information was difficult, or where access to corporate insiders was high.

Interviews conducted by the researchers echoed these findings. Analysts explained that they use AI for information retrieval, extraction, summarization, and sentiment analysis to save time. They then spend that secured time conducting deeper research on covered companies, analyzing new firms, and expanding interactions with corporate insiders.

Of course, this study did not track every analyst's day minute-by-minute to directly measure that "one hour saved from disclosure analysis moved exactly to task X." It analyzed the relationship between a firm's AI investment level and subsequent analyst behavior, supplementing the mechanisms through interviews. Therefore, rather than concluding that AI directly caused these behavioral changes, it is more accurate to understand that the results are consistent with the theory that "the automation of public information processing via AI expands the capacity for additional, human-centric information acquisition."

Nevertheless, the direction this study points to is clear. AI does not stop at merely replacing what humans used to do; it can alter the very composition of the work a human performs in a day. It can reduce the proportion of repetitive tasks like reading and summarizing public documents, and increase the proportion of finding new targets for analysis, communicating with stakeholders, and interpreting additional information that machines struggle to acquire.

However, it would be dangerous to generalize this case into a naive optimism that "if we just hand employees AI, they will automatically find valuable things to do."

Sell-side analysts are in a profession where performance—such as forecast accuracy, timeliness, and the scope of covered industries—is relatively clearly measured. There is also a high probability that deeper research and the acquisition of differentiated information will directly lead to personal performance evaluations and competitiveness. The Bank of Korea survey also observed that groups with a strong link between performance and income, such as the self-employed, or those with high job autonomy, like professionals, tended to translate the time saved by AI into an actual increase in work output.

Conversely, in an organization where performance criteria are ambiguous, there are no rewards for extra effort, and employees lack the authority to redesign their own workflows, the incentive for saved time to migrate to more valuable tasks is weak. Technology can buy us time, but directing that time to places that lead to organizational performance and innovation is, ultimately, the role of management.

The 3 Stages of Value Creation

To manage the process by which AI investments yield actual payoffs, the economic value generated by AI must be divided into three distinct stages.

Stage 1: Release This is the stage where AI is utilized to reduce the time, wait times, and information processing costs previously dedicated to existing workflows, thereby securing new capacity. Evaluations like "Report writing speed increased by 30%" or "Document review time decreased by 5 hours a week" mostly fall into this stage. While this Release is a vital technical achievement, it is not yet an economic one. It simply means the company has gained available time; what that time will create has not yet been determined.

Stage 2: Reallocate This is the stage of deciding where to use the secured time. The organization must choose whether to reduce overtime and outsourcing, cut planned headcount additions, process backlogged demand, or shift resources to higher-priority activities like sales, innovation, analysis, or control. The true make-or-break point of an AI project often lies in this stage rather than the initial technology adoption. If you don't dictate where the saved time should go, it will simply dissipate into the gaps of existing workflows.

Stage 3: Realize This is the stage where reallocated time materializes into actual economic outcomes. It must be verified through results such as decreased overtime pay and outsourcing costs, incremental contribution margins from additional contracts, working capital recovery due to shortened processing cycles, loss prevention from reduced errors, or improved customer retention rates. Just because secured time was reallocated doesn't mean it all becomes valuable. Spending more time on sales activities doesn't guarantee more contracts, and spending more time on analysis doesn't always improve decision-making. Therefore, the link between the reallocated activity and the final economic outcome must be rigorously verified.

Applying the logic of these three stages completely changes a company's AI investment scorecard. If AI saved 100,000 hours annually, one should not simply multiply that by the average wage rate to calculate cost savings, but must ask the following:

Out of those 100,000 hours, how much transitioned into an actual cost reduction?
How much time was used to process additional demand?
How much time was migrated to higher-value tasks?
How much time was intentionally left for employee learning and recovery?
And how much time is still unaccounted for?

If a company cannot answer these questions, what it has measured is not the performance of AI, but merely the possibilities provided by AI.

Design the Destination of Saved Time Before Adopting AI

The actions management must take to bridge the gap between the time AI creates and actual economic outcomes are surprisingly straightforward. Rather than adopting AI and subsequently asking, "What did everyone do with their extra time?", the destination of the time to be saved must be determined before the project begins.

Suppose you are introducing an AI tool that reduces the proposal writing time for the sales team by 5 hours per employee per week. Simply stating "Proposal writing time cut by 50%" on the project approval document is insufficient. It must be accompanied by a reallocation plan like this:

"Out of the 5 hours saved, a minimum of 3 hours must be obligatorily allocated to additional face-to-face meetings with existing clients and prospecting for new clients."

And it cannot stop there. You must be able to verify whether client contact time actually increased, and track whether that additional contact led to an increase in consultation opportunities, proposal submissions, win rates, and contribution margins.

This does not mean every minute of the organization's time must be suffocatingly micromanaged. The saved time could be used for employee retraining or learning new technologies. It could be set aside as recovery time following highly intensive tasks, or converted into shorter work hours. If the organization consciously made these choices, that too is a perfectly rational management decision.

The key point is that not all time has to be immediately converted into revenue. The organization simply must know why it set time aside and where it moved it.

Calculating ROI by merely multiplying the average wage rate against secured time without deciding how to use it is entirely different management from proactively defining the destination of surplus capacity and tracking the results.

Therefore, the first question an executive should ask when receiving a report on an AI project is perhaps not "How many hours did we save?" After hearing that number, they must inevitably move to the next question:

"So, where did that saved time go?"

AI saving an employee's time is the achievement of technology. However, routing that time to places that yield money, growth, learning, and recovery is entirely the responsibility of management.

References

Seo, D., Oh, S., & Yoon, J. (2026). Does AI adoption increase productivity? Analyzing the effects of the first 3 years (BOK Issue Note No. 2026-12). Bank of Korea. https://www.bok.or.kr/portal/bbs/P0002353/view.do?nttId=10098322&menuNo=200433

Keynes, J. M. (1963). Economic possibilities for our grandchildren. In Essays in persuasion (pp. 358–373). W. W. Norton & Company. (Original work published 1930)

Shanthikumar, D., & Yoo, I. S. (2026). Beyond automation: AI and the human value of sell-side analysts. Journal of Accounting Research. Advance online publication. https://doi.org/10.1111/1475-679x.70083