Values in the AI Era: "Why Isn't It Cheaper Even Though AI Was Used?" The Pitfall of Unseen Efficiency in Invoices

"This document, AI can create it quickly, right?"

When such expectations are placed, there are circumstances that are difficult to explain for those creating it. Indeed, the time until text appears on the screen has shortened. However, the time to read that text, investigate the basis, correct mistakes, and decide whether it is okay to hand it over has not disappeared.

The client feels they are paying for something "created in seconds." The provider believes they are being paid for a "job finished to a state where it can be responsibly handed over."

This gap in perception leads to distrust regarding AI-driven efficiency.

The essay by lawyer and writer Stephen Embry, published in the U.S. legal media "Above the Law," delves into this issue from the perspective of legal work.

The question is clear. Are we so focused on the time it takes for AI to output text that we overlook the work humans spend afterward?

"From 5 hours to 6 minutes"—what part of the work does this number represent?

Embry presented a hypothetical comparison for creating a document. It is not an average value obtained from investigating actual law firms.

Traditionally, it took 4 hours for drafting and 1 hour for review, totaling 5 hours.

With AI, drafting can be done in 0.1 hours, or 6 minutes. However, it requires 2 hours for review and correction.

If only drafting is considered, time is reduced by 97.5%. However, the total process until completion is a 58% reduction.

Even a 58% reduction is a significant improvement. Yet, it differs from the impression that "most of the work disappears."

If the starting and ending points of the comparison are not aligned, the effect of AI can easily appear larger. If the human side includes everything from research to final review, while the AI side only measures text generation, it is not comparing the same work.

What should be measured is the time until a product of the same quality is in a usable state.

Even if time is reduced by 58%, the amount decreases by only about 9.7%

Another easily overlooked point is that the cost of time changes depending on who performs the work.

The original article assumes the hourly rate for a junior lawyer is $250, and for a partner lawyer, it's $800.

Traditionally, a junior would spend 4 hours drafting, and a partner would spend 1 hour reviewing. After AI adoption, the drafting time is reduced to 0.1 hours, while the partner's review time increases to 2 hours.

The reduction is $175. In percentage terms, it's about 9.7%.

Here, much of the effect of reduced time is offset by the high-cost review work.

However, it is not appropriate to read this number directly as "the actual cost of a law firm" or "profit." The hourly rate used for billing and the internal labor cost calculated from salaries are different figures. The table also does not include AI usage fees or training costs.

What this hypothetical example teaches is not that AI will always only reduce costs to this extent.

The reduction rate of time, the reduction rate of fees paid by customers, and the reduction rate of the provider's cost need to be measured separately.

It's not just incorrect text that requires verification

When thinking of AI verification work, it's easy to imagine finding non-existent precedents or incorrect numbers.

However, there is a more fundamental issue. Even if the text is correct, it is not known to be correct before reading it.

Even if there are no errors in the document, work remains to ensure it aligns with the case's premise, that no necessary conditions are missing, and that it meets the client's objectives.

In a contribution referenced by Embry from Legal IT Insider, Harshit Garg, founder of legal AI company Lexi, also focuses on the time spent re-reading a correct draft.

Garg points out the fragmentation of records. AI has records of generation and usage fees, while the office has records of the staff's work hours. However, if the verification time is not linked to which AI output it corresponds to, it's difficult to grasp the total cost involved in a single deliverable.

The fact that an AI invoice is cheap and the cost of completing work using AI is cheap are not the same.

On social media, "the work has shifted" and "it should be cheaper" intersect

This issue is also reflected in discussions on social media.

What is introduced here is not a direct response to the original article but related posts published about AI work hours and legal service fees. The qualifications and practical experience of the posters have not been independently verified, nor is it a survey representing the entire industry's opinion.

In Reddit's business-oriented AI community, there was a post stating that since starting to use AI, time has been spent on verifying, adjusting, and selecting the output instead of writing text.

The poster expresses a sense not of "the work has disappeared," but rather a feeling that "the focus of effort has shifted."

This is something that anyone who has experienced pressing the button to generate text would understand. The initial output is quick. However, while correcting details, it can become unclear how much correction is needed to complete it.

Meanwhile, in Reddit's community dealing with electronic discovery, there were differing perspectives on the efficiency and price reduction of legal work.

One comment suggested that when including AI tool usage fees, additional implementation into existing systems, and team training time, the cost savings are currently offset.

Another comment questioned whether, in work billed by time, if the necessary time decreases, the amount paid by the customer should also decrease, and if prices are not dropping, whether it is truly becoming more efficient.

This conflict is not simply about being for or against AI.

The provider sees the burden from introduction to operation. The customer sees the invoice they receive. As long as both parties are looking at different numbers, both are likely to feel that "the other side doesn't understand."

Will "experts are verifying" alone be enough to convince?

There is a reason for the explanation that verification takes time. However, that alone may not resolve doubts about fees.

If the pre-advertisement emphasizes "overwhelming time savings" and at the billing stage says "expert verification is necessary," it may sound like a convenient explanation to the customer.

What is needed is not to appeal to the value of verification abstractly but to specifically show what was verified.

For example, in a contract, there are checks for consistency with transaction terms, omissions in responsibility sharing, and alignment with negotiation policies. In research materials, there are checks for the source of numbers, the target period, and comparison conditions.

Where AI shortened time, where human time was used, and what was improved as a result. With that explanation, customers can more easily judge the fees.

At the same time, verification is also necessary for the provider. The mere fact that verification time is long does not mean all of it was necessary and efficient work.

Is AI being used for tasks it is not good at? Are instructions and materials insufficient? Are the same verifications being repeated among staff? It is important not to wrap operational issues in the phrase "expert judgment."

In Japanese workplaces, it may appear as waiting for a superior's confirmation

This issue is not unique to Japanese law firms. It broadly applies to work where deliverables are checked by someone before being released, such as proposals, sales materials, public relations texts, translations, and programs.

For example, suppose a person in charge can quickly create a proposal using AI. However, if the number of managers reviewing it does not change and only the number of submitted materials increases, the waiting time for approval may become longer.

The time for the creation process has decreased. Still, the number of days until it reaches the customer remains unchanged. In some cases, only the overtime of managers increases.

This is a scenario that can be considered when applying the original article's points to Japanese workplaces, and it does not represent a confirmed reality across all Japanese companies.

Nevertheless, it is a useful perspective when considering the evaluation of AI introduction. Simply adding up the time savings for each person in charge does not tell whether the overall workflow has sped up.

Along with "time to create materials," it is necessary to look at "time waiting for confirmation," "time dealing with rejections," and "days until final approval."

Effect measurement should use "one usable deliverable" as the unit

So, what should be recorded?

If we apply the original article's question to practice, it is easier to think of changing the unit from "how many times AI was used" to "whether one usable deliverable was completed."

It is also necessary to align the difficulty and quality standards of the work being compared as much as possible. If AI is used for simple tasks and difficult tasks are processed in the traditional way, the effect of AI might be overestimated.

Conversely, if only the unfamiliar period right after introduction is measured, the potential for improvement might be overlooked. Separating the initial costs for learning and setup from the recurring costs in normal operation makes judgment easier.

The important thing is not to align measurements with either the expectation that "AI should be useful" or the impression that "verification is so difficult that it's meaningless."

Whose benefit will the freed-up time be turned into?

Even if AI truly reduces time, that alone does not reduce cash expenditures.

Even if the work of salaried employees is reduced by an hour, their salary for that month does not automatically decrease. The effects may appear in other forms, such as reduced overtime, handling additional cases, shortened delivery times, or improved quality.

What becomes necessary here is deciding how to use the freed-up time.

Will it be used for price reductions for customers? To increase the number of cases that can be handled? To spend time on careful consultation and proposals? To lighten the burden on workers?

The results to be observed change depending on which option is chosen.

If the goal is price reduction but only the number of uses is tracked, it is unclear whether the customer benefited. If the goal is to reduce the burden but only the number of cases processed is increased, workers may find it hard to feel the benefits of efficiency.

What do we want to improve before and after AI introduction? Without that agreement, even if numbers