AIで生産性2倍。じゃあAIが止まったら？

Using AI can double productivity. This is a common narrative we hear these days. In reality, the benefits of AI assistance are evident in various tasks such as code generation, content creation, and data analysis.

However, one question arises: What happens to a person’s performance the moment they can no longer use AI?

The answer is not “it goes back to normal.” It actually becomes worse than before. This is a structure revealed by several recent studies. For small and medium-sized enterprises, this is not just a theoretical concern.

—

数字で見る「依存の罠」

ケース1：AIチューターで学んだ学生、一人では解けない

A research team from Carnegie Mellon University (Bastani et al., 2024) conducted an experiment involving around 1,000 high school students in Turkey. They compared a group that used a GPT-4-based AI tutor for math practice problems with a group that did not.

• While using the AI tutor: The accuracy rate significantly improved.
• Final exam without the AI tutor: The performance of the group that had used AI was no different from that of the group that had not used it. In some cases, it was even lower.

In other words, when AI is available, performance is fast. But without AI, learning does not stick. The speed gained did not translate into personal growth.

ケース2：AIでアイデアの質は上がるが、全員同じアイデアになる

A joint study by Harvard Business School and Wharton School (Doshi & Hauser, 2024) examined the impact of AI assistance on creative tasks.

• Ideas generated with AI assistance had higher average quality.
• However, looking at the group as a whole, the diversity of ideas decreased by up to 40%.
• Everyone was drawn to the AI’s “optimal solution,” resulting in similar outputs being produced.

What does this mean? While individuals using AI may feel they are doing “good work,” when viewed from an organizational perspective, the source of differentiation disappears. For small and medium-sized enterprises, uniqueness is vital. In a world where everyone consults ChatGPT and produces the same proposals, where can one find a competitive edge?

ケース3：AI支援後に独力作業させると、パフォーマンスが「元以下」に

Research by Microsoft Research and others (Chopra et al., 2024) measured the performance of developers using AI coding assistance tools when they worked without the tools.

• During AI assistance, task completion speed improved by an average of 56%.
• After removing AI assistance, the tasks showed higher error rates compared to the group that did not use AI.

Those who became faster had skipped the process of thinking for themselves. As a result, when returning to an environment without AI, their judgment had deteriorated compared to their previous selves. This is the essence of the “dependency trap.”

—

中小企業にとって、これがなぜ致命的か

For large companies, there are alternative measures if AI stops working. They have specialized teams and manuals. But what about a company with only 10 or 20 employees?

If one person who can effectively use AI leaves, operations can come to a halt.

This is a matter of dependency, but it differs in nature from traditional dependency. Previously, it was the “intuition and experience of veterans” that created dependency. Now, it is the “ability to use AI tools and to judge AI outputs” that becomes personalized. Moreover, this ability becomes obsolete every six months due to tool updates.

隠れたコストを計算してみる

When small and medium-sized enterprises adopt AI tools, there are visible and invisible costs.

Assuming a company with monthly sales of 5 million yen, if an AI-dependent employee leaves and operations are disrupted for a month, the impact could be several million yen. If the annual cost of AI tool usage is 600,000 yen, the “savings” that were supposed to be realized could vanish in an instant.

—

じゃあ、どうすればいいのか

This is not a call to “stop using AI.” The cost-saving effects of AI are real. The issue lies in the design of its usage.

1. 「AI付き」と「AIなし」の業務を意図的に分ける

Do not base all operations on AI. Set aside one day a week, or designate specific tasks as “AI-free days.” It’s like strength training; without time to remove the training wheels, your muscles for independent thinking will weaken.

There are examples of this in practice. At a local production company, there is a rule that the first draft of proposals must always be handwritten. AI is used for finishing later. This separates the “thinking process” from the “finishing process.”

2. AIの出力を「判断する力」を仕組みで担保する

The most dangerous practice for small and medium-sized enterprises is using AI outputs as they are. Sending a quote generated by ChatGPT directly to a client or deploying code written by Copilot without review.

The solution is simple: incorporate a rule into the workflow that mandates a human check on AI outputs. This is not just a double-check; it formally defines the process of questioning AI outputs as part of the business operations.

3. 属人化を防ぐ「AIの使い方マニュアル」を残す

The way AI tools are used, including how to write prompts, continues to evolve. Therefore, create a system to record “how AI is currently being used in this operation” once a month for five minutes. This can be done in Notion or a spreadsheet. Even if an employee leaves, the next person can maintain the same quality from the next day.

4. 「AIで浮いた時間」の使い道を決める

This is often overlooked. If AI reduces work time by half, what will you do with the extra time? If this is not decided, the extra time will dissipate aimlessly or lead to deeper dependency on AI.

The extra time should be allocated to dialogue with customers, on-site observations, and prototyping new ideas—in other words, to “human work” that AI cannot perform. This will become the competitive edge for small and medium-sized enterprises.

—

AIは「道具」だが、道具は人を変える

The proliferation of calculators has led to an increase in people who struggle with mental arithmetic. The widespread use of GPS has resulted in people who no longer remember routes. The same phenomenon is about to occur on a broader scale and at a faster pace.

AI indeed dramatically reduces costs. We are entering a world where operations that used to cost 3 million yen can now be done for 50,000 yen. However, to reap these benefits, it is essential to have “humans who can make judgments even without AI” within the team.

The strength of small and medium-sized enterprises lies in their quick decision-making, proximity to the field, and collective understanding of the situation. If this strength becomes hollow due to “AI dependency,” they will relinquish their only competitive advantage against large corporations.

When implementing AI, it is crucial to simultaneously design a system that maintains “operational capability without AI.” This may sound contradictory, but it is the most cost-effective risk hedge.

First, I encourage you to try just one thing next week. Ask someone on your team, “Can you try doing this task without AI today?” Observe what happens during that time. This will be the first step in measuring your company’s “dependency level.”