Conclusion First: “Writing Fast” and “Finishing Fast” Are Two Different Things

With AI, code can be written 100 times faster. This is no exaggeration. A study by GitHub Copilot reported that developers’ task completion speeds improved by up to 55%. Google DeepMind’s AlphaCode ranks in the top 54% in competitive programming. When looking solely at the speed of “code generation,” the areas where humans can compete are rapidly diminishing.

So, I want to ask: Has development in your company become faster?

Garry Tan, CEO of YCombinator (a prominent U.S. accelerator), recently pointed out, “The amount of code written by AI has exploded, but the quality of the products shipped has not increased proportionately.” A survey by GitClear presents an even harsher reality. In 2023, repositories after the introduction of AI assistance saw an increase in “code changes,” while the proportion of “code movement and deletion”—essentially, the rate of revisions—also surged. Writing and revising, revising and writing. It’s unclear whether this has made the process faster or slower.

Even though code can be written 100 times faster, results do not increase 100 times. Let’s delve into the nature of this structural gap from the perspective of small and medium-sized enterprises (SMEs).

—

“Tokenmaxxing”—The Counterproductive Effects of Mass Generation

Recently, the term “Tokenmaxxing” has been spreading among developers. This refers to a development style that pushes for quantity by making AI churn out a massive amount of tokens (units of code).

What does this lead to? Let’s look at a specific example.

At a contract development site, AI was tasked with writing API connection code. In just 10 seconds, it produced 200 lines of code—an amount that would take a human 2 to 3 hours to write. However, the code contained missing error handling, unnecessary library calls, and security issues. As a result, it took 4 hours for review and correction. Writing it manually might have been faster.

According to GitClear’s data, projects after AI implementation saw a significant year-on-year increase in “added lines of code,” while the “refactoring ratio” also showed an upward trend. In other words, generation is fast, but the increase in revisions offsets that speed.

This is particularly painful for SMEs. Large companies may have dedicated code review teams, but a development team of 5 or 10 lacks the human resources to scrutinize code generated by AI. Consequently, they fall into a binary choice: either “use the code written by AI as is” or “write it manually without AI.” Neither is an optimal solution.

—

The Speed of Code and the Speed of Products Are Different

I want to discuss another fundamental issue.

In software development, the proportion of time spent on “writing code” is actually not that large. A study by Microsoft Research found that developers spend only about 30-40% of their working hours writing code. What occupies the rest? Requirements definition, design, review, testing, debugging, and communication. In other words, most of the time is spent on “deciding what to create” and “verifying if what has been created is correct.”

AI has made a portion of this 30-40% faster. However, the overall bottleneck may not lie there.

To illustrate, consider this analogy: the most time-consuming aspects of cooking are “deciding what to make,” “going to buy ingredients,” and “plating the dish.” Even if the speed of “cutting with a knife” increases 100 times, dinner won’t be served 100 times faster.

What truly becomes a bottleneck in the development field of SMEs? In many cases, it is:

• Uncertainty about “what to create” (ambiguity in requirements)
• A lack of people who can assess the quality of what has been created (absence of a review system)
• Frequent specification changes (slow decision-making)

Even if AI code generation is introduced here, the only speed increase will be in “writing.” If the preceding and subsequent processes are bottlenecked, the overall throughput will not change. This embodies the principle of TOC (Theory of Constraints) that states, “improving anything other than the bottleneck will not speed up the overall process.”

—

Design and Architecture—Areas Where AI Cannot Step In

Another often-overlooked issue is that AI generates “code,” not “design.”

In this context, design refers to aspects such as the structure of the database, the design philosophy of the API, the logic of screen transitions, and the user experience during errors—essentially, the parts that determine “how it should function for the user.”

AI can produce code when instructed, but it still cannot adequately answer questions like “What is the optimal data structure for this business flow?” or “What does this user want to do next?”

A case study from a company developing SaaS for local manufacturing is illustrative. They quickly built a management screen using AI. While it looked decent, it was not designed with the end-users—workers using tablets—in mind. The buttons were too small, there were too many input steps, and the error messages were technical. As a result, the field workers returned it with complaints of “it’s hard to use.” It took two weeks to redo what had been created in three days with AI. This is the reality.

The ability to create quickly and the ability to create correctly are entirely different skills.

—

So, What Should SMEs Do?

Vague statements like “cooperation between AI and humans is important” will not bring about change. Here are three specific actions that can be taken starting today.

1. Narrow Down the Processes Where AI is Used

Trying to delegate everything to AI leads to failure. AI excels at “pattern-based routine code.” CRUD (Create, Read, Update, Delete) processes, validation, and templates for test code are areas where AI should be concentrated.

Conversely, the core parts of business logic, security-related aspects, and data design should be handled by humans. Just clarifying this division can significantly reduce the effort required for revisions.

One contract development company reported that after implementing this division, the rate of revisions for AI-generated code dropped from about 60% to 15%. Narrowing the scope of what to delegate to AI ultimately speeds up the overall process. Paradoxically, this is the reality on the ground.

2. Cultivate “Reviewers”

What SMEs lack the most is personnel who can assess the quality of the code produced by AI. In the AI era, the ability to read and judge code is becoming more valuable than the ability to write it.

This may sound like a high-level skill, but it can actually be checklist-based. “Are there any security holes?” “Are there unnecessary dependency libraries?” “Is the error handling appropriate?” “Are naming conventions consistent?” Even just these four items can significantly improve the quality of reviews.

By using affordable AI code review tools (like CodeRabbit), the human review burden can be further reduced. For a team of five, an investment of 10,000 to 20,000 yen per month could potentially halve the revision workload.

3. Spend Time on “What to Create”

With AI compressing the time spent on writing code, where should the saved time be allocated? The answer is “requirements definition” and “user hearings.”

The strength of SMEs lies in their proximity to customers. Unlike large corporations, which must navigate through multiple layers of organization, SMEs can hear the voices of users directly. By leveraging this strength to enhance the accuracy of “what should be created,” rework can be reduced. When rework decreases, the speed of AI can directly translate into results.

Instead of using the time saved by AI to “write more code,” use it to “define the right things.” This is the only way to convert speed into results.

—

Understanding the Structure Provides Opportunities for SMEs

Finally, I want to share an encouraging thought.

The “gap between speed and results” becomes more severe in larger companies. Why? Because as organizations grow, processes other than coding—approvals, adjustments, meetings—become bloated. Even if AI speeds up code generation, if approvals take two weeks, it becomes meaningless.

On the other hand, SMEs can make decisions quickly. When the CEO says, “Let’s go with this,” they can act the next day. They already possess an organizational structure that can leverage AI’s speed.

In other words, the ones who can maximize the benefits of AI are actually SMEs. However, there are conditions. They must be able to correctly determine “what to create” and possess the ability to assess AI’s output. If these two conditions are met, while large corporations are stuck in meetings, SMEs can ship products.

AI has brought the cost of “writing” close to zero. Therefore, the value of “what to write” has skyrocketed. SMEs that recognize this structural change will begin to win.