Zero Programmers, Yet Code is Being Written

There are no programmers in the company. Yet, code for business systems is generated daily, and workflows are automated.

This is not a science fiction story. It is happening in a small to medium-sized enterprise that has implemented AI coding agents.

In one company, a non-technical team adopted AI coding agents and achieved a 100% adoption rate. This means that everyone on the team is using AI to write code, despite having zero programming experience. Still, they have achieved a 96% task completion rate.

The essence of this story is not that “AI can write code.” It is that “dependency is structurally eliminated.”

The True Cost of Dependency

One of the most serious issues in small and medium-sized enterprises is dependency.

“Only Tanaka can do this task.” “This Excel macro was created by Suzuki, so it can’t be fixed without him.” “Only Sato, who left three years ago, knows how to configure this system.”

This is not a joke; it is a reality that many SMEs face daily. The costs are often invisible but are surely eroding management.

• If the responsible person is absent, operations come to a halt → Opportunity Loss
• If the responsible person leaves, know-how disappears → Reconstruction Costs
• Dependency on individuals leads to disadvantages in salary negotiations → Rigid Labor Costs
• Handovers can take months → Decreased Productivity

According to a survey, about 40% of operations in SMEs depend on specific individuals. Data shows that if that person suddenly disappears, it can take an average of 3 to 6 months to recover operations.

Dependency is not a “people problem”; it is a “system problem.” AI coding agents have the potential to fundamentally solve this system problem.

The Mechanism by Which AI Coding Agents Eliminate Dependency

In traditional programming, code relied on “the mind of the person who wrote it.” Naming conventions, logic structuring, and exception handling policies—all were influenced by individual styles. Therefore, if the person who wrote the code is no longer available, no one will touch it.

AI coding agents reverse this mechanism.

1. Code is generated from “natural language instructions”

“At the end of each month, transfer sales data from CSV to a spreadsheet, calculate the month-over-month change, and send the results via email”—this instruction in Japanese becomes code as is. There is no need to read the code. By reading the instructions, one can understand what is being done.

2. The same code is generated regardless of who gives the instruction

Whether Tanaka or Yamada gives the instruction, if the business logic is the same, the same code will be output. It does not depend on individual styles. This is “reproducibility.”

3. Modifications can also be made in natural language

If you say, “Add the year-over-year comparison as well as the month-over-month comparison,” the AI will modify the code. There is no need to ask the original person or search for specification documents.

These three elements combine to structurally eliminate dependency. Business logic remains not in “the minds of individuals” but as “natural language instructions.” Anyone can read and modify it.

Centralized Management of Over 50 Repositories with Governance File “Crag”

“But is it really okay for AI to write code on its own?”—This question is valid.

What is noteworthy here is the governance file method called “Crag.”

Crag is a system that manages multiple AI coding tools with a single governance file. Specifically, it defines rules such as:

• Style guidelines for code generated by AI
• Libraries and versions that may be used
• Security constraints (e.g., restrictions on external API calls)
• Conditions for automatic test execution
• Review process flow

By setting up this governance file, it has become possible to centrally manage over 50 repositories and operate AI tools with 96.4% accuracy.

This is not about “letting AI do whatever it wants” but rather “letting AI operate within rules.” Humans set the boundaries, and AI operates within them. This balance is key to safely utilizing AI coding even in non-technical teams.

The True Benefits for Small and Medium-Sized Enterprises

Let’s summarize the benefits of adopting AI coding agents in the context of small and medium-sized enterprises.

Cost Benefits

• Elimination of hiring costs for programmers (annual salary of 4 to 6 million yen)
• Significant reduction in outsourcing development costs (50,000 to 200,000 yen per project)
• Near-zero external costs incurred for maintenance and modifications

Speed Benefits

• Modifications that would take two weeks if outsourced can be completed in a few hours
• The time from “Can you create something like this?” to “It’s done” can be within a day

Organizational Benefits

• A business structure that does not depend on specific individuals can be established
• Since business logic remains in natural language, handover costs are drastically reduced
• Even new employees can understand and modify existing automation flows

Where to Start

1. List the “tasks that cannot function without this person” within the company.

Excel macros, internal tools, data aggregation—identify tasks that are dependent on individuals.

2. Write down what that task entails in Japanese.

“Every Monday, download data from System A, calculate the total in Column B, and email it to C.”—This Japanese instruction becomes the directive for the AI.

3. Try reproducing it with an AI coding agent.

There are several options like Claude, Cursor, and GitHub Copilot. Start with one task.

4. Set up minimum governance rules like those in Crag.

“Connecting to external APIs is prohibited” and “Generated code must always be reviewed by one person”—these two rules are sufficient to start.

Dependency, if left unchecked, becomes a risk for the organization. However, using AI coding agents can eliminate that risk through a structured approach. It is cheaper than hiring programmers and faster than outsourcing.

Transforming “This cannot function without this person” into “Anyone can handle it” is the true value of AI coding agents.