5 Years of Work in 2 Weeks. Cost: Approximately 1.8 Million Yen.

Asana utilized OpenAI’s Codex to refresh its legacy test code base in just two weeks. This was a task that traditional engineering estimates suggested would take five years. The cost incurred was approximately $12,000 (around 1.8 million yen).

When faced with this figure, do you simply say, “That’s amazing,” or do you wonder, “Can we do that too?” This is where the divergence occurs.

The key point is that what Asana accomplished was not merely “having AI write code.” It was about breaking down the tedious work that would take humans five years into a form that could be fully delegated to AI, creating an automated system. The migration of test code is a job that no one wants to do, but someone has to do it. It was the ideal form of “automatically completed.”

Now, let’s pose a question. Can a company with 10 employees and an IT budget of 50,000 yen per month reproduce this structure?

To conclude, it is not possible to do exactly what Asana did. However, the same structure—”delegating tedious repetitive tasks to AI to have them completed automatically”—can be reproduced starting today.

—

Breaking Down What Asana Actually Did

First, let’s clarify the facts.

Asana’s engineering team was tasked with migrating from a legacy test framework. They needed to rewrite thousands of files of test code into a new format. If done by humans, it would involve reading, understanding, rewriting, and verifying each file—an overwhelming task.

What Codex was tasked with was this “patternizable rewriting work.” There are three key points:

1. The work was patternizable. The conversion rules from “old format to new format” were clear.
2. Each task was independent. It could be processed file by file, with minimal dependencies.
3. Success or failure could be automated. Whether the tests passed could mechanically determine if the conversion was correct.

With these three conditions met, AI could automatically execute the loop of “delegate, process, verify, and move to the next.” Humans only needed to design the loop and handle exceptions.

Conversely, if these three conditions are met in a task, the same can be done in a company with 10 employees.

—

“Five Years of Tedious Work” in a Company of 10

While Asana’s case was about code, there are countless tasks in small and medium-sized enterprises that are “unwanted but necessary” to keep operations running.

• Inputting and reconciling invoices. Manually entering paper or PDF invoices into accounting software every month. 20 hours a month.
• Classifying inquiry emails and providing initial responses. Writing similar responses to similar questions. 15 hours a month.
• Formatting product data and registering it on e-commerce sites. Rearranging CSV items, resizing images, and writing descriptions. 30 minutes per product × 50 products a month = 25 hours a month.
• Aggregating daily and weekly reports and creating summaries. Copying and pasting reports from each member. 3 hours a week.

All of these tasks meet the earlier three conditions—patternizable, independently processable, and verifiable for success. In other words, they have a structure that can be fully delegated to AI.

—

What Can Be Done with a Monthly Budget of 50,000 Yen? — A Detailed Cost Calculation

When people hear “using AI,” they might envision expensive system development. However, as of 2025, the cost structure has dramatically changed.

Let’s take a closer look.

Total: 10,000 to 20,000 yen per month. With a budget of 50,000 yen, there’s plenty left over.

What to do with the remaining 30,000 yen? Use it for experimentation. Try out new APIs, compare different models, and learn from failures. Whether you have this “experiment budget” or not can determine whether automation takes root.

—

Four Steps to Reproduce

Step 1: Create a “List of Unwanted Tasks”

Compile a list of tasks that “no one wants to do but are done every month” within the company. The key is to prioritize based on psychological weight rather than work hours. Tasks that take “only 5 hours a month but are always dreaded” are where automation can have the greatest impact. Tasks that cause people to leave are costing more than just time.

Step 2: Filter Using the Three Conditions

Select tasks from the list based on Asana’s success criteria.

• Can it be patternized? (Is it almost the same procedure each time?)
• Can each task be processed independently? (Are there minimal dependencies?)
• Can success be mechanically verified? (Are the numbers correct, is the format right, etc.?)

If all three are “yes,” it can be automated immediately. If two, it can be done with some ingenuity. If one or none, it can be postponed.

Step 3: Automate at the Minimal Unit

Do not attempt to automate everything at once. Automate only one step of one task.

For example, in invoice processing, first automate just “extracting text from PDFs.” A human can verify the extraction results, and inputting into the accounting software can remain manual. This is sufficient.

Why? If you insist on full automation, you’ll get stuck on designing exception handling. Small and medium-sized enterprises do not have dedicated engineers for designing exception handling. Therefore, clarify the boundary between what AI does and what humans do, allowing AI to handle only the “patterned parts.” Humans will catch the exceptions. This clear division is the key to making automation work in a company of 10.

Step 4: Create a “Self-Running” System

Once the minimal unit of automation is successful, set triggers using workflow tools like n8n or Make (formerly Integromat). Triggers could be “when an email arrives,” “when a file is added to a folder,” or “every morning at 9 AM”—this will allow the system to operate automatically without a human pressing the start button.

Once you reach this point, you can arrive at the office in the morning and think, “Oh, it’s already done.” This mirrors the structure of Asana. The scale may differ, but the experience is the same.

—

Advantages of Small and Medium-Sized Enterprises

When large corporations attempt to implement AI automation, they face security reviews, approvals, vendor selection, PoCs, and production transitions—taking anywhere from six months to a year.

In a company of 10, you can start this afternoon.

The distance between decision-makers and implementers is nearly zero. The time from “Let’s try it” to “It’s working” can be as short as one day. This is a structural advantage that large corporations cannot replicate.

Another point is that the operations of small and medium-sized enterprises have fewer exceptions compared to large corporations. The complexity of patterns is vastly different between having 100 business partners and having 10. The simpler the patterns, the higher the accuracy of AI. In other words, small and medium-sized enterprises are better suited for AI automation.

—

“So, what should we do in the end?”

There is only one thing to do tomorrow.

Find the most “unwanted” repetitive task within the company and let AI handle just the first step of that task.

It can be ChatGPT or Claude. You don’t even need to hit the API. Start by manually entering prompts and checking the outputs. If you think, “Oh, this is usable,” then the next step is to automate it via the API. After that, set triggers using workflow tools.

The cost of automating one step is almost zero. All that’s needed is two hours of trial and error.

Asana completed five years of work for $12,000. Your company can take the first step for zero yen.

“Automatically completed” is not a privilege of large corporations.