Conclusion

Let’s get straight to the point: “An AI that can completely handle computer tasks has become a reality.”

OpenAI’s GPT-6 Astra is not just an AI that answers questions in a chat. It opens browsers, inputs numbers into spreadsheets, and clicks buttons in business software. In other words, the tasks that humans used to perform with a mouse and keyboard are now done by AI.

Previous AI focused on “writing text” and “summarizing.” Astra is different. It observes the computer screen and operates it. It is also distinct from RPA (Robotic Process Automation), which required humans to write out all the steps, such as “click this button on this screen.” With Astra, you can simply say, “Process this invoice,” and it will assess the screen, make decisions, and proceed with the operation.

What does this mean? For a company with 10 employees, the work of one administrative staff member could potentially disappear entirely. At the same time, there is a risk of mistakenly sending the wrong invoice to a client due to operational errors.

Let’s take a closer look at both the expectations and risks.

—

What Can Be Automated?—An Estimate of “1,200 Hours” Disappearing

In a company with 10 employees, where can Astra be effective? Let’s estimate in three areas of business.

1. Invoice Processing: From 600 Hours Annually to Almost Zero

Assuming a company processes 100 invoices a month, with each taking 30 minutes, that amounts to 50 hours a month and 600 hours a year. This translates to approximately 1.2 million yen in labor costs (calculated at 2,000 yen per hour).

Astra can read PDF invoices, input amounts, dates, and client names into accounting software, and bring them to a state of awaiting approval. The only task left for humans is the final check. The processing time per invoice shrinks to 2-3 minutes. Annual hours drop from 600 to 50. That’s a reduction of 550 hours, or about 1.1 million yen.

However, this is contingent on the AI functioning “accurately.” The risks will be discussed later.

2. Inventory Ordering: AI Can Handle Decision-Making

Inventory management can be divided into “counting items” and “deciding when and how many to order.” RPA could only automate the former. Astra can handle the latter as well. By analyzing past sales data, seasonal fluctuations, and lead times, it can determine that “50 units should be ordered by next week” and operate the ordering screen accordingly.

If the 20 hours spent monthly on inventory management can be halved, that results in an annual reduction of 120 hours, equating to about 240,000 yen. While the numbers may seem modest, the essential value lies in preventing critical issues for small businesses, such as “missing sales due to stockouts” or “cash flow problems from excess inventory.”

3. Email Sorting and Standard Replies: 30 Hours a Month Disappear

Incoming emails can be automatically classified into categories such as “quotation requests,” “complaints,” “sales emails,” and “internal communications,” with standard replies drafted as well. This saves 30 hours a month, totaling 360 hours a year, or about 720,000 yen.

In total, these three tasks alone could free up approximately 1,030 hours annually, amounting to around 2.06 million yen in labor costs. In a company of 10 employees, that means one person could be assigned to entirely different work—such as sales, developing new products, or meeting clients—which would be far more beneficial for revenue.

—

“It’s Already Done”—Existing Achievements

Several initial case studies of Astra’s implementation have been released.

• Legora (Legal Tech): Completed 41 contract reviews in just a few minutes. The error detection accuracy was equal to or better than that of humans, with processing speed improved by about 40%. Previously, this work took a lawyer half a day.
• Playco (Game Development): Reduced manual corrections in the prototype production process by 50%. Designers shifted from “fixing tasks” to “creative tasks.”

Both cases share the commonality that “while the AI was working, humans were engaged in other activities.” This is the essence of automation. The best automation is when you realize the task is completed without your involvement.

In the context of small businesses, imagine arriving at the office in the morning to find yesterday’s sales reports already compiled, all invoices entered into the accounting software, and purchase orders lined up awaiting approval. Such a world could be attainable for just a few tens of thousands of yen in monthly API fees.

—

Now, the Main Issue: What Are the Dangers?

If you implement it based solely on expectations, accidents are guaranteed to occur. Let’s outline three specific risks.

Risk 1: Operational Errors—AI Runs “Incorrectly Until the End”

Astra makes judgments based on what it sees on the screen. However, if the screen layout changes, a popup appears, or the internet is slow and loading is interrupted—these are things that a human would notice instantly, but the AI may malfunction.

It could input an invoice amount incorrectly by one digit, order ten times the required quantity, or forward an email from Client A to Client B. What a human might consider a “slip-up” could result in the AI making “systematic errors across the board.” If all 100 invoices are off by one digit, it could lead to trust issues.

Countermeasure: Final approvals must always be conducted by a human. In batch processing, include sample checks. Design the system as “95% automated + 5% human check” rather than fully automated.

Risk 2: Security—Data Leaks

Astra operates on OpenAI’s cloud. It cannot run on your own servers. This means that invoice amounts, client names, customer lists, and inventory data—all pass through OpenAI’s servers.

Large companies have security departments to review this. Small businesses typically do not have such structures. Therefore, it is essential to establish minimum guidelines for what should be done.

Countermeasure: 1) Clearly define the scope of data to be handled (e.g., excluding personal information). 2) Review the API’s terms of use and choose a plan that does not use data for training. 3) Identify the potential impact in case of a data leak in advance.

“Handing everything over because it’s convenient” is the worst decision you can make.

Risk 3: Dependency—”If the AI Stops, Work Stops”

If you rely entirely on Astra for your operations, the moment the AI stops due to a malfunction, all work comes to a complete halt. There have been instances in 2024 where OpenAI’s services experienced outages lasting several hours.

What if a company with 10 employees depended on Astra for invoice processing, inventory ordering, and email sorting? A half-day outage could completely halt operations.

Countermeasure: Always maintain manual backup procedures. Ensure that “a minimum operational state can be maintained without AI.” While it is correct to eliminate dependency on individuals, simply replacing it with “AI dependency” is meaningless.

—

Why Small Businesses Will Win by Being the First to Engage

Large companies take at least six months for security reviews, internal approvals, PoCs, and full implementation. Small businesses can start testing next week. This is the key point for reversal.

The API usage fees for Astra are currently estimated to be in the same price range as GPT-4o. If you can automate the work of one administrative staff member for a monthly fee of just a few tens of thousands of yen, the ROI is clear.

However, it is essential to start with “tasks where mistakes have minimal impact” rather than trying to automate everything. Begin with internal sales aggregation, email sorting, and organizing meeting minutes. Test these areas first to assess accuracy and reliability.

For tasks like invoicing and ordering, where mistakes could inconvenience clients, it’s best to wait until accuracy is confirmed.

—

So, What Should You Do?

1. What to Do This Week: Experiment with a demo of Astra. List three “tedious tasks that are repeated every month” in your company.
2. What to Do This Month: Try Astra on one low-risk task. Record accuracy and time taken.
3. What Not to Do: Do not attempt to automate all tasks at once. Avoid removing human checks. Do not hand over sensitive data without restrictions.

GPT-6 Astra is not about “AI taking jobs away.” It’s about drastically reducing the time spent “sitting at a computer and moving your hands” and allowing more time for “decisions and relationship-building that only humans can do.”

For a company with 10 employees, the impact of freeing up 1,000 hours annually is incomparable to that of a large corporation. The productivity per employee will visibly change.

Start by experimenting. You can make decisions from there.