Morning Arrival: It Was Already Done

“Who completed that aggregation I requested yesterday?”
“No one. The AI finished it overnight.”

This is not science fiction. In 2025, it is already happening in some workplaces.

The issue of personalization in small and medium-sized enterprises (SMEs). For years, companies have been saying, “Let’s create manuals” and “Let’s systematize,” yet they still rely on veteran employees like Tanaka-san to keep things running—there are countless such companies.

However, what truly kills personalization is not ‘systems.’ It is the experience of ‘finishing on its own.’ Systems require human intervention. But ‘finishing on its own’ operates even without people. This difference is significant.

The evolution of AI agents is making this ‘finishing on its own’ a reality. Let’s take a closer look at what has changed.

—

26 Minutes vs. 33 Seconds—What This Difference Means

A joint study by Harvard University and Perplexity (2025) produced some interesting numbers.

• Average autonomous work time per session for AI agents: 26 minutes
• Average time using traditional search engines: 33 seconds

About 47 times longer. This is not just a matter of being “faster.” It represents a qualitative change where ‘substantial work can be completed all at once without human intervention.’

In 33 seconds, you can at best “search and return results.” With 26 minutes, you can gather information, compare it, organize it, and even draft a report.

Consider the daily operations of SMEs. Every morning, spending 30 minutes on sales aggregation, competitor checks, and compiling daily reports. Imagine an AI agent completing this in 26 minutes overnight, with results arriving in Slack by morning.

30 minutes a day × 20 business days a month = 10 hours of liberation per month. That’s 120 hours a year. For an employee earning 2,000 yen per hour, that’s 240,000 yen annually. But the real value is not in cost reduction. It is that the employee can focus on ‘the work they should be doing.’

—

AI That ‘Chooses Its Own Tools’

Now, onto the main topic. The most impactful evolution of AI agents is the concept of dynamic tool selection.

Traditionally, using AI required humans to instruct, “Use this tool to process this data.” In other words, while AI is an excellent worker, the planning is done by humans.

The AutoTool framework (introduced in 2025) changed this. AI agents can assess the task at hand and choose the most suitable tools on their own.

Here’s how it works in practice:

• You instruct, “Compile last month’s sales data.”
• The agent decides for itself: “Since this is a CSV aggregation, let’s use Python,” “I’ll create the graph with matplotlib,” “I’ll output the report in Markdown.”
• Humans don’t even need to know about the tools used.

What’s beneficial for SMEs here is that “it can operate even without someone knowledgeable in IT.”

In large companies, the IT department handles tool selection and setup. SMEs don’t have that luxury. The reality is often that the president learns how to use ChatGPT from YouTube and tries to spread it in the workplace.

Dynamic tool selection has the potential to eliminate the handicap of not having IT-savvy personnel. AI can autonomously choose tools, combine them, and process tasks. Humans only need to communicate what they want done.

—

An AI That Remembers, “We Did This Before”

Another subtle yet powerful evolution is agent memory.

With the MemToolAgent framework, AI agents can remember past task execution experiences and apply them to future tasks.

What changes?

• First time: “Create a monthly report” → The agent creates it through trial and error (30 minutes)
• Second time: “Create a monthly report” → The agent remembers the previous steps and creates it without hesitation (10 minutes)
• From the third time onward: Further optimization and increased accuracy

This mirrors human ‘expertise.’ The difference is that this ‘expertise’ is not tied to an individual.

The operational know-how that Tanaka-san acquired over ten years disappears when he leaves. But the memory of the AI agent remains as an asset of the company. It can be passed on to new agents.

The essence of personalization is that “experience and judgment reside only in an individual’s mind.” Agent memory structurally resolves this issue.

—

So, How Much Will It Cost?

Discussing technology alone is not enough. What SME owners want to know is, “So, how much will it cost, and when can we start using it?”

To be honest, as of June 2025, there are still few complete service offerings that can fully implement the aforementioned “dynamic tool selection + memory functions + workflow automation.” Many frameworks are still in the research stage and not yet ready for immediate deployment.

However, partial implementation can start today.

| Task | Tools Used | Estimated Monthly Cost |
|—|—|—|
| Daily sales aggregation and report auto-generation | ChatGPT API + Google Sheets + Zapier | 5,000–15,000 yen |
| Automatic classification and drafting of inquiry emails | Claude API + Gmail integration | 3,000–10,000 yen |
| Monitoring competitor pricing and social media trends | AI agents (like Dify) + scraping | 5,000–20,000 yen |
| Automatic summarization of meeting minutes and task extraction | Whisper API + GPT-4o | 2,000–8,000 yen |

Monthly costs range from 10,000 to 50,000 yen. Annually, that’s 120,000 to 600,000 yen—less than the cost of one part-time employee.

Whether this is seen as “cheap” or “still expensive” depends on how much labor costs are currently incurred for those tasks. If you can automate a task that costs 40,000 yen a month for just 10,000 yen, you’ll break even in three months.

—

Start with “Finishing One Task on Its Own”

Trying to automate everything at once will lead to failure. I can say that with certainty.

What you need to do is simple.

1. Choose one task that is done daily and yields the same result no matter who does it.

Sales aggregation, inventory checks, compiling daily reports, sending standard emails—these repetitive tasks that require no judgment are the best starting points.

2. Assign it to an AI agent and observe for one week.

The accuracy doesn’t need to be 100% from the start. 80% is fine. Humans can check and correct the remaining 20%. This feedback will improve the agent’s accuracy. If there’s a memory function, this learning will accumulate.

3. Share the experience of ‘finishing on its own’ within the company.

This is the most important part. It’s crucial for not just the president but also the employees on the ground to experience the realization that “the aggregation was finished by the time I arrived in the morning.” This experience will foster motivation for further automation.

Rather than top-down orders of “We will implement AI,” it’s 100 times more effective to let the staff experience the surprise of, ‘Wait, it’s already done?’ just once.

—

Why SMEs Stand to Gain the Most from ‘Finishing on Its Own’

Large companies can absorb personalization through ‘organization.’ If Tanaka-san is absent, Sato-san can back him up. They have manuals, training systems, and handover periods.

SMEs lack these resources. If Tanaka-san falls ill, the operations come to a halt.

That’s why the value of “automation that doesn’t rely on people” is overwhelmingly higher for SMEs.

The evolution of AI agents is not just for large corporations. Rather, it is evolving to directly address the weaknesses of SMEs—those that lack personnel, have knowledge that is personalized, and don’t have IT-savvy individuals.

Dynamic tool selection resolves the issue of “not having IT personnel.” Agent memory addresses the problem of “know-how being tied to individuals.” Autonomous work solves the issue of “not having enough people.”

All of these are the very challenges faced by SMEs.

—

What to Watch for in the Future

Three key areas to keep an eye on:

1. Evolution of Inter-Agent Collaboration
We are moving from a stage where one agent handles one task to a stage where multiple agents collaborate to complete a series of workflows. A world where “order receipt → inventory confirmation → ordering → customer notification” flows entirely automatically will become visible in late 2025 to 2026.

2. Further Cost Reductions
API usage fees are dropping every quarter. The token price for GPT-4o is about half of what it was a year ago. If this trend continues, we are on the verge of an era where practical agents can operate for just a few thousand yen a month.

3. No-Code and Low-Code Agent Building Tools
Tools like Dify, n8n, and LangFlow are rapidly maturing, allowing AI agents to be built without writing code. The excuse of “It’s impossible because we don’t have programmers” will soon be obsolete in just six months.

—

Finally

Personalization will not die with ‘systems.’ Even if manuals are created, someone still needs to read them. Even with training, someone needs to remember.

But if it is ‘finishing on its own,’ then there’s no need for someone to read or remember.

AI agents are not about “hiring excellent employees” for SMEs; they are a technology that “eliminates the work itself.”

Start with just one. Tomorrow morning, try creating a task that ‘finishes on its own.’