Arriving at the Office in the Morning to Find Everything Done

“I just gave instructions last night, and by morning, the estimate was ready.”

This is not a story from the future. In 2025, it is already happening.

The era of AI agents that “finish work on their own” is here. And it costs just 50,000 yen a month. Tasks that once cost 3 million yen when outsourced or 5 million yen a year when hiring employees are now being handled by automation tools for just 50,000 yen a month.

The question is not “What can AI do?” but “What costs disappear, and what value increases due to AI?” This structural change is more disruptively positive for small and medium enterprises than for large corporations.

Why? Large companies have numerous “unchangeable systems.” Small and medium enterprises do not. The time has finally come when agility can be a weapon.

—

Development Speed Improved by 28-38%. However, Teams That Don’t Set Up Properly See a 53% Increase in Debt

First, let’s talk about coding agents.

A recent large-scale survey of 441 repositories found that teams that implemented coding agents saw their development speed improve by 28% to 38%. The number of commits increased, and the speed of processing pull requests also rose.

Up to this point, it sounds impressive. However, there is another crucial piece of data in this survey.

Teams that did not set up AI properly experienced an increase of about 53% in cognitive complexity of their code.

In other words, teams that thought, “Let’s just try it out,” became faster but ended up with a mountain of unreadable code. Technical debt accumulates. This can be fatal for small and medium enterprises, as they lack the engineering resources to repay that debt.

So, what should be done?

The answer is simple: Take time for the initial setup. Specifically, this means creating prompts for coding standards, clarifying review criteria, and specifying output formats. Just by doing this, a “runaway AI” can be transformed into a “reproducible system.” The tacit knowledge of veteran employees that was previously personalized can be converted into AI prompts. This is the true “DX” in small and medium enterprises. It may not sound glamorous, but it works.

With Cursor Business or GitHub Copilot Enterprise at 50,000 yen a month, this level of automation is entirely achievable. That’s 600,000 yen a year—cheaper than hiring one part-time employee. Moreover, AI doesn’t quit, doesn’t take breaks, and doesn’t require handovers.

—

MCP Tools—Protocols for AI Agents to Connect with the “Outside World”

Next, let’s focus on MCP (Model Context Protocol).

For those wondering, “What is MCP?” here’s a quick explanation: It’s a common standard that allows AI agents to “use external tools and databases on their own.”

Until now, AI could only “answer questions when asked.” What MCP has changed is that AI can now open Google Drive, read spreadsheets, post to Slack, and update databases on its own. In other words, AI now has “limbs.”

In an experimental study called MCPAgentBench, multiple AI agents were given MCP tools to measure their completion rates for complex tasks. The results showed that it’s not the “number” of tools but the “selection” of tools that makes a difference in outcomes.

The practical implications for small and medium enterprises are as follows:

• The first automation should focus on “repetitive tasks that don’t require judgment.” Transcribing invoices, aggregating daily reports, updating inventory data. Just incorporating MCP tools here can eliminate 20 to 40 hours of labor per month.
• The next automation should focus on “tasks that require judgment but can be patterned.” Rough estimates, initial classification of inquiries, screening of job applications.
• Finally, the tasks that should remain with humans are “building relationships” and “handling exceptions.” Inserting AI here can actually increase costs.

Getting this order wrong can lead to implementation failure. The key point for small and medium enterprises in utilizing AI is not “What to have AI do?” but “What to not have AI do?”

—

Agent Harness and Framework—Focusing on “What to Delegate” Rather Than “What to Choose”

Now, let’s discuss some technical aspects, but at a level that business owners should understand.

To operate AI agents, there are two concepts: “Agent Harness” and “Agent Framework.”

• Agent Harness is the execution environment that actually runs the AI model as an “agent.” It manages conversation state, calls tools, and handles error processing. In car terms, it’s the “chassis and engine.”
• Agent Framework is the set of components used to assemble agents. Examples include LangChain and CrewAI. In car terms, it’s the “parts catalog.”

The key point for small and medium enterprises is simple: “Don’t build it yourself. Buy something that works.”

Building from scratch using a framework is something for companies with five engineers. Small and medium enterprises should choose services that are already complete as harnesses. Dify, n8n, Make (formerly Integromat). These can be used to build AI agents with no-code for a monthly fee of a few thousand to 50,000 yen.

The selection criteria are straightforward:

1. Does it stop when there’s an error, or does it go haywire? Choose tools that will stop.
2. Does it keep logs? Tools that don’t allow tracking of what was done will only create an AI version of personalization.
3. Can it connect with existing tools? kintone, freee, Chatwork. The ability to connect with what your company is currently using should be the top priority.

—

The True Conditions for Eliminating Personalization

Having read this far, you might be wondering, “So, how exactly can personalization be eliminated?”

Here’s the answer:

Personalization disappears not when AI is introduced, but when “what humans did is documented into prompts and workflows.”

The task of verbalizing the “this is how we do it in such cases” that resides in the minds of veteran employees to feed it to AI is the very act of eliminating personalization. AI is merely a catalyst that accelerates this process.

Conversely, if you introduce AI tools without doing this verbalization, nothing will change. 99% of companies that say, “We implemented AI but it didn’t work” skip this step.

The specific steps are as follows:

1. Record the work of veterans for a week. Screen recording is fine. Loom is free.
2. Feed that recording to AI to automatically generate a procedure manual. Claude or GPT-4o will suffice.
3. Verify whether newcomers can reproduce the process based on the manual. If they can’t, revise the manual.
4. Transfer the procedures that can be reproduced, starting with those that require no judgment, to AI agents.

These four steps require no special technology. All that’s needed is the decision to do it.

—

Small and Medium Enterprises Can Win

Finally, there’s one last thing I want to convey.

The era of AI agents is not a time for small and medium enterprises to “catch up.” It is a time to “surpass.”

Large corporations are bound by existing systems, approval flows, and vendor contracts. It takes three months for them to approve the introduction of one AI agent and six months for security reviews. In that time, small and medium enterprises can conduct ten experiments.

50,000 yen a month. 600,000 yen a year. The question is whether this investment can create a state where the routine tasks of three employees are “finished by the time you arrive at the office in the morning.” This is the management decision for 2025.

Let’s conclude the technical discussion here.

Tomorrow morning, I encourage you to ask yourself one question: “What would happen if I delegated this task to AI?” The answer will likely come back faster than you think.