Conclusion

Let’s get straight to the point: The cost of ‘hiring people’ is fundamentally changing.

Imagine your company needs a research officer, a document preparation officer, a data organization officer, and a report writing officer. Hiring four people would cost around 20 million yen annually. Even with temporary staff, it would be about 8 million yen. What if that could be replaced by AI agents costing only a few thousand to tens of thousands of yen per month?

Moonshot AI’s ‘Kimi Work’ poses exactly that question. It can run up to 300 sub-agents simultaneously, automatically handling tasks while accessing local files on the user’s desktop. By simply instructing it in natural language, like ‘Please do this,’ it breaks down tasks, divides the work, and completes it on its own — this is the envisioned world.

However, jumping on the bandwagon just because it sounds impressive is risky. Does multi-agent technology really work? Is it worth the cost? Should small and medium enterprises (SMEs) adopt it right away? Let’s take a calm look.

—

What’s New About Kimi Work — From ‘Beyond the Cloud’ to ‘Right on Your PC’

Traditional AI agents — such as ChatGPT plugins or AutoGPT — primarily operate in the cloud. This means they don’t directly interact with files on your PC. You have to upload them to Google Drive or connect via API, which adds extra steps.

Kimi Work aims to eliminate this bottleneck. By operating directly on your desktop, it can access local Excel files, PDFs, and internal documents seamlessly. Picture the reality in small and medium enterprises: most operational data is not organized in the cloud. Excel files scattered in shared folders, PDFs lying dormant on employees’ PCs. Whether you can directly access these files is a decisive factor in practicality.

And then there are the 300 sub-agents. Essentially, this means that ‘when you give one instruction, the AI autonomously organizes a team internally to process tasks in parallel.’ Research teams, analysis teams, and summarization teams can all operate simultaneously. In human organizational terms, a project team can be formed in an instant and disbanded in minutes.

—

But Does It Really Work? — Uncomfortable Research Findings on Multi-Agents

Here, we need to pause.

Multiple studies released in 2024 (e.g., Are More LLM Calls All You Need?) have raised doubts about the superiority of multi-agent systems (MAS). The main points of criticism include:

• Even if it costs ten times more, the accuracy can be equal to or less than that of a single-agent system (SAS).
• The more ‘interactions’ there are between agents, the more likely hallucinations (false outputs) are to propagate.
• For certain types of tasks, it may be faster and more accurate to have one excellent agent handle everything.

In other words, while ‘we can operate 300 agents’ is technically impactful, the question remains: how many tasks actually require 300 agents?

Considering the reality in SMEs, 80% of daily operations are tasks that ‘can be sufficiently handled by one AI in sequence.’ Creating estimates, summarizing meeting minutes, aggregating data — these tasks do not require 300 agents.

Conversely, multi-agents truly shine in tasks that involve ‘simultaneously researching and integrating a large number of information sources.’ For example, price surveys of 100 competitors, cross-referencing patent documents, and classifying and trend-analyzing dozens of customer feedback. Such ‘parallel tasks that take humans three days’ are precisely where multi-agents come into play.

In short, the idea that ‘multi-agents are for everything’ is incorrect; it’s essential to discern which tasks require multi-agents and which can be handled by a single agent.

—

A Closer Look at Cost Structures — The Break-Even Point for SMEs

While Kimi Work’s pricing structure is not fully disclosed at this time, we can estimate based on the cost of similar AI agent services.

• Running a GPT-4o-based agent for one task at full capacity incurs API costs of approximately $0.5 to $5 (about 75 to 750 yen).
• For large-scale tasks running 300 sub-agents simultaneously, the cost per instance could range from $50 to $150 (about 7,500 to 22,500 yen).
• If you run 20 such tasks a month, that totals to 150,000 to 450,000 yen.

On the other hand, what if the same work is done by a human team?

• Outsourcing research and analysis: 50,000 to 300,000 yen per project.
• Part-time administrative staff: 150,000 to 200,000 yen per month.
• Full-time employee: 350,000 to 500,000 yen per month (including social insurance).

For simple administrative tasks, there may still be cases where humans offer better cost performance. However, when considering the value of speed — ‘completing a three-day investigation in 30 minutes’ — the conversation changes. For SMEs, the speed of decision-making is directly linked to survival. The quality of decisions made in three days versus those made in 30 minutes is fundamentally different.

The break-even point is determined by ‘how much you value time.’ If the CEO spends three days on research, the opportunity cost of those three days is far more than just a few hundred thousand yen.

—

Memory Management — Why ‘AI That Remembers’ Resonates with SMEs

Another point to note is the evolution of agents’ memory capabilities.

Current AI chat systems essentially ‘forget once the conversation ends.’ Each time, you need to explain everything from scratch. This is the biggest stress in the workplace: ‘Do I have to explain my company from the beginning again?’

Recently proposed memory management frameworks like G-Long and MemRefine tackle this issue. The key is simple: structure and save past interactions and business contexts, allowing the retrieval of only the necessary information when needed.

What happens when this is implemented?

• The AI retains tacit knowledge such as ‘Our major clients are Company A, Company B, and Company C, where Company A prioritizes price, and Company B prioritizes delivery.’
• When assigning new tasks, it provides suggestions based on past contexts.
• Even if the person in charge changes, the AI functions as ‘the company’s memory.’

This is the very essence of eliminating dependency on individuals. It complements the biggest structural weakness of SMEs — ‘things don’t run without that person’ — with the AI’s memory. It could liberate businesses from the nightmare of operations collapsing due to the retirement of veteran employees.

However, it should be noted that many of these memory technologies are still in the research phase, and it needs to be verified how much of this is implemented in Kimi Work. There is always a gap between ‘it should work’ and ‘it works in practice.’

—

What SMEs Should Do Now — Three Specific Actions

‘It sounds impressive, but what should we actually do?’

Here are three concise recommendations for those who have read this far.

1. First, Identify Tasks That Can Be Handled by a ‘Single Agent’

Before jumping into multi-agent systems like Kimi Work, take stock of tasks that can be automated with single AIs like Claude, ChatGPT, or Gemini. Summarizing meeting minutes, drafting emails, organizing data — this can be started for 2,000 to 3,000 yen per month. This is where you can build the ‘muscle’ to delegate work to AI.

2. List Up ‘Tasks That Require Parallel Processing’

Count how many tasks there are that involve ‘parallel work that takes humans a full day or more,’ such as competitive research, market research, and classifying large volumes of documents. If there are more than five such tasks a month, it’s worth considering the introduction of multi-agents. If there’s only one a month, it may still be too early.

3. Start Documenting ‘Tacit Knowledge’ of the Company

In preparation for when memory management technology becomes practical, begin documenting your company’s operational rules, decision-making criteria, and characteristics of clients. This is the first step in eliminating dependency, regardless of whether AI is introduced. Before ‘teaching’ AI, humans must first ‘write it down.’

—

Organizational Charts Won’t ‘Melt’ but Will ‘Lighten’

What we see at the end of the maturation of tools like Kimi Work is not the extreme notion that ‘organizational charts will become unnecessary.’ Rather, organizational charts will ‘lighten.’

Work that used to require five people can now be handled by two people plus AI. The resources saved from three people can be redirected to tasks that only humans can do — building relationships with clients, making on-the-spot decisions, and seeding new business ideas.

Large companies, with their heavier structures, will take longer to make this transition. SMEs, being more agile, can move faster. A three-person company can leverage AI agents more swiftly than a large corporation with a department of 300.

This is not a story of ‘catching up to large enterprises.’ It’s a story of ‘being able to reach ahead precisely because you are a small enterprise.’

However, I must reiterate: ‘Implementing just because it sounds impressive’ is the worst decision. Examine your own operations and consider where AI can provide the most leverage. Start small. Measure the effects numerically. If it doesn’t fit, stop. Only through this cycle can truly usable AI applications emerge.

Technology won’t wait. But there’s no need to rush. Start in the right order and the right size. That is the essence of the AI strategy for small and medium enterprises.