AI・テクノロジー
Marubeni Cuts 1.2 Million Hours | The '4 Keys to Success' with Generative AI
機械翻訳 / Machine-translated
A
AI Friends
2026年4月27日・48分で読める
機械翻訳 / Machine-translated
@aifriends
AI Friends（https://aifriends.jp）のクロスポスト公式アカウント。AIツールの紹介・使い方・できることを、中学生でもわかるやさしい日本語で届けます。
Many Japanese companies struggle with the frustration of "We introduced AI, but our Excel workload hasn't actually decreased." Against that backdrop, the major trading company Marubeni has achieved an annual reduction of 1.2 million work hours through generative AI.
We explain in plain language — accessible even to a middle schooler — what the internal chatbot "Marcha" used by roughly 16,000 employees really is, how it was built by just one person in a single month, and how the company selectively uses ChatGPT and Claude for different purposes.
You're sure to find hints here for your own company's generative AI strategy.
Let us first organize what the April 2026 news reported from three angles.
On April 27, 2026, ITmedia Business ONLiNE published an article titled "Annual Reduction of 1.2 Million Hours: The 4 Reasons Marubeni's Generative AI Efforts Are Delivering Results."
The source was a seminar held at "Japan DX Week Spring 2026" from April 8–10, 2026.
The presenter was Hiroya Uenishi and colleagues from the Data Analytics Section of Marubeni Corporation's Digital Innovation Department.
The theme was "How Are 20,000 Marubeni Group Employees Making the Most of Generative AI?"
This was "a moment when a major Japanese trading company drew attention from companies nationwide as a model student in AI adoption." Trading companies tend to evoke images of "old-fashioned paper-and-stamp firms," but Marubeni had in fact been seriously committed to AI since April 2023.
The three years of effort culminated in the concrete figure of an annual 1.2-million-hour reduction — and that is the news.
Many people may struggle to grasp just how significant "1.2 million hours" really is. One office worker logs roughly 2,000 hours of work per year.
1,200,000 hours ÷ 2,000 hours = the equivalent of 600 employees' annual working time. In other words, Marubeni effectively generated the same output as hiring 600 new employees — through generative AI.
Think of it as "the same impact as a cafeteria worker serving 600 people's meals all at once."
Assuming 8 hours a day and 250 working days a year, it's like having 600 extra pairs of hands free. At an average annual salary of 5 million yen, that represents economic value equivalent to 3 billion yen in labor costs.
What's more, this figure grew in stages: 90,000 hours in 2024, 900,000 hours in 2025, and now over a million.
It is a prime example of steady accumulation transforming into an astronomical number.
The centerpiece of Marubeni's generative AI efforts is the company's internal-only chatbot "Marubeni Chatbot," affectionately nicknamed "Marcha."
In April 2023, shortly after GPT-4 launched, the beta version went live. Within just one year, 7,000 employees had signed up; today the user count stands at approximately 16,000.
The system has processed a cumulative 1.2 million queries and has been trained on 60,000 internal documents.
Think of it as "having one ChatGPT-expert new hire permanently stationed inside the company."
It is used as a work partner for everything — translating contracts, creating meeting minutes, searching internal regulations, answering questions about Excel formulas.
Even conservative departments that were initially in the "don't want to use it" camp had switched to "can't do without it" within six months.
It has become a model for Japanese companies as a success story that spread organically from the bottom up.
We explain each of the "4 reasons" mentioned in the article title one by one.
"Introducing DX and digital technology is nothing more than a means to an end. We put business impact first." — This statement by Hiroya Uenishi neatly captures Marubeni's philosophy.
Many companies treat AI adoption as the "goal" itself, but Marubeni treats it squarely as a "means."
The difference is like "choosing shoes that won't hurt your feet rather than buying trendy sneakers."
For example, when frontline staff complained that "translation work takes too long," the response wasn't "let's build a translation feature" but rather "let's first quantify how much time a translation tool would free up."
They rolled out the tool starting with tasks that showed results, and did not force it into areas where the impact was slim.
The goal is set on "changing work," not on "using AI."
This is the decisive difference from companies that lose their way after adopting AI.
Marubeni has a dedicated "Digital Innovation Department" to drive DX.
This department is an operational unit that handles AI, data analysis, and app development entirely in-house.
When GPT-4 appeared in 2023, development started from a casual conversation between two junior employees who wanted to "do something with generative AI."
As a result, the initial version of the Marubeni Chatbot was completed by essentially one developer in about one month.
Think of it as "a separate task force that skips the heavy decision-making of a large corporation and moves at startup speed."
Development that would have taken months and tens of millions of yen if outsourced was deployed at lightning speed using only internal talent.
And because it is in-house, when a user says "I want this feature," it can be reflected the following week.
"Being able to ask the developer directly" lowered the psychological barrier for employees.
Speed and flexibility are the in-house DX organization's greatest weapons.
Marubeni tackled head-on the psychological hurdle of employees who felt "AI is scary."
About a month after GPT-4 appeared, a ChatGPT study session was held, and since then beginner-oriented workshops have continued monthly.
The policy of "providing a place to learn the moment an employee gets curious" is consistently upheld.
Even more important is the company-wide sharing of the idea that "making mistakes is the same for people and for AI."
The thinking goes: "It's natural for a new hire not to be perfect — that's why veterans double-check. AI is the same."
The awareness that "even without 100% accuracy, significant work-hour savings are possible" has been internalized across the organization.
The direction has shifted from "AI makes mistakes, so we won't use it" to "because it makes mistakes, we build a habit of verifying."
As a result, a culture in which even cautious employees start using AI has taken root.
Breaking down the "mental wall" rather than the technical one is the most critical factor.
"Marcha" is characterized by a flexible design that can keep up with the rapid advances in the AI world.
In 2023 it relied on GPT-4 alone; Claude 3 appeared in 2024; Gemini 1.5 Pro and GPT-4o arrived in 2025 — LLMs (large language models, the AI engines behind ChatGPT and similar tools) change generations every six months.
Marubeni built its architecture from the start to "switch between multiple LLMs."
Today it selectively uses five models — Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, GPT-4o, and GPT-4 Turbo — assigning each to different tasks.
Think of it as "a chef who switches between a French knife, a Japanese knife, and a Chinese cleaver depending on the dish."
For example, Claude for translation, GPT-4o for summarization, a separate model for image generation — the best AI for each task can be chosen.
The database and UI are also designed to withstand frequent feature additions and specification changes.
Rather than building it once and being done, making it a continuously evolving system was the right call.
This is a real-world example showing that the "build it and forget it" approach no longer works in the AI era.
We look at the technology powering "Marcha" from three angles.
The backend of "Marcha" is Amazon Web Services (AWS)'s generative AI service "Amazon Bedrock."
Bedrock is like "the kitchen of generative AI" offered by Amazon — it lets you call multiple AI models through a common API.
It integrates with numerous AWS services: S3 (cloud storage), Lambda (serverless processing), DynamoDB (high-speed database), Transcribe (speech recognition), and more.
Think of it as "a system kitchen where the refrigerator, stove, and microwave are all integrated in one."
The vector database "Pinecone" is also incorporated, enabling high-speed search across 60,000 internal documents.
Because OS maintenance and server management are unnecessary, the development team can focus entirely on improving AI features.
Marubeni has explicitly cited "fully managed yet highly extensible" as the reason for choosing AWS.
The configuration achieves both startup-level development speed and enterprise-grade operational quality.
"Marcha" has three standout features.
The first is "File Chat" — upload Word, Excel, PowerPoint, or PDF files and get summaries or translations.
A Spanish contract or a Chinese quote can be summarized in Japanese in one go.
The second is "Speech Recognition Chat" — meeting audio is transcribed by Transcribe and automatically turned into meeting minutes.
Minutes for a one-hour meeting have gone from 30 minutes to 3 minutes to produce.
The third is "Custom Bots" — users can register their own department-specific FAQs and manuals to create a dedicated AI agent.
It's like "being able to add a 'reheat lunch box' button to your microwave yourself."
This freedom sparks creativity on the frontline and has been the catalyst for wider adoption.
Rather than ending as a single general-purpose tool, the ability for each department to customize it to their own specifications is a key strength.
The concrete work-hour reduction effect ranges from 25% to 65% depending on the task.
The greatest impact is seen in translation work — because Marubeni handles many overseas transactions, the reduction in translation tasks has been overwhelming.
Next come minute-taking, summarizing meeting materials, and searching internal regulations — broadly, any task involving written documents shows significant gains.
It's like "the difference between using an electric screwdriver versus a manual one on a factory floor — assembly speed differs by a factor of five."
On the other hand, for creative planning and interpersonal negotiation, the effect is limited.
Identifying "where to apply it" determines the success of company-wide rollout.
This is a realistic approach that abandons the illusion that "AI will make all tasks more efficient."
The result is the enormous figure of 1.2 million hours annually.
We look at "what the other major trading companies are doing" across three dimensions.
In contrast to Marubeni, Mitsubishi Corporation emphasizes collaboration with external partners.
It builds its generative AI foundation through partnerships with NTT Data, Microsoft, and others.
The policy is "team up with specialists rather than build it yourself."
The contrast is like "the difference between eating out and cooking at home."
The advantage is the ability to incorporate cutting-edge technology quickly; the disadvantage is less freedom to customize.
Because Mitsubishi Corporation is larger, each business sector pursues separate AI strategies.
It does not have a symbolic company-wide tool like "Marcha."
It is a representative example of a "decentralized AI strategy" — the polar opposite of Marubeni's approach.
Mitsui & Co. takes a strategy of developing AI solutions specialized by industry.
It partners with AI companies in individual fields — healthcare, energy, food, and so on — to deploy tailored solutions.
"Marubeni, covered by one general-purpose tool" vs. "Mitsui & Co., optimized industry by industry."
Think of it as "the difference between an all-you-can-eat conveyor-belt sushi restaurant and a high-end omakase sushi counter."
The advantage is deep industry expertise; the disadvantage is that company-wide efficiency improvements tend to lag somewhat.
It is a strategy uniquely suited to a diversified trading company with a wide range of business areas.
Neither approach is definitively correct — the right choice depends on corporate culture and operational characteristics.
The reality is that Japanese companies' AI strategies come in a wide variety.
Itochu Corporation has taken the bold approach of deploying OpenAI's ChatGPT Enterprise across the entire company at once.
Full-scale rollout began in 2024, with the official paid version from US-based OpenAI distributed to all employees.
The advantage is fast deployment; the disadvantage is that monthly costs can run high.
It's a choice akin to "the generous approach of equipping every employee with a premium laptop."
This is the exact opposite strategy from Marubeni's in-house "Marcha" — "buy and use" rather than "build."
However, since Marubeni also uses OpenAI and Anthropic APIs under the hood, the two are not in complete technical opposition.
"In-house platform + external APIs" vs. "off-the-shelf tool company-wide rollout" are the two main options among Japan's large enterprises.
Rather than asking which is superior, each company should choose based on its own culture and organizational capabilities.
We consider "what will spread across Japanese companies as a whole" from three angles.
Marubeni's case is giving other large Japanese companies the courage to think "maybe we can do this too."
It is an example that overturns the images of "large corporations = slow-moving" and "Japanese companies = bad at AI adoption."
After ITmedia's report, inquiries from banks, manufacturers, insurers, and others wanting to "learn from Marubeni" surged.
If a company of 16,000 employees can roll it out company-wide, a mid-sized company of 3,000 can adapt the approach.
It's like "when a quiet senior colleague wins a major award, the junior staff think 'I can do that too' and step up."
From 2026 onward, AI democratization among Japan's large companies is expected to accelerate sharply.
"In-house AI platform development" is becoming a new axis of competition for Japanese companies.
The gap between companies that can keep up and those that cannot will directly affect business performance.
Marubeni's "4 reasons" are a success pattern that mid-sized and small companies can apply directly.
Business-impact-first + dedicated team + monthly study sessions + flexible operation — these four points can be practiced at any company.
However, since "one person building it in one month" is unrealistic for a small company, a practical starting point would be services like ChatGPT Team, Claude Pro, or Gemini Business, available for a few thousand yen per month.
The difference is like "a bespoke suit for a large corporation vs. a smaller company wearing off-the-rack clothes well."
Even with off-the-shelf tools like ChatGPT Team or Microsoft 365 Copilot, the Marubeni philosophy can be practiced on a smaller scale.
What matters is not "which tool" but "how it is operated."
2026 is the inaugural year of AI adoption for small and mid-sized companies, and Marubeni is the model.
It is a textbook case from which companies of any size can learn.
Companies wanting to build "in-house AI development teams" like Marubeni's are rapidly increasing.
As of April 2026, approximately 30% of Japan's large companies are considering in-house development of generative AI platforms (Ministry of Economy, Trade and Industry survey).
As a result, job postings for engineers who can work with AWS Bedrock, Azure OpenAI, and generative AI prompt design have tripled year-on-year.
Salary ranges are also rising — at major Tokyo firms, starting salaries of 7 million yen for new graduates are no longer uncommon.
It mirrors the structure where "when a new sport becomes popular, demand for both players and coaches rises at the same time."
Conversely, for those with AI engineering skills, this is an outstanding career opportunity.
The years 2026–2028 are the golden era for "in-house AI engineers."
This is a turning point at which Japan's entire labor market is being restructured.
Kenta, a fourth-year employee at a mid-sized trading company in Tokyo, was drowning every day in translating contracts with overseas clients.
"One English contract eats up my entire morning; the afternoon goes to meeting minutes, and evenings to document organization — leaving on time is a dream."
One day, the company introduced an internal chatbot inspired by Marubeni.
Upload a PDF of a contract and get a Japanese summary in 30 seconds, with risk clauses automatically extracted.
He experienced "the shock of a task he had done every day for over three years being replaced by the press of a button."
Translation time shrank from 4 hours to 30 minutes; the time freed up let him focus on creating proposal materials for new clients.
Within six months, his sales ranking in the department jumped from 3rd to 1st.
He felt firsthand that this is "an era where a junior who can use AI outperforms a senior who cannot."
A prime example of junior employees thriving.
Yumiko, who has worked in accounting at a mid-sized manufacturer in Aichi Prefecture for 20 years, used to spend three days every month end processing invoices.
"Checking 500 invoices one by one, entering data in Excel, double-checking for errors… stiff shoulders and eye strain became chronic."
The company introduced generative AI combined with OCR (technology that reads text from paper).
Now, simply scanning an invoice auto-populates the data, and AI summarizes the content and flags anomalies.
It was the experience of "a robot perfectly replicating the skill of a seasoned expert."
Month-end processing shrank from three days to half a day, and overtime was eliminated.
With the time freed up, she began producing management analysis reports, which the company president praised as "the most useful report ever."
Proof that veteran employees benefit the most from AI.
A real-world example of "not having your job taken by AI, but having AI raise the quality of your work."
Makoto, who works in the HR department of a major financial institution, was overwhelmed handling inquiries from 10,000 employees.
"'How do I take paid leave?' 'What are the expense reimbursement rules?' 'What are the conditions for the housing allowance?' — 100 inquiries a day."
Inspired by the Marubeni case, he built a custom bot trained on 60,000 internal regulation files.
Employees can now ask questions directly to the chatbot without going through an HR staff member and get an instant answer.
The experience is like "a librarian who instantly flips to the right page in the company handbook whenever you're stuck."
Inquiry volume dropped dramatically from 100 per day to 10, and the HR team's working hours were cut in half.
With the time freed up, he focused on planning employee engagement initiatives, and the turnover rate declined.
HR work shifted from "administrative processing" to "strategic planning."
A successful example of DX in a back-office department.
A. "Marcha" itself is exclusive to the Marubeni Group, but an external version has been launched as the sister service "I-DIGIO next-AI Chatbot."
It is operated by Marubeni I-DIGIO Holdings and has been available to corporations since 2024.
Based on the features proven at Marubeni, it can be customized to each company's needs.
Think of it as "the cafeteria's popular dish now being available at a restaurant outside."
Target customers are mid-sized to large companies, with monthly fees ranging from approximately 100,000 to 1,000,000 yen depending on scale.
"Being able to borrow the knowledge of a major trading company" is the differentiating point from other AI services.
It stands as an option distinct from ChatGPT or Microsoft 365 Copilot.
Its strength lies in being optimized for business practices unique to Japanese companies.
A. Scaled down appropriately, the "Marubeni Method" is fully applicable even for small and mid-sized companies.
For a company of 100 people, a realistic starting point is ChatGPT Team ($25/person/month) or Microsoft 365 Copilot ($30/person/month).
What matters is emulating the spirit of the "4 reasons": prioritize business impact, designate a specialist, hold monthly study sessions, operate flexibly.
Think of it as "a local youth baseball team adopting the training methods of a professional team."
Even with off-the-shelf tools, applying Marubeni's operational approach will yield solid results.
Conversely, even with premium tools, poor operation will produce less impact than Marubeni achieves.
"It's the way you use the tool, not the tool itself" is the essence of AI adoption for Japanese companies.
Small and mid-sized companies especially should learn from Marubeni's philosophy.
A. Based on Marubeni's experience, there are no cases of "people whose jobs were taken by AI" — instead, there are more and more people "whose work quality has improved through AI."
Those whose translation work was replaced by AI have shifted to higher-level negotiation and strategic planning.
Secretaries freed from minute-taking now concentrate on supporting executive planning.
It mirrors the structure where "when calculators became widespread, the number of accountants doing mental arithmetic decreased, but the number doing financial analysis increased."
However, the reality is that "those who cannot delegate work to AI" will be at a competitive disadvantage.
2026 is a turning point where the gap between "those who can use AI and those who cannot" is widening.
Rather than cultivating "skills that AI cannot replace," focus on developing "skills to master AI."
In practice, every Marubeni employee has already set their course in that direction.
A. Marubeni achieves high security through the combination of "in-house development + AWS Bedrock."
All data is under Marubeni's control and is designed so that it is not sent to external ChatGPT servers.
When using OpenAI or Anthropic APIs, corporate contracts include a commitment that the data will not be used for training.
The difference is like "using your own company's dedicated ATM rather than a public one."
Using the free version of ChatGPT for business purposes is risky because of the potential for confidential information leakage.
With enterprise-tier plans — ChatGPT Team or above, Claude Pro or above, Microsoft 365 Copilot — data is not used for training.
By 2026, "no personal-plan usage for business purposes" has become standard practice for corporate AI use.
Setting up a safe operating environment is a prerequisite for AI adoption.
A. The standard approach is to list three tasks in your work that are time-consuming, then try starting with the one most likely to show results.
Examples: translation, minute-taking, handling internal inquiries — tasks that are "routine but time-consuming."
Start by trying the free versions of ChatGPT or Claude to experience the impact, then move to paid plans or dedicated tools once you see results.
It's like "when taking up a new sport, trying it out before rushing to buy expensive equipment."
Starting small and building a success story is the biggest key to company-wide rollout.
Trying to roll it out to the entire company at once leads to failure — even at Marubeni, it started with two junior employees chatting.
"Take a small first step, a bold second step" is the iron rule for AI adoption at Japanese companies.
If you don't try it within 2026, you risk falling two years behind your competitors.
Marubeni's case overturns the image that "large Japanese companies are bad at adopting AI." In April 2026, the concrete achievement of reducing 1.2 million work hours annually gave major momentum to AI democratization in Japan.
Business impact first, in-house DX organization, ongoing education, flexible architecture — these four reasons are a universal success pattern applicable at any company.
The essence is not "buy ChatGPT and be done with it," but rather "building a culture of operating AI in your own company's way."
"Try asking AI to handle the most time-consuming task in your work today" — that small first step is the first move that will guide your company toward the Marubeni Method.
This article is a cross-post from AI Friends.
