Japanese Companies Are Using AI—but Few Have Rebuilt Work Around It

Imagine the first morning back after Obon. An employee asks generative AI to summarize the notes from a meeting. A job that once took 30 minutes takes three. Then the summary is pasted into a spreadsheet, copied into a different template, sent up the same approval chain and stored in the same folder as before. One task became faster. The work around it barely moved.

This is not a reported scene from one named company. It is a way to understand the pattern visible in “DX Trends 2026,” released in July by Japan’s Information-technology Promotion Agency, or IPA. AI has reached the individual desk. It has not yet consistently redrawn the blueprint of work: who decides, which data are authoritative, where a human checks the machine, what can be removed, and how the result reaches a customer.

Two revolutions hiding inside one adoption rate

The sentence “we adopted AI” can describe two very different revolutions.

The first is a tool revolution. AI summarizes and translates text, drafts emails, searches for information and writes code. Distribution is easy. An employee can save time on the same day an approved tool becomes available. That is where Japanese business is moving quickly. In IPA’s survey, the leading use case was summarizing, translating or proofreading text and audio, at 82.5%. Creating documents and reports followed at 80.5%, then search, collection, analysis and reporting at 77.0%.

The second is a work revolution. A company examines the entire sequence from a customer question to a resolved answer, from order to shipment, or from an equipment warning to completed maintenance. It prepares the inputs, decides what AI may recommend and what a person must decide, removes redundant approvals, designs exceptions and measures the result. That cannot be completed by distributing software. It moves departmental boundaries, authority, incentives, budgets and sometimes the relationship with suppliers.

Among organizations using, testing or considering generative AI, 63.3% said it was used by individuals for work and 44.8% said individuals or departments were testing it. Only 11.9% said it was embedded in a departmental business process, while 18.6% said it was embedded in a companywide service. These were multiple-response questions, so the figures are not additive. The direction, however, is clear: Japan has accelerated the first revolution and is standing at the entrance to the second.

A long history of ambition—and the harder problem of implementation

Japan did not arrive at AI for lack of ambition. In 1982, the Ministry of International Trade and Industry launched the Fifth Generation Computer Project. It pursued knowledge-based information processing, built parallel inference machines and invested about ¥54 billion over 11 years. The technology was far removed from today’s large language models. But the project matters because it complicates the easy story of Japan as a country that merely “fell behind” in AI. The national effort to make machines work with knowledge began more than four decades ago.

In 2016, the government proposed Society 5.0: a human-centered society in which economic development and the resolution of social problems are made compatible by integrating cyberspace and physical space. In 2018, the Ministry of Economy, Trade and Industry’s DX Report warned that complex, aging core systems could block transformation—a danger that became known as the “2025 Digital Cliff.” AI Strategy 2019, the AI Act of 2025 and Japan’s Second AI Basic Plan in 2026 moved the policy vocabulary from research to social deployment and then to “AX,” or AI transformation.

The second plan defines the challenge in unusually organizational terms. It calls for organizations to fundamentally review decision-making and operations on the premise that AI exists. The government document and the corporate survey are, in effect, pointing at the same gap from opposite sides. Policy calls for a fundamental review. Corporate use is still led by summaries and documents.

Why AI stops at “personal convenience”

The easiest explanation is a talent shortage. Half of respondents—50.1%—identified a lack of specialists as a challenge in adopting and operating AI. Another 45.8% cited inadequate understanding of generative AI’s effects and risks, and 39.7% said it was difficult to create appropriate rules and standards. But “we do not have enough AI people” is not the whole diagnosis.

Rebuilding work requires more than model expertise. It requires the person who knows the exceptions on the factory floor, the person who can define what the customer is actually waiting for, legal and security staff, a data owner, and an executive with the authority to change an approval chain. IPA found that heads of IT led AI use at 46.0% of companies, ahead of executives at 32.1% and business-unit heads at 27.5%. Chief AI officers appeared at 2.9%; other dedicated AI organizations, at 9.8%.

IT leadership is not the problem. Isolation is. If AI remains an information-systems project, it can accelerate a piece of the existing process without removing the process itself. The owner of the workflow and the executive who controls decision rights must also be present.

Departmental use makes that boundary visible. Generative AI had reached IT systems at 71.9% of relevant respondents, corporate planning at 69.0%, and general affairs or legal at 68.4%. Procurement stood at 31.3%, manufacturing—including design and production control—at 35.7%, and distribution at 23.9%. AI enters text-heavy administrative work first. It moves more slowly into work tied to inventory, quality, delivery, machines and physical accountability.

Internal benefits have arrived. External value is next.

Among companies that had adopted or were testing AI, 31.8% said results met or exceeded expectations. Another 50.6% said there had been some effect, but less than expected. That is not a picture of universal failure. It is a picture of many firms harvesting small efficiencies without yet building the bridge to business performance.

For firms that reported an effect, 91.6% cited faster or more efficient operations. Improvements in the quality or speed of proposals reached 48.9%, and reduced overtime reached 29.2%. By contrast, only 4.5% cited higher customer satisfaction and 3.9% cited higher sales or profit.

Measurement is part of the divide. It is easy to observe that AI saved five minutes on an email. It is harder to connect AI to fewer returns, shorter customer waits or revenue from a new service. That requires a baseline and an outcome metric spanning the process. Yet only 23.9% of companies pursuing DX said they had set performance indicators for the results. An unmeasured transformation tends to revert, at budget time, to “a useful tool.”

Behind the smooth interface, the data are fragmented

The interface of a generative-AI service may be frictionless. Corporate data rarely are. Customer names differ across departments. Product codes do not line up. The basis for decisions is scattered through old PDFs, spreadsheets, email and paper. Before debating model intelligence, a company may first have to decide which number is true.

Only 7.0% of respondents said they had prepared training data and were using it for AI. Another 28.4% had prepared data but said it was insufficient; 27.6% recognized the need but had not done the work. Companywide or departmental data use reached 59.9%, but companywide data use alone stood at 20.0%. Among companies with 100 or fewer employees, data use fell to 38.6%.

Cleaning identifiers, defining customers and products, and documenting the history of work is not glamorous. It is what turns AI from a personal assistant into an organizational capability. A firm’s proprietary data are also harder to copy than a model anyone can buy. That is one reason the 2026 AI plan emphasizes Japan’s field data and domain-specific “vertical AI.”

For smaller companies, lag is not destiny

The size gap is substantial. Across the survey, 75.7% of companies were pursuing DX at some level, from a companywide strategy to department-level initiatives. The share was 98.7% among firms with more than 1,000 employees and 41.6% among those with 100 or fewer. For the smallest group, 34.6% were interested in AI but had no specific plans.

Smallness can also be an advantage. A company with fewer departments and a short distance between owner and frontline may find it easier to change one process end to end. An IPA case study from Tokyo’s Ota Ward describes how small factories extended “nakama-mawashi”—the local practice of passing specialized work among trusted workshops—through a digital network. The example predates the current wave of generative AI, but its lesson fits: transformation can begin with a real constraint understood by operators, not with a giant technology program.

An SME does not need a miniature version of a conglomerate’s AI strategy. It can choose one high-friction flow among quotations, inquiries, maintenance, ordering or inspection. It can judge results by response time, shortages, rework, margin and overtime—not by the price of the model. If the design works, it can extend the process one adjacent step at a time.

Governance is not the brake. It is the road.

The deeper AI moves into operations, the more consequential its errors become. Confidentiality, copyright, discrimination and accountability matter more when a model is connected to a customer record, a production decision or a payment. IPA found that 32.5% of relevant organizations had companywide AI guidelines or rules and 28.8% had an internal AI policy. Only 14.2% said they conducted risk management specific to AI or generative AI.

It is easy to read those figures and decide to wait. But companies that reported stronger AI effects also tended to report more risk management. That association does not establish which caused which. It does suggest that guardrails may enable broader use: employees know what data they may enter, what must be verified, when to escalate to a person and who owns the outcome.

The AI Guidelines for Business, version 1.2, compiled by METI and IPA/AISI in March 2026, similarly treats innovation and risk mitigation as tasks to pursue together. It emphasizes management leadership, risk-based controls and accountability to stakeholders. Even a small company can move beyond a one-page prohibition list by documenting, for one workflow, the purpose, permitted data, validation, logs and responsible owner.

From adding AI to rebuilding work

Add AI	Rebuild the work	Measure
Summarize meeting notes	Connect pre-meeting issues, the decision record and assignment of actions	Time to decision; open actions
Draft inquiry responses	Join classification, response, human escalation and the updating of shared knowledge	First-contact resolution; wait time; repeat inquiries
Write quotation text	Connect customer requirements, cost, delivery and approval through common data	Response speed; margin; win rate
Format maintenance reports	Link anomaly detection, parts, work orders and service history	Downtime; repeat failure; inventory

“Giving work to AI” can sound like removing people. Good redesign usually makes human responsibility clearer. AI gathers information, proposes options and handles defined routines. People set the objective, judge ambiguous exceptions, protect customer relationships and remain accountable. If that boundary is vague, accelerating the process also accelerates its mistakes.

What executives should ask on Tuesday morning

At the first meeting after Obon, asking “How many employees use AI?” will reveal the spread of adoption. It will not reveal the strength of the business.

The sharper questions are: Which customer outcome changed? Which step disappeared? Whose decision became faster? Do we have a before-and-after baseline? Who can stop the process when the system is wrong? Are the company’s own data fit to use?

Japanese business is leaving the stage in which caution meant doing nothing. The jump in generative-AI adoption from 22.6% to 44.0% is evidence of that. The next step is not merely to make each employee a little faster. It is to escape the double cost of paying for a new tool while preserving every old procedure, and to redraw work from end to end.

From the giant national project of 1982 to the small text box in a browser in 2026, the distance between Japan and AI has narrowed dramatically. The remaining distance is not between people and computers. It is between new technology and old organizations.

Sources and reporting notes

1. IPA, “DX Trends 2026” publication page, July 30, 2026; updated August 4.
2. IPA, “DX Trends 2026” main report. Unless otherwise stated, figures on adoption, uses, outcomes, data, talent and methodology come from this report.
3. Information Processing Society of Japan Computer Museum, “Start of the 5th Generation Computer Project”.
4. Cabinet Office, “Society 5.0”.
5. METI, Comprehensive Report of the Committee on Legacy System Modernization, for the history of the 2018 DX Report and the Digital Cliff.
6. Cabinet Office, Japan’s Second Artificial Intelligence Basic Plan, provisional translation, as of July 14, 2026.
7. METI and IPA/AISI, AI Guidelines for Business, version 1.2, March 2026.
8. IPA DX SQUARE, Ota Ward digital manufacturing-network case study.

Editor’s note: The opening back-to-work scene is an illustrative composite, not an account from a named company. Survey denominators vary by question, and percentages may not total 100 because of rounding. The exchange rate was supplied by the publisher and its UTC timestamp was converted to Japan Standard Time.