Generative AI in Business: How to Get Results, Use Cases by Function, and the Shift to AI Agents

"We deployed generative AI, but the front line isn't using it as much as we hoped." "The PoC we tried worked, but it never spread across the company." "We don't even know where to start, and no one internally really understands this stuff." When companies come to us about generative AI, these are almost always the first things they say. Signing up for a tool is no longer the hard part. The hard part is turning it into results for the organization.

Let me give you the conclusion up front. What separates the companies that get results from those that don't is not the raw performance of the model. It is the literacy of the people using it and the design of the work itself. This article walks through it at a level of detail you can actually act on: from the basic vocabulary of what generative AI is, to the common root causes of stalled results as of 2026, to concrete use cases by department, to how to choose among the major services, and finally to AI agents, the trend coming next.

Key Terms to Know First

If you leave the terminology vague, internal discussions talk past each other. Let me define the words in this article briefly, up front.

• Generative AI -- AI that creates new text, images, audio, code, and so on. Conversational tools like ChatGPT are the familiar example.
• LLM (large language model) -- The core technology behind generative AI. It learns from vast amounts of text and works by predicting the next word.
• RAG (retrieval-augmented generation) -- A method that searches external information, such as internal documents, and has the AI answer based on that content. You use it when you want answers that follow your own policies and manuals.
• AI agent -- An AI that, given a goal, plans multiple steps on its own and executes a task end to end while operating tools. What sets it apart from conventional generative AI is that it does not stop at a single generation.
• Hallucination -- The phenomenon where AI produces content that is factually wrong but sounds plausible. It is the single most important property to watch when using generative AI at work.
• Human in the loop -- A design philosophy in which a person checks and corrects what the AI generated before it is used. It is the foundation of any defense against hallucination.

Of these, the one you cannot ignore in the 2026 conversation is the AI agent. The center of gravity is shifting away from the 2023-2024 pattern of "have the AI write a draft and let a person fix it" toward handing the work itself to AI.

In 2026, the Question Shifts from "Have You Adopted It?" to "Are You Getting Results?"

Generative AI adoption among Japanese companies has already entered a high plateau. PwC Japan's Generative AI Survey 2025 (published March 2025) found that 64.4% of responding companies had adopted generative AI, and of those, 38.8% said they use it company-wide. The benefit companies feel most is "operational efficiency" (52.3%), followed by "improved quality and accuracy" (33.7%) and "reduced labor and operating costs" (30.4%).

And yet, in the same survey, only 13% of companies said the effect exceeded their expectations. Having adopted it and getting results from it are two entirely different things. What the numbers describe is a reality in which many companies are stuck at "we put it in, but it isn't working as well as we thought."

There is one more thing you cannot overlook: the international gap in usage rates. International comparisons in Japan's Information and Communications White Paper (from the Ministry of Internal Affairs and Communications) repeatedly point out that generative AI usage in Japan, for both individuals and companies, remains low relative to the United States and China. A gap with other countries can translate directly into a gap in productivity. Closing that gap is the homework in front of Japanese companies.

These figures reflect the point in time when they were published. Statistics on adoption rates and outcomes are updated every year, so when you use them for management decisions, we recommend confirming against primary sources such as PwC's latest edition and the most recent Information and Communications White Paper.

Why Results Don't Come: The Real Causes of PoC Fatigue

"We tried it, but it didn't spread." This so-called PoC fatigue has clear, common causes. They are problems on the organization's side, and switching tools will not solve them. The root causes come down to roughly four.

The first is that internal data is not in order. Even if you want RAG to reference your internal documents, if files are scattered and old and new versions are mixed together, the AI cannot answer correctly. Getting your data in order is unglamorous, but skip it and every downstream step falls apart.

The second is that it does not take hold on the front line. Handing out a tool does not make people change how they work. Unless it clicks for them where and how it fits their own job, the accounts sit dormant.

The third is a shortage of talent and AI literacy. Across many surveys, the barriers that repeatedly top the list for company-wide rollout are "we have no one who can use it" and "the front line won't use it." Writing prompts, understanding the limits of AI, verifying the output -- if these basic skills have not spread through the organization, investment does not turn into results.

The fourth is weak governance. If everyone uses it however they like with no rules, you get information leaks and uneven quality, and eventually you slide back to "it's dangerous, so we're banning it."

Three of the four, pushed to their root, are problems of "people" and "how it is used." That is exactly why I said the key to results is not the performance of the model but literacy and work design.

Where to Start: The Implementation Priority Matrix

Starting everything at once is not realistic. Decide the order in which to begin along two axes: the size of the impact and the ease of adoption.

	Easy to adopt	Requires planning
High impact	Top priority: drafting emails and documents, FAQ handling	Plan carefully: demand forecasting, building an internal knowledge base
Medium impact	Next tier: meeting-minute summaries, translation and multilingual support	Evaluate: contract review, code generation
Limited impact	Watch and wait: image generation, social posts	Defer: building your own model from scratch

The first step is the top-left quadrant: "high impact, easy to adopt." If you can produce one clear number for time saved here, the mood inside the company changes. What pulls in the next budget and the next round of cooperation is not "this looks useful" but the fact that "this task dropped by X hours a month." For the big picture of adoption, The AI Adoption Roadmap lays out the stages in order.

Use Cases by Business Function

From here, let's look at approaches that are actually producing results, department by department. The amounts and timeframes are representative benchmarks. As you read, notice the one thing the success stories have in common: without exception, the companies that are doing well do not use the AI's output as-is. The person using it puts in the final touch.

Sales -- Making Proposal Drafting More Efficient

A 50-employee staffing agency rolled out a proposal-drafting workflow to all 10 salespeople, cutting first-draft time from an average of 60 minutes to 15 minutes. The point was not "let the AI write the proposal and send it as is," but a flow in which the salesperson layers their own knowledge onto the AI's draft. Over a year, that removed roughly 920 hours of work.

Customer Support -- Automatically Updating the FAQ

A 120-employee SaaS company built a mechanism using an internal knowledge AI that analyzes past inquiry history to auto-generate and update its FAQ. The self-resolution rate for inquiries rose from 35% to 58%. The reason quality held up was that the whole team stuck to one rule: never take the AI's answer at face value.

Accounting -- Automating Invoice Processing

A 300-employee wholesaler connected OCR with generative AI and cut the time to process 800 invoices a month by 60%. Manual-entry errors fell from an average of 12 a month to 2.

HR -- Making Recruiting More Efficient

An 80-employee IT company used AI to draft job postings (from 30 minutes down to 5), to generate interview questions, and to draft standard replies to applicants. It cut the two-person recruiting team's monthly workload by about 25% and redirected the freed-up time to candidate interviews, which they say raised the offer-acceptance rate.

Legal -- Assisting with Contract Review

A 200-employee manufacturer introduced a system in which AI analyzes contract clauses and flags high-risk passages and suggested revisions, cutting review time per contract from 90 minutes to 35 minutes. But it never lets the AI make the final call on its own. It keeps to a workflow where a legal team member always checks and approves. The heavier the legal risk, the more you thicken human involvement -- that is the ironclad rule.

Manufacturing -- Auto-Generating Quality Reports

An 80-employee precision-parts maker built a system where feeding inspection data to the AI produces a report in the prescribed format, cutting the quality-control team's (5 people) report writing from 40 hours a month to 12. Inspectors now only need to check the content and make small corrections.

Marketing -- Producing Content

A 40-employee B2B company put AI to work on blog outlines, newsletter drafts, and social post copy, doubling its monthly output. It set an operating rule that fact-checking and adjusting to the company's own tone remain with people.

Construction -- Supporting Estimate Preparation

A 150-employee construction firm auto-generates draft estimates for new projects based on past estimate data, cutting preparation time by 60%. The big payoff was that, with veterans' know-how accumulated in the AI, even junior staff could produce estimates of consistent quality.

Customer Support -- Multilingual Handling (An Example Rebuilt After Failure)

A 500-employee manufacturer automated English- and Chinese-language inquiries with an AI chatbot. At first the response accuracy was poor, and mistranslation of technical terms drew three complaints from overseas customers. So the team manually added an industry-terminology dictionary and added a rule to automatically escalate complex inquiries to a person. After the rebuild, customer satisfaction exceeded pre-deployment levels. The turning point was that, when things weren't working, they did not conclude "AI is useless" -- they fixed the design.

Company-Wide -- Putting Internal Knowledge to Work

A 250-employee food manufacturer set up a system to search and summarize its policy manuals, procedure documents, and past meeting minutes with AI. "Where's that document again?" inquiries dropped from about 200 a month to 30, and the risk of knowledge draining away when veterans retire eased as well. When you want everything to stay within internal data, a RAG that runs entirely on domestic servers -- such as enterprise AI (ZEROCK) -- becomes an option.

Line up all ten examples and what comes into view is not a difference in tools but a difference in the people using them and in how the work is designed. With the same tool, a front line with literacy produces results, and one without it leaves the tool dormant.

Key Points by Industry and Company Size

The shape of adoption is, to a degree, determined by industry and size. Starting from whatever is closest to your own situation makes it easier to move forward.

• Manufacturing -- Generating standardized documents such as inspection reports and work procedures, and searching past trouble cases, works well. For processes that handle drawings and design data, judge based on fit with a purpose-built system.
• Retail and services -- Writing product descriptions, handling inquiries, and mass-producing promotional copy are easy areas to enter.
• Finance -- Because regulation is strict, prioritize confidentiality and audit readiness, and design on the premise of a configuration that stays within domestic data and includes log management.
• Construction -- Loading veterans' tacit knowledge into drafts of estimates, construction plans, and reports also helps develop junior staff.
• Large enterprises -- With many departments, a uniform company-wide rollout is hard. The approach that fits is to create success stories in a leading department and expand horizontally.
• Small and mid-sized companies -- Precisely because people are stretched thin, the per-person time savings show up big. The realistic move is to start small with one task, using free or low-cost tools.

Comparison of Major AI Services

Here is a summary of the services companies most often consider. Each provider frequently updates its plan structure, pricing, and model generation, so treat amounts as a rough guide and be sure to confirm the latest information on the official site before contracting.

Service	Provider	Strengths	Data handling
ChatGPT (Business / Enterprise)	OpenAI (US)	Highly versatile, rich set of extensions	Business plans do not use input data for training
Claude (Team / Enterprise)	Anthropic (US)	Strong at long documents and safety	Business plans do not use input data for training
Gemini (Workspace / Enterprise)	Google (US)	Integration with Google products	Can integrate with data inside Workspace
Microsoft 365 Copilot	Microsoft (US)	Integration with Office products	Integrates with the Microsoft 365 environment
ZEROCK	TIMEWELL (Japan)	RAG specialized for internal data, GraphRAG, domestic servers	Runs entirely on domestic servers, securing data sovereignty

There are three key points for choosing. For highly confidential work, a domestic-server setup like ZEROCK gives you peace of mind. If general-purpose document creation and research are the center of gravity, the cost-performance of ChatGPT and Claude comes through. If you already run a Google or Microsoft environment, Gemini or Copilot is the natural choice. If you're still at the stage of deciding on a tool, our AI Vendor Selection Guide is worth reading alongside this.

Risk and Governance by Application Area

Putting generative AI to work always comes with risk. Judge how heavy the risk is for each area, and prepare countermeasures ahead of time.

Application area	Primary risk	Risk level	Essential safeguard
Drafting internal documents	Hallucination	Low to medium	Human review
Customer communications	Misinformation, tone mismatch	Medium	Template management, approval flow
Contract review	Overlooked legal risk	High	Legal team always makes the final check
Data analysis	Wrong trends extracted, bias	Medium	Cross-check against statistical validation
Code generation	Vulnerabilities, license issues	Medium to high	Mandatory code review, license verification
Image generation	Copyright infringement, brand damage	Medium	Choose services cleared for commercial use

The practical issue that has surfaced especially in recent years is "shadow AI." It refers to the risk that an employee, without company permission, enters work data into a personal account, and information leaks from there. Ban it outright and, if anything, underground use grows. It is a more realistic countermeasure to provide an official tool people can use safely, give that behavior a legitimate outlet, and put rules and an approval flow in place for entering confidential information.

For organizations handling highly confidential information, a configuration in which data never leaves for overseas servers -- so-called domestic data sovereignty -- also becomes a consideration. If you keep RAG entirely on domestic servers, you can put internal knowledge to work while holding down leak risk. For the full picture of rules and structure, The AI Governance Framework covers it in detail.

From Generative AI to AI Agents

The biggest trend of 2026 is the AI agent. Until now, generative AI was a tool where a person gave instructions and had it generate text or images one at a time. An AI agent, given a goal, plans multiple steps on its own and executes a task end to end while operating tools.

For example, take an instruction like "total up last month's sales, put it into a report, and share it with the relevant people." Picture the AI retrieving the data, tallying it, writing the text, and sending it out -- carrying the process forward, including the judgment calls in between. From one-off draft generation to autonomous execution of a business process: the axis of use is moving up a level.

Even so, the foundation does not change. The wider the range you hand to an agent, the more weight there is in separating out which work is safe to delegate, in the mechanisms to detect errors, and in a design where a person holds final responsibility. Here too, what pays off is the organization's AI literacy. Step into autonomous execution while skipping the basics and you invite failures you cannot control. The healthy order is to climb in stages -- from efficiency as your foundation, to value creation, and then to autonomous execution.

Four Principles for Successful Adoption

Building on the use cases and root causes, here are four principles for producing results.

First, design a "human + AI" workflow. Generative AI carries the risk of hallucination. AI creates a draft, a person checks and corrects it, and the final judgment rests with the person. Building this human in the loop into the work is the starting point.

Second, start with efficiency and then move toward value creation. Rather than aiming for "innovation with AI" from the first move, first shore up your footing by making existing work more efficient. A success experience of time saved pushes the next challenge forward.

Third, measure impact quantitatively. "It feels more convenient" is no basis for an investment decision. Task time, volume processed, error rate, customer satisfaction, amount of cost saved -- get to where you can speak in the before-and-after numbers. How to measure is laid out in A Guide to Calculating AI ROI.

Fourth, invest in talent and AI literacy. This is the principle many companies leave out. As we saw earlier, the great majority of the root causes of stalled results were "people" problems. Investing in the people who will master the tool -- as much as or more than investing in the tool -- is the shortcut to results.

How to Think About Cost Estimation

Adoption cost varies enormously by purpose and scale. Having a rough sense of the going rate keeps your internal budget discussions from wobbling.

Level of use	Rough initial cost	Rough monthly operating cost	Representative purpose
An extension of individual use	0 to 100K yen	10K to 50K yen	Document creation on individual accounts
Team use	100K to 1M yen	50K to 300K yen	Shared team tools, prompt libraries
Business-process integration	1M to 5M yen	300K to 1M yen	API integration with existing systems, building RAG
Company-wide platform	5M yen and up	1M yen and up	Company-wide knowledge base, integrating multiple functions

What tends to get overlooked is hidden cost. Plan for more than tool fees: data cleanup (which can reach 30 to 50% of the initial cost), employee training, and the effort of changing business processes. Training cost, in particular, is an investment in getting results -- not somewhere to cut.

Embedding AI Adoption in Your Organization with WARP

As I have said repeatedly, what separates results in generative AI is the literacy of the people using it and the design of the work. TIMEWELL's WARP is a hands-on AI-talent development program led by working AI developers, designed precisely to solve this "people" challenge. It has served more than 250 participants to date, with a 90% completion rate.

Depending on where your organization stands, you can build up in this order, without straining:

• WARP 1Day -- For the stage where you first want to raise the whole company's AI literacy. A short, intensive session that turns the basics of generative AI and where it fits in real work into shared language for everyone.
• WARP BASIC -- An e-learning-style AI learning platform. At 8,500 yen per person per month, people learn at their own pace from content updated every month. It becomes the foundation for taking root a habit of learning across the company.
• WARP NEXT -- A three-month intensive talent-development program. It pairs 32 hours of coursework with one-on-one mentoring, at 250,000 yen per person. Signing up includes one year of WARP BASIC at no charge. It suits the stage where you want to develop the core people who will drive AI use inside the company.
• WARP ENTRE -- A program for going beyond efficiency into new business and even founding companies. It has supported the launch of more than 15 companies to date.

"We don't know where to start" and "we have no one internally who can use it" are exactly the challenges WARP has been built to face. You can gauge where your own organization's literacy stands in a few minutes with the AI Literacy Check. From there, if you'd like to talk through the approach that fits your company, reach out via a personal consultation.

Frequently Asked Questions

Can small and mid-sized companies use generative AI too? Yes. You can start with free or low-cost tools, so the upfront investment is minimal. Companies short on people tend to feel the per-person time savings most acutely. Rather than a company-wide rollout, the realistic move is to start small with one task in one department.

What is the difference between the free version and a paid business plan? The biggest differences are how your input data is handled and what management features you get. Most business plans are contractually committed to not using your input to train the model, and they let you manage logs and set access permissions. For business use, a business plan or a setup that keeps everything on domestic data is the safe choice.

How do you reduce the risk of information leaks? Three basics: choose a business plan that does not use your data for training, set rules for entering confidential information, and put an approval flow in place. Since shadow AI -- employees using it without permission -- is the main risk, providing an official tool people can use safely and giving it a legitimate outlet is more effective than a ban.

Is training necessary? If you want results, we believe it is. The great majority of the root causes of stalled results are on the talent side: "we have no one who can use it." Rather than handing out the tool and calling it done, an investment in spreading real fluency through the organization ends up speeding your return.

Is it better to build in-house or bring in outside support? A combination of both is realistic. Accelerate the raising of baseline literacy and the development of core talent with outside training, and internalize day-to-day operation and improvement. Try to carry the whole thing in-house from the start and you tend to burn out at launch.

Summary

• Generative AI adoption among Japanese companies has plateaued high, and the question has shifted from "have you adopted it?" to "are you getting results?"
• The root causes of stalled results are four -- unprepared data, weak front-line adoption, a lack of talent literacy, and weak governance -- and most of them are problems of "people" and "how it is used"
• Start from the "high impact, easy to adopt" quadrant of the priority matrix, and the first step is showing time saved in numbers
• The success factor common to all ten examples is not using the AI's output as-is. It is the people using it and the design of the work, not the tool, that separate results
• In 2026 the AI-agent trend gets serious, and the axis moves from one-off generation to autonomous execution of work
• As much as investing in the tool, investing in talent and AI literacy is the shortcut to results

Generative AI is not a magic tool that works the moment you install it. It turns into results only once you develop the people who master it and build it into the work. Start by finding out where your own organization's literacy stands.

---

Related articles:

• The AI Adoption Roadmap -- The phased adoption plan for turning these use cases into reality
• Improving Organizational AI Literacy -- How to develop the "people" power that separates results
• The AI Governance Framework -- The rules and structure that generative AI use requires
• Change Management for AI Adoption -- Management methods for embedding use into the organization
• A Guide to Calculating AI ROI -- How to measure the impact of these use cases correctly

References (Primary Sources)

• PwC Japan, Generative AI Survey 2025 (published March 2025): https://www.pwc.com/jp/en/knowledge/thoughtleadership.html
• Ministry of Internal Affairs and Communications (Japan), Information and Communications White Paper (generative AI usage rates for companies and individuals, international comparison): https://www.soumu.go.jp/johotsusintokei/whitepaper/
• TIMEWELL, WARP AI Talent Development Program official page (program details and pricing): https://timewell.jp/en/warp