DX in the Era of Generative AI and AI Agents生成AI・AIエージェント時代のDX

It's been over two years since ChatGPT captured the world's attention in 2023, and the term "Generative AI" has now become one of the defining keywords of our era.

As dizzying technological evolution continues daily, foundation models continue to be updated further with new models equipped with reasoning capabilities and open-source models. Many experts position 2025 as "the first year of AI agents," predicting that AI utilization will become even more critical in corporate DX strategies.

This time, under the title "DX Promotion in the Era of Generative AI and AI Agents," I would like to write about important points that each company should grasp in promoting their own AI utilization.

Introduction

When generative AI first gained attention, the focus was mainly on chat-type services that general users could easily use.

However, in just a few months afterward, cloud services aimed at corporate use, open-source models, as well as plugins, APIs, and open-source libraries appeared one after another.

The utilization of generative AI has gone through the PoC (proof of concept) stage in various industries and business formats, and is now transitioning to the phase of "how to make it take root internally" and "how to connect it with our company's unique strengths." Since the latter half of 2024, many experts have been emphasizing the importance of "AI agents," but on the other hand, there is still a large gap with the actual field in this dawn of generative AI, such as "We deployed ChatGPT but it's still not being used much internally" and "Even if you say AI agents, we don't specifically know where to start." Regarding generative AI-related news, with the US government's Stargate Project, Chinese companies' announcement of DeepSeek, and OpenAI's planned release of GPT-4.5 and GPT-5, there is no shortage of new topics daily.

On the other hand, information including things that are not very important has come to be widely disseminated, making it very difficult to understand what points should be fundamentally grasped in the generative AI era, and what points companies should really be focusing on in DX promotion.

In the current situation where technological evolution is fast and new information is being announced in rapid succession, information that merely summarizes the latest trends has almost no value. The shelf life of the latest information on generative AI is too short, as it becomes "that information is already old" a few weeks later.

This time, rather than content that relies on specific topics (branches and leaves) that change with the latest information, I would like to think about the "trunk" part of what needs to be fundamentally addressed in corporate DX promotion when considering the medium to long term span.

Why Isn't ChatGPT Used Much Even When Distributed Internally?

This has already been described in "The Real Reason Why ChatGPT Isn't Used Much Even When Distributed Internally", but since many people may not have read it, I'll explain just the key points.

ChatGPT captured the world's attention, but actually, how much is it being used within companies?

Thinking "Our company must not fall behind this AI trend," many companies, mainly large enterprises, started deploying ChatGPT for their internal employees from 2023.

They actively incorporated ChatGPT usage training and practical training, and while expecting significant improvements in employee work efficiency through the introduction of generative AI, post-implementation surveys showed usage rates were only around 10% internally, greatly falling short of the initial expectations of the implementers.

So why isn't ChatGPT, which attracted so much attention worldwide, being used that much internally?

While this phenomenon is often attributed to insufficient ChatGPT performance or lack of employee IT literacy, the fundamental reason is much simpler: "ChatGPT knows nothing about internal matters." As a basic premise, since ChatGPT is trained on past open data, it can solve general problems in programming, multilingual translation, mathematics and physics, etc. with high accuracy as is.

And since web search functionality has also been incorporated, ChatGPT has become able to basically solve problems that can be handled with past open data + web search.

Due to this nature, which job types actually use it frequently includes IT engineers, researchers, and web marketers.

Programs are a universal language worldwide and are included in the training data of foundation models, so for IT engineers, it becomes a great ally for programming code generation and bug fixes, and it can be said that work is no longer possible without ChatGPT.

For researchers too, the scope of business use is quite wide, including literature surveys, translation, and writing support, and for web marketers, there are many opportunities to use it in many situations such as persona analysis, brainstorming, copywriting creation, and web article writing support.

On the other hand, in the case of large companies, etc., how many people in such job types are there internally - they would be positioned as only a minority. Most job types, including the majority of front office sales people and back office people in general affairs, human resources, legal, and accounting, are conducting business centered on internal information, not open data.

For sales people, customer information and transaction information registered in CRM, etc. are essential for conducting business, and for back office people, internal business processes, regulations, organizational information, etc. are also essential as prerequisite information.

On the other hand, ChatGPT knows nothing about such internal information. In other words, for job types that make up the majority internally, there are not many tasks they can request support for from ChatGPT, which knows nothing about internal matters.

When thinking in anthropomorphic terms about what state it is to introduce ChatGPT as is internally, it's a state where an extremely excellent person has been hired, but this person is being made to work in an environment isolated from internal information.

Employees can ask this person anything or make any request via chat, but this person knows nothing about internal matters. In other words, for people doing work where internal information is essential, no matter how smart this person is, there aren't many things they can request.

So on the other hand, what would happen if this person could access internal information?

This person is extremely smart, can handle any language, and can even program. On top of that, they know internal matters better than anyone, and become an existence that responds to chats 24/365 without complaint.

In this case, they would probably become a superman-like existence that everyone relies on. In other words, from the perspective of internal use, the key is how much internal information ChatGPT knows.

This is not a problem that will be solved no matter how much ChatGPT's performance improves in the future, or how much employee IT literacy improves, so it's necessary to advance individual countermeasures internally.

Specifically, using methods called RAG and Fine-Tuning, it becomes a matter of creating a state where ChatGPT can access internal information, but since Fine-Tuning has difficult costs and evaluation for retraining, and information control, in many cases RAG will be central.

In advanced companies, there were many cases of developing internal information search using RAG from the latter half of 2023 to 2024, but internal information search is suitable as a first step initiative, and through this initiative, the importance of internal information management in the AI era and a sense of the characteristics of generative AI come to be cultivated among stakeholders.

For companies where specific initiatives regarding generative AI have not yet been advanced, without getting stuck in abstract discussions, as one guideline it would be good to first set a goal of minimizing to zero as much as possible the acts of "searching for information and asking people," which do not themselves generate added value, by connecting ChatGPT to internal information.

Even in the social trend where personnel mobility is increasingly high and hybrid work styles such as remote work are further expanding, a state where there is no internal information search function by generative AI and there are many things that can only be known by asking people will greatly reduce work efficiency and become a factor in declining corporate competitiveness.

Development of "Specialized AI Agents" That Will Be the Main Axis of DX Promotion from 2025 Onward

2023 was when ChatGPT was first distributed internally and deployment was promoted, 2024 was when implementation of internal information search using RAG, etc. was advanced, but 2025 is said to become the first year of AI agents.

Why AI agents become important is because this AI agent, especially specialized AI agents, will be a major factor determining corporate competitiveness going forward.

In a situation where foundation model updates are fast and various generative AI-related services are being released in rapid succession, while catching up with this information is also important, what companies should really focus on in DX promotion is the development of their own specialized AI agents. This is for the following two reasons.

1. Both Foundation Models and Generative AI-Related Services Will Eventually Become Commoditized Currently, various information is updated daily, and it's already impossible to catch up with all the latest information, but foundation models themselves will eventually become commoditized, and generative AI-related services will also converge to a form where only truly good ones remain through natural selection.

Foundation models initially had large differences depending on each company's model, but that difference is already becoming smaller, and from the user's side, it's no longer possible to make judgments about whether each model is fundamentally good or bad.

Just as from the user's perspective there is now essentially no fundamental difference in the performance of home appliances, cars, and smartphones released by various companies, it will soon become a situation where there are generally no problems if you use any of the latest models from each company.

Also, when the shift from on-premise to cloud/SaaS occurred, a large number of products were introduced to the market, but natural selection has progressed since then, and now major players in each area are gradually becoming fixed.

When Office was first introduced, being able to use the latest tools Excel, Word, and PowerPoint was valued, but now everyone uses them as a matter of course, and similarly, when generative AI-related services become commoditized with major players fixed, using them will become a matter of course for everyone.

Both foundation models and generative AI-related services are, in the end, things that anyone can buy if they pay money for products on the market. Things that anyone can buy by paying money cannot become fundamental differentiation factors or competitiveness.

Of course, the latest information should be caught up with and the latest tools and products should be used, but since foundation models provided via API and SaaS tools can be used by anyone who pays money, it's fine to respond while watching market conditions.

In any case, catching up with the latest information on foundation models and tools and rushing to implement them even a step ahead of other companies does not lead to fundamental competitiveness strengthening.

2. Both Foundation Models and Generative AI-Related Services Are General-Purpose Products Another important perspective is that foundation models and generative AI-related services are basically general-purpose.

From the product development side's standpoint, they develop centering on shared functions so that as many people as possible can use them.

As can be seen from the fact that many generative AI-related services promote meeting minutes creation, summarization, email draft creation, voice transcription, etc., when trying to increase users as much as possible, it inevitably becomes this kind of common area function provision.

Industry-specific tools are also emerging, but even if industry-specific, when viewed within the industry, they become general-purpose, so these also do not directly lead to competitiveness. Because if they are truly good products, each company in the industry will come to use them, and these will also become commoditized.

Thus, while products on the market will be commoditized and are basically for general-purpose use, catching up with and using these latest tools is essential, but this itself cannot be called a DX strategy.

The source of differentiation and competitiveness strengthening is, after all, that company's unique business model, business processes, knowledge, resources, etc. In other words, if you can build dedicated specialized AI agents that can maximize your company's strengths, you can greatly leverage your company's unique strengths, and they become strengths that other companies cannot follow.

For non-core operations and general-purpose areas, while using market products and SaaS tools well, how strong specialized AI agents dedicated to your company you can create in core areas will become one of the most important DX strategies (turning point) going forward.

Specialized AI agents are basically not subject to constraints like human resources and can be replicated as much as you want. Therefore, if you develop and refine your own high-performance specialized AI agents, depending on the industry, there is potential to become a state of one company dominating.

Management resources are often said to be people, things, and money, but in the future they will probably be people, things, money, and AI, and companies without AI will greatly lose competitiveness.

From 2025 onward, said to be the first year of AI agents, it will probably become a competition of how strong specialized AI agents each company can create, which can now be called one of the management resources.

What Companies Should Prepare in the Era of Generative AI and AI Agents

From here, as a more concrete discussion, I will explain three important points that companies need to prepare in the era of generative AI and AI agents.

1. Development of Specialized AI Agents

First, the most important thing is to advance the development of specialized AI agents that leverage your company's unique strengths, as described up to this point.

This needs to be considered separately from general-purpose areas and product utilization such as so-called ChatGPT utilization training or Microsoft Copilot utilization.

Rather than the perspective of efficiency improvement and improvement of general-purpose operations, it's important to formulate a concept from the perspective of "What AI agent best leverages our company's strengths?" and advance the development of specialized AI agents.

The difference in corporate competitiveness will be clear between companies that generally use ChatGPT, Microsoft Copilot, etc., and companies where, in addition to utilizing ChatGPT, Microsoft Copilot, etc., multiple specialized AI agents unique to their company are operating.

Below, I'll describe several points that become important in planning and developing specialized AI agents.

Point ① Aim for No-Prompt as Much as Possible

Since generative AI became the focus of attention, the term "prompt engineering" has gradually gained attention.

In order to properly draw out the functions of generative AI, short chat instructions alone are insufficient, and it's important to construct appropriate prompts so that instructions and purposes are clear.

This is the same story as when a superior gives instructions to a subordinate - with vague and ambiguous instructions, just as a subordinate doesn't know what to do, it's important to give instructions to AI as detailed and specific as possible.

Prompt engineering itself is of course important, and there's no doubt that it's better to know it as knowledge, but when developing AI agents, what's important is rather the opposite - how to make it so that users don't have to input prompts.

Even if you loudly proclaim "Prompt engineering is important" and actively implement training and education, in most cases it will not become widely established among employees. This is because while there is no objection to the importance of giving detailed and specific instructions, it's simply bothersome. As a basic premise, no one wants to type long chats.

In the first place, excellent human resources can, to put it bluntly, be rephrased as "people who take care of things appropriately." In other words, human resources who move on their own without giving detailed instructions, and who in some cases move on their own before instructions are given from here.

Human resources who cannot move without being given concrete and detailed instructions cannot be called high performers, but are rather positioned as low performers. Everyone seeks people who move on their own proactively without giving various instructions.

In other words, in constructing excellent AI agents, what becomes important is how much you can reduce the burden of instructions through user prompt input.

In some cases, if you can build an AI agent that can be used with no prompt input required, so-called no-prompt, there's nothing better than that.

Making it work with as short a chat as possible is one thing, but a configuration where options are presented and you only need to select a button would also have high usability.

After all, when presented with a chat field, everyone hesitates about what kind of content to input.

Even with the premise that prompt engineering is important, if you build an AI agent where users have to input many prompts, probably not much usage will become established.

What should be aimed for is rather the opposite - to build specialized AI agents that embed your company's business processes, know-how, unique data, etc., that users can easily use without typing difficult chats, and that raise users to the level of best practices.

High freedom of input also means that it can become better or worse depending on how the user uses it, so this is rather an area that general-purpose AI will handle.

The development of specialized AI agents has a major premise that you are specializing for your company's unique purpose, so rather than having high freedom, having a design where appropriate output is always obtained with minimal input from the user is closer to the ideal form as a specialized AI agent.

In specialized AI agents, the prompt already contains your company's business processes, know-how, unique data, etc., so the prompt users need to input should be minimal.

Point ② Be Conscious of Human-in-the-Loop Design

When it comes to AI agents, the image of "autonomous type" is strong for many people. Autonomous AI agents are AI that autonomously execute multiple processes to solve user requests and complete tasks.

This autonomous type is more evolved than chat-based generative AI in that it can get one step closer to the final output from input, and certainly, this is one of the important features of AI agents.

However, for the time being, there are still issues with the reliability of generative AI and AI agents, and there is still considerable distance to complete end-to-end autonomous execution in real business.

The important thing in developing specialized AI agents is to think about "how much to rely on AI," and while having an understanding of the latest technologies and capabilities of AI agents, it's important to decide the scope while keeping a firm grasp on "whether it can withstand use in business." It's not necessary to be overly particular about fully autonomous type here, and it becomes effective to have Human-in-the-loop type, that is, to build in appropriate human intervention as part of the design.

In other words, since AI, which still has issues with reliability and is in a developmental stage, is difficult to trust completely, include perspectives of human confirmation within a series of processes as part of the design.

This is the same story even with humans - even if you hire the most excellent human resources, you would be anxious if they just said "I finished everything" on the first day regarding the work you entrusted.

The movement expected here, probably the most excellent human resources would be those who move without giving various instructions, but who properly report, communicate, and consult.

In other words, even if autonomous type, not fully autonomous type, but by appropriately designing reporting, communicating, and consulting at necessary timings, users can use AI with peace of mind.

Including myself as an engineer, I think everyone has a desire to make it as autonomous as possible from the technical interest and impact, but in the end, if it's not used by people in the field, it's meaningless. Because if it's just technically interesting, the added value is zero.

Many people have a strong image of "AI agent = autonomous type," but without being overly caught up in this image, while talking with the people who will be actual users, it becomes important to find a good landing point including the design of the scope to entrust to AI and the scope for human intervention.

2. Maintenance of APIs and Data Infrastructure

I've described the importance of specialized AI agents up to this point, but then, if you decide "Let's develop our own specialized AI agents!", can you just go ahead and advance AI agent development?

Actually, what becomes equally important as your company's AI concept formulation and planning that leverages strengths is "Is an environment where AI can move freely already in place internally?"

The construction of specialized AI agents naturally requires access to internal information, but whether AI agents can access internal information, in other words, means whether API integration with systems is possible, and whether data is maintained in the data infrastructure.

To put it bluntly in technical terms, it means whether AI can interact with systems via API, and whether necessary data can be retrieved from the data infrastructure with SQL.

Unlike humans, AI agents don't operate system screens (GUI), so when retrieving data from systems or registering data in systems, they do it via API, and when accessing company data, etc., they retrieve data by issuing SQL to the data infrastructure.

Therefore, in a situation where there is still a lot of paper internally, only on-premise systems that can only be operated via GUI, and data is scattered everywhere, AI cannot access data in the first place, so development of specialized AI agents cannot be advanced.

Also, no matter how strong foundation models emerge in the future, what AI can do will be quite limited.

This is the same as how no matter how fast a car is developed, it cannot demonstrate its functions on unpaved roads or roads that cannot be traveled on in the first place.

Some people may hope "We haven't actively advanced digitalization until now, but if we work hard on AI from now, can't we make a comeback all at once?", but there is no leapfrogging here, and it becomes a state where companies that have steadily advanced DX until now can start earliest and also increase that speed.

While DX promotion up to now appears to have been game-changed by AI, it should rather be understood as a positioning where companies that have properly done DX until now (companies that have steadily prepared their footwork) can further accelerate.

However, some may think "In terms of order, shouldn't preparing APIs and data infrastructure come first, and AI agent development come later?", but there's a reason I wrote about the importance of specialized AI agent development first.

That is because "DX should be advanced output-first." This is because, in cases where there is distance between business departments and IT departments, which is typical, there are often cases where the IT department first prepared the data infrastructure, but what specifically to use it for is not decided (actually nobody is using it).

This is a typical input-first approach, and is the same as studying something thinking you might use it someday, but never getting an opportunity to use it.

Simply accumulating data doesn't mean you can make some kind of good AI, and actually, unless what kind of AI is needed is specifically decided, what data is needed cannot be defined.

Many people say "AI needs data so we have to prepare data anyway" or "We can't implement AI because our data isn't prepared," but actually it's the opposite - "We want to make this kind of AI, so we need this kind of data" is the correct order of examination.

In other words, when you can concretely form an image of the specialized AI agents you want to make at your company, what data is needed to make this AI becomes concretely decided.

Regarding data granularity as well, whether monthly is fine, whether daily is needed, whether batch processing is fine, or whether access to the latest information in real-time is needed, etc., cannot be concretely decided unless the actual output functionality is decided.

Maintenance of APIs and data infrastructure is of course important, but if you just say "Let's maintain APIs and data infrastructure," the scope cannot be decided with this itself, so it's better to first define the output of specialized AI agents, and then prepare the necessary APIs and data infrastructure by working backward.

With input-first thinking, data prepared thinking "we might use this too" often ends up not being used, so advancing output-first allows you to run the shortest distance without waste.

3. Preparation of Integrated UI

This is actually not limited to AI, but what needs to be seriously examined in the coming AI era is this examination of integrated UI.

This is because while general-purpose and specialized AI agents, etc. are expected to further increase going forward, already at this point systems and data continue to increase, and from employees' perspective it has become a state of "not really knowing where what is." Particularly in large enterprises, cases where links to systems that can be used are listed on so-called internal portal sites are common, but probably no one fully understands all internal systems and functions.

Conversations among employees like "I didn't know there was such a system..." "I can get this data from here... I didn't know..." are everyday situations, and monitoring whether systems are properly recognized and used in the first place has become an important issue for DX and IT departments, equal to or more than the purpose of improving system convenience.

And since data and systems that can be utilized will continue to increase going forward, it can be said that they have completely exceeded the level that each individual can recognize.

In the first place, the number of systems that can be properly recognized and used at the individual level is probably around 5. Up to about 10 is still manageable, but when it comes to 20, 30, or more systems, it exceeds the limits of recognition and will become a state with many systems that are simply not known.

Already in such a situation, even if new AI functions are added to each system or individual specialized AI agents are developed, it can be easily imagined that the walls of recognition and diffusion cannot be overcome in the first place.

No matter how convenient systems or AI are built, if their existence and convenience itself are not known, naturally there is no point in building them.

On the other hand, if each system person in charge or AI construction person in charge independently focuses on internal recognition and diffusion, even though within the same company, it becomes a structure where each person in charge competes for the limited recognition resources of employees.

When this happens, only things from departments with loud voices or departments skilled at internal marketing stand out.

Ideally, DX and IT departments should devote efforts to system and AI concept formulation and planning, development and testing, usability improvement and quality improvement, but the man-hours that must be allocated to internal marketing for recognition and diffusion continue to swell.

In an era where systems and data continue to increase and AI agents are also newly built, "how to deal with employees' recognition problem" becomes an unavoidable problem.

One effective solution here is to create an integrated UI with AI as a concierge, making it a touchpoint with each system and specialized AI agent groups.

Each system and specialized AI agent are naturally in an optimal UI configuration to achieve their respective purposes, so it's not realistic to integrate these themselves into one.

In such a fast-changing era, a monolithic architecture will eventually collapse, so it's necessary to maintain a state like microservices.

Without touching existing systems, etc., preparing an integrated UI that users access first, and from here guiding them to necessary systems and AI agents according to their purposes becomes one effective form.

To put it simply, this means "creating a state where you can always ask someone who is fully familiar with all systems and AI agents that can be used internally." The advantage of this state is for both the employee side who are users and the DX/IT department side that deploys systems and AI agents.

First, for the employee side, as various new systems and AI agents are deployed going forward, it becomes a state where they just need to access here first, so they no longer get lost in utilizing systems and data, and it becomes very easy.

Including internal systems and SaaS, aren't there many employees who are fed up with the internal IT environment, with many systems in use, chatbots proliferating, systems migrating to new ones before they know it, etc.?On the other hand, there are also many voices saying "If there was such a system, I wanted to know earlier..."

In other words, having a concierge-like AI that teaches "Ah, if you want to do that, use this system/AI, and do it like this" in the shortest way is very welcome for the employee side.

Especially as personnel mobility increases, it would be a strong ally for new employees and people who came to new departments through personnel transfers, wouldn't it?

And not only for the user side, but the advantages are large for the DX/IT department side as well, the reason being that they can concentrate on system and AI development where they should originally focus their efforts.

What does this mean? As explained earlier, since systems, apps, and AI already exceed the level that each individual can recognize, DX and IT departments are not in a state where they can concentrate only on development, but activities like internal marketing have become necessary.

In other words, especially in the case of new BI and AI apps, it doesn't end simply by developing and releasing, but the weight of activities to make people know and use, such as holding explanation sessions after release, study sessions, and steady diffusion activities, is getting larger and larger.

I myself was confident in the team that we made something quite good, but after release the utilization rate didn't rise as expected, and thinking "Maybe there are issues in UX or functionality...", when I took individual surveys, I was surprised that there were many opinions saying "If there was something this good, I wanted to be told earlier." I thought I had properly done explanation sessions gathering stakeholders and notifications, but still, in the internal IT becoming complex, the hurdle of making people know is getting bigger and bigger.

This trend will become even stronger going forward, so a change in thinking is necessary. The point here is that while there is a limit to the number of systems and apps each individual can recognize, AI has no such limit.

In other words, giving up on having each employee recognize all systems and use them properly, and making AI recognize everything and recommend what's necessary, thereby effectively eliminating the need for internal marketing.

For the employee side, they first access the integrated UI, tell the AI concierge what they want to do, and it guides them to the optimal system or AI agent, and for the DX/IT department side, it becomes a form where they just need to register information about new systems and AI agents, etc. in the information that this integrated UI's AI concierge references.

In this way, the integrated UI plays a role like a lubricant for communication between employees and the DX/IT department, so for the employee side they no longer get lost in utilizing internal IT, and for the DX/IT department they can concentrate on developing systems and AI.

That said, you might think "Isn't developing this integrated UI difficult?", but rather here it's better to keep high independence from each system and AI agent, so you can start from a form where you first create a list of each system and AI agent list information, and guide URLs along with introducing the functions of necessary systems and AI agents in response to requests from users.

If you start bringing the functions of each system and AI agent to the integrated UI side, it will expand as much as possible and become something heavy and large before you know it, so it's better to position it as covering a thin layer as the role of a hub for guiding each system and AI agent.

It's better to cut it off in a form where authentication, access control, detailed usage manuals, etc. are also once transferred to URLs and then left to each system and AI agent. In any case, it's important to maintain a loosely coupled state with each system and AI agent.

Without having individual specialized functions and maintaining independence, since it becomes a touchpoint with all employees, it becomes important to focus on improving UX, including responsiveness, stability, and screen design.

On the other hand, it's also fine to have a function to widely search internal information. This, without making dedicated elaborate implementations, becomes an internal information search function emphasizing recall rather than precision.

Regardless of accuracy, for the purpose of reaching internal information broadly for now, when you think "I want to do something like this, is there any related material?", it returns information that seems highly relevant across the board, so to speak, like a strengthened version of full-text search within the company by AI.

The most common pitfall in building internal information search apps with RAG is trying too hard to return information pinpoint, and not being able to release because that precision doesn't come out.

When trying to increase precision in a specific area, rather the precision in other areas decreases, and it becomes a state where if you stand here, you can't stand there.

Also, there are many cases where precision came out with current documents, but when documents are updated, precision suddenly drops.

For the employee side, even just widely picking up information that seems related is sufficiently appreciated, so it's good to include an internal information search function emphasizing recall, giving up on precision to some extent, as a function of this integrated UI.

Of course, if you want to search information with high precision (increase precision) for a specific area or purpose, you can develop and release it as a specialized AI agent.

As an image of this integrated UI, Google's announced AgentSpace is a reference. As it becomes a touchpoint with internal apps and AI, and NotebookLM also makes it easy to search internal information, etc. Microsoft's Copilot also has a function to easily access built AI agents, and will be in the same positioning going forward.

Then, you might think "Why not just use these as is?", but there's one caveat.

Since the integrated UI becomes a touchpoint with all systems and AI agents, if you're not careful you can easily get locked in.

As I also emphasized the importance of maintaining a loosely coupled state earlier, if this integrated UI is built with external tools and the degree of integration with internal systems is increased, it will probably become difficult to switch in the middle.

Before you know it, including surrounding tools, that vendor's related products will increase even if unintended. Precisely because it becomes a touchpoint with all systems and AI agents, it's important to increase its independence so as not to get locked in.

Also, since the optimal form of the integrated UI naturally differs depending on that company's business model and employee/organizational structure, it's desirable to maintain a state where the UI can be freely customized.

Vendor tools, for better or worse, have UIs almost fixed so that anyone can use them easily, so if your company's unique custom requirements come up after release, it becomes a state where it's rather more expensive through individual SI.

Also given that recently there are many products with per-employee billing, considering running costs and the risk of lock-in, regarding the integrated UI, it's better to build something unique to your company so as not to depend too much on specific vendors.

I've written at length up to here, but in DX promotion in the era of generative AI and AI agents, "development of specialized AI agents" that maximize your company's strengths, "maintenance of APIs and data infrastructure" that supports that development, and "construction of integrated UI" to not let employees get lost become important.

Summary

How was it?

For better or worse, AI has become widely noted in the world, and information that is not very important about AI has also come to be widely disseminated.

Many people may be confused by the amount of new information daily, including both important and unimportant things, but the points that should be fundamentally grasped in corporate initiatives are not that many.

I would be happy if this helps to thicken the trunk part of DX promotion without being swayed by transient information.