May 22, 2025

Following on the heels of the generative AI that is taking the world by storm, Agentic AI is a new technology that is drawing attention. It is a system in which the AI can think and work autonomously just by specifying a goal without issuing detailed instructions. NEC recently announced that they have built some of the functions of this Agentic AI (Note 1). To learn more about the nature of this groundbreaking system and the advantages that it provides, we spoke at length with an NEC researcher.

Agentic AI thinks about the tasks required to fulfill the instructions and automatically finds the necessary information

―Please tell us more about the Agentic AI recently announced by NEC.

With an Agentic AI, the user only needs to input a goal for it to act autonomously and deliver results. This is an area of active research worldwide. The Agentic AI that we have developed is based on this technology which can be broadly and universally used. However, for this project we built a system that focuses on information search and report creation as an application that is expected to meet the needs of many people.

―The main functions are searching for information and creating reports, but how does Agentic AI differ from generative AI?

Text generation using generative AI is fundamentally relied on information contained within the AI. In contrast, the Agentic AI that we have developed is able to go and search for information scattered around the Internet and internal web sites on its own as needed. For example, if a user specifies, "I would like to check what rules exist within the company regarding overseas business trips," the AI will autonomously reference various databases and answer the question. The user does not need to enter search keywords or specify the search destinations. Just by entering the goal, the AI will think about the search keywords and destinations, execute the search, and effectively retrieve the required information that is scattered across various sources. When answering, the AI shows the reference source, which enables the user to verify the reliability of the information.

In addition, the system can also handle more complex instructions. For example, if you tell the AI, "Survey the companies that compete with NEC," it will first need to search for and create a list of competitors and then summarize the information about each company in a two-step process. Agentic AI is able to handle instructions even when they generate multiple tasks such as these. The AI appropriately divides up the tasks and executes them in the correct order.

Utilizing the unique know-how of NEC cultivated through AI and LLM development

―What kinds of know-how and technologies were used to achieve complex processing?

There are three main categories. The first category is the know-how cultivated during the development of the large language model (LLM) which forms the core of the cotomi generative AI developed by NEC. An LLM is a system that behaves in unique ways and requires special skills to get it to do what you want. In addition, the LLMs made by each company have different characteristics, and their behaviors can significantly change between consecutive versions. You have to devise ways to make the models run well according to each of their characteristics. Accordingly, NEC is sufficiently aware of how to handle its own LLM and thoroughly researched other LLMs to develop its own. I believe that this enabled us to accumulate know-how regarding the handling of LLMs, which led to the development of this technology.

The second category is the task decomposition algorithm. For the task decomposition, we considered two methods. The first method builds the overall flow and then proceeds in a top-down manner while the second method achieves the goal by handling each individual step in a sequential manner. Ultimately, we developed a hybrid approach that combines both methods. Initially, we proceeded with a method that designed the overall flow in a top-down manner, but we realized that with this method, the AI was unable to recover when it failed and would not reach the goal. That being said, in a flow where the tasks are carried out sequentially in a comprehensive manner, you cannot see what lies ahead, and it is impossible to predict whether you will reach the goal or not. Therefore, we are developing a new algorithm that designs the overall flow in the beginning and then reviews it as needed. This technology was summarized in a research paper and announced at IEEE BIG DATA 2024, a top international conference, in December of last year.

The third category is a mechanism that presents the information sources. This is also related to the know-how concerning the LLM characteristics mentioned earlier, and we made many adjustments to design it to be able to output the appropriate source as expected. The main purpose of this function is to prevent LLM hallucinations. However, NEC is developing other technologies to suppress hallucinations, so it is possible to combine these during implementation.

One more point that I would like to emphasize is the user interface (UI) of this system. We were quite particular about this until right before the release. Because the key to this system is that you only need to input the goal. You do not need to painstakingly write out the procedure. Our goal with the UI was to create a design that would enable users to understand a new system such as this in an intuitive way. In addition, it is designed not just to display the results during output but to also review the steps that it followed. This allows the user to review the process in the same way that a manager reviews the work of a subordinate. They can also realize, "Maybe this part of the instructions was wrong" and discover points for improvement.

Creating an AI that supports daily tasks by linking to internal systems

―What are the expected use cases?

I believe that it can be used for very generic purposes. For example, it can be used by a salesperson to draft a written proposal when thinking of making a new proposal. When preparing a proposal for Company A, if a user entered “I want to create a sales proposal document for Company A,” the system could investigate relevant information about Company A and industry trends while simultaneously summarizing what proposals were made internally in the past to create the main points of the written proposal. It can also show the items that should be emphasized in the proposal. As a result, even new employees and other workers with limited experience can instantly gather only useful information from across the web and prepare the basis of a highly accurate proposal.

In addition, while it is difficult from a confidentiality standpoint to use an AI agent running in the public cloud to search internal information, NEC's Agentic AI can also be installed on-premises. Compared to large, cloud-based systems, we are building a more lightweight system that is being productized to provide it in a packaged form. Based on the "client zero" approach of putting cutting-edge technologies into practice with NEC itself as the zeroth client, we started using the Agentic AI within NEC, where it has been well received by the employees.

―What do you think you will be able to achieve in the future with this technology?

That's a good question. Eventually, I would like to build a system that can not only search for and display information but also execute actions. For example, if you instructed the system to create a business trip itinerary, it would make the arrangements and reservations for your passport and the hotel. If this were to happen, I think we could depict a future in which the Agentic AI assists the user like a concierge while working together.

In addition, I think that we also need to consider collaboration and automatic negotiation between Agentic AIs. For example, a settlement of interests would be needed when Company A's agent wants to sell a product for as much as possible and Company B's agent wants to purchase that product as cheaply as possible. To handle such cases, we probably need to develop a protocol that makes mutual adjustments.

In any case, what we have created so far only realizes a subset of our intended vision. Our ultimate goal is to realize a system that acts autonomously and produces results once the goal is determined. As of right now, the application specializes in search and report creation functions, so I would like to gradually expand the scope of functions that it supports. It would be great to develop it into a system that a wide range of users could use on a routine basis to work in a more creative way.

Go back to Featured Technologies