With the widespread adoption of generative AI, such as ChatGPT, competition is intensifying worldwide to develop “AI agents”—systems that can understand objectives, formulate plans, and autonomously perform tasks by interacting with external tools.While “patents” are the means to protect this cutting-edge technology from imitation, many developers and executives have questions such as, “Can AI agents really be patented?” and “What kinds of patents are other companies securing?”
In this article, a patent attorney with expertise in the AI field will provide an explanation by comparing actual patent cases and official examination examples from Japan (JPO), the United States (USPTO), and Europe (EPO).Using verified real-world examples—such as Google’s Transformer patent, OpenAI’s tool integration patent, and DeepMind’s external memory patent—we’ll outline specific points for securing patent rights for AI agent technology.
Table of Contents
First and foremost, AI inference algorithms (mathematical formulas) themselves are generally not patentable in any country. However, if that information processing is configured as a technology specifically implemented using hardware resources such as computers, it becomes patentable as a “software-related invention.”
In the case of AI agent technology, the main focus of patenting efforts lies not in improvements to the model itself, but rather in the system architecture and processing flow—specifically, “how LLMs were utilized to build an autonomous system.” Below, let’s take a closer look at specific examples of what has been patented across three jurisdictions (Japan, the U.S., and Europe).
To understand the cases in each country, let’s review three common perspectives.
| Perspective | Japan (JPO) | United States (USPTO) | Europe (EPO) |
|---|---|---|---|
| Basic Approach | Concrete information processing using hardware resources | Alice/Mayo Test (Whether the Idea Is Abstract) | Technical Nature + Technical Contribution (COMVIK) |
| Key Questions | Is the information processing concretely implemented? | Is it a technical solution to a technical problem (practical application)? | Are the technical features conducive to the technical objective? |
| Weak Structure | Mere presentation of data or the automation of human tasks using AI | Execution of abstract ideas on a general-purpose PC | Mathematical or business methods “as such” |
In Japan, AI inventions are examined as a type of “computer software-related invention.” The core criterion for determination is whether “information processing is concretely implemented by software using hardware resources,” as specified in Appendix B of the Examination Handbook.
The Japan Patent Office (JPO) published an official collection of AI-related case studies in January 2019 and added 10 new cases in March 2024 to address the era of generative AI and large language models (LLMs). The conclusions for representative cases are as follows (all are hypothetical cases presented by the JPO).
JPO Case: Granted
A case where “prediction accuracy was significantly improved” by a new input variable (estimation of hydroelectric power generation)
While a configuration that merely replaced a conventional regression model with a neural network was deemed to lack inventive step, a configuration that added “river water temperature” as an input variable to account for snowmelt inflow—thereby significantly improving prediction accuracy—was judged to possess inventive step. The key lies in technically meaningful innovations regarding input features and training data.
JPO Case Study: Rejection (NG)
Case of Simply Replacing Human Work with AI (Calculating Cancer Severity)
An approach that merely replaced the manual calculation of cancer probability based on blood markers—previously performed by physicians—with a pre-trained neural network was deemed to lack inventive step, as it was considered “merely the automation of human tasks using AI.”
JPO Case Study: 2024 · LLM
Inventive Step in Configurations Using Generative AI (LLM)
While a system that automatically generates answers by inputting questions into an LLM (such as automated customer service responses) was denied inventive step, a specific process that extracts multiple keywords from relevant documents to generate prompts and produce more appropriate text was granted registration on the grounds that it provided a significant effect.
Let’s look not only at the Japan Patent Office’s hypothetical examples but also at actual registered patents.
Actual Patents | Japan
JP 7282070 (Mitsubishi Electric) | “Compressed Data” for Neural Networks
A patent regarding data that compresses the configuration information of a neural network through quantization (registered in 2023). This is an example of securing intellectual property rights for “data structures” in the AI field.
Actual Patents | Japan
JP 7177608 (Komatsu) | Pre-trained Position Estimation Model for Construction Machinery
A system patent (registered in 2022) using a pre-trained model to estimate the position of a hydraulic excavator’s working equipment based on camera images. This is an excellent example of securing rights not only for the system and method but also for the manufacturing method of the pre-trained model and the training data, serving as a reference for autonomous machine control (an area adjacent to AI agents).
In the U.S., patent eligibility (35 U.S.C. §101) is determined using the two-step Alice/Mayo test. The key questions are: “Is the invention directed to an abstract idea?” and “If so, does it include an inventive concept (a specific application to a technical problem) that goes beyond the abstract idea?”
On July 17, 2024, the USPTO published examination guidance specifically focused on AI and provided three concrete examples (Examples 47–49).
| USPTO Examples | Ineligible Configuration | Eligible Configuration (Registerable) |
|---|---|---|
| Example 47: Anomaly Detection | Method for training a neural network (application of mathematical formulas only) | Specific Application: Real-Time Blocking of Malicious Packets Using a Pre-Trained Neural Network |
| Case Study 48: Speech Segmentation | Simply Calculating Embedded Vectors Using Mathematical Formulas | Specific Process for Separating Speaker Voices Using Clustering and Masking |
| Case Study 49: AI in Healthcare | Simply Calculating Risk Scores from Genetic Data | A specific application: administering a specific treatment (eye drops) to high-risk patient groups |
The lesson is clear. The key to securing rights for AI agent inventions in the U.S. is to specifically demonstrate a “technological solution to a technological problem,” along with how the AI “operates” and “what it improves.”
Actual Patents | U.S.
US 10,452,978 (Google) | Transformer Patent
A patent (granted in 2019) for a neural network based on the self-attention mechanism, corresponding to the paper “Attention Is All You Need.” This architecture forms the foundation of all modern LLMs and AI agents.
Existing Patent | U.S.
US 11,922,144 (OpenAI) | External API Integration (Tool Utilization)
A patent (registered in 2024) in which an LLM reads an API schema (manifest) and generates function calls to external tools (such as shopping, databases, and email) without the need for retraining. This is a core patent for “AI agents,” corresponding to ChatGPT’s plugins and function calls.
In Europe (EPO), computer programs and mathematical methods are excluded from patentability “as such” (Article 52 of the European Patent Convention). However, this exclusion can be easily circumvented by referring to hardware (computers, processors, etc.), and the real challenge lies in the assessment of inventive step (the COMVIK approach).
Under the COMVIK approach, only features that contribute to the technical nature of the invention are considered in the assessment of inventive step; features that serve purely mathematical or business purposes are not taken into account. In other words, even if a feature is mathematically novel, it makes a “zero” contribution to inventive step if it lacks a technical purpose.
EPO Case: Rejection Not Accepted
T 0702/20 (Mitsubishi Electric) | Sparsely Connected Neural Network
Regarding a neural network structure that uses error-correcting codes to create sparse connections between layers, the Board of Appeal acknowledged that it was “novel and non-trivial” but rejected the application on the grounds that it merely defined a class of mathematical functions (an exclusion).This is an important case demonstrating that an objective such as “preventing overfitting” alone is insufficient, and that it is difficult to obtain protection in Europe for improvements to core AI.
EPO Case: Key Ruling
G 1/19 (Enlarged Board of Appeal, 2021) | Pedestrian Crowd Simulation
Regarding computer-based simulations of pedestrian crowd movement, the Enlarged Board of Appeal ruled that “computer-implemented simulations are assessed using the same criteria as other software inventions” and that “numerical output can constitute a technical effect.” This provides guidance for simulation and planning-type AI inventions.
Actual Patents | Europe
EP 3398117 (DeepMind) | Neural Network Extended with External Memory
A European patent (registered in 2023) for an architecture that combines a neural network with external memory (storage). This registered case is directly linked to the “memory” function of AI agents, which retain information and utilize it for subsequent actions.
| Comparison Items | Japan (JPO) | U.S. (USPTO) | Europe (EPO) |
|---|---|---|---|
| Decision-Making Framework | Concrete Implementation Using Hardware Resources | Alice/Mayo Two-Step Test | Section 52 + COMVIK (2-stage) |
| Core AI (Model Improvement) | Acceptable if Ingenuity + Significant Effect is Demonstrated | Acceptable if technical improvements are demonstrated | Strict (Rejected in T0702/20) |
| Type that is easy to obtain a patent for | Specific processing flow and novel input features | Specific application to the technical problem | Linked to the technical objective and technical implementation |
| Official Guidelines | AI Case Studies (2019, 2024) | AI Guidance Case Studies 47–49 (2024) | Guideline G-II 3.3.1 and G1/19 |
| General Trends | Relatively Lenient | Tightened after Alice → Improved predictability by 2024 | Most Stringent Requirements for Technical Contribution |
Common Essence: What the three regions have in common is not “what AI achieves (business idea),” but rather the need to specifically describe “how to technically achieve it.” This single point determines the success or failure of AI agent patents.
Here we summarize the practical points for successfully obtaining AI agent patents, derived from cases in Japan, the U.S., and Europe.
① File the application before public disclosure: Disclosure via academic papers, open-source software (OSS), or press releases will result in a loss of novelty. Ensure the application is filed before any such information is made public.
② Describe the “Process Flow”: Rather than focusing on prompt text or business ideas, describe the system as an information processing procedure—specifying how it behaves dynamically under specific conditions.
③ Clearly state the technical problem and its solution: Present technical problems—such as “how to prevent hallucinations”—alongside the specific means to solve them and the significant effects achieved.
④ Avoid over-reliance on specific technologies: Claims that are entirely dependent on specific external APIs (such as OpenAI) will lose their protection if the model is replaced in the future. Extract your company’s core value using universal concepts.
⑤ Align with each country’s examination standards: The approach to drafting claims that are “effective” differs between Japan, the U.S., and Europe. For global applications, a strategy well-versed in the practices of each jurisdiction is essential.
Patenting AI agent technology—start by consulting with us.
Patent attorneys with expertise in the IT and software fields provide comprehensive support—from technical consultations and prior art searches to patent prosecution strategies in each country. We offer a free initial consultation.
Schedule a Free Initial Consultation IT & AI Intellectual Property ServicesQ. Can AI agents be patented?
A. Yes, they can. While the inference algorithms of AI agents themselves are mathematical methods, if they are described as systems or methods that specifically implement information processing using hardware resources such as computers, they may be eligible for patent protection in Japan, the U.S., and Europe.In fact, numerous foundational technologies for AI agents have been patented, such as Google’s Transformer patent (US10,452,978) and OpenAI’s tool integration patent (US11,922,144).
Q. Can you obtain a patent based solely on the design of a prompt?
A. In reality, it is difficult to do so. The prompt text itself is often regarded as a “human instruction (presentation of information),” and there is a risk that it will not be deemed an invention.In a 2024 case from the Japan Patent Office, a configuration that simply inputs a question into an LLM to generate a response was found to lack inventive step. On the other hand, the likelihood of obtaining a patent increases if the method is described as an information processing process that dynamically generates and synthesizes prompts based on context.
Q. Which region—Japan, the U.S., or Europe—is the easiest to obtain a patent in?
A. While it’s difficult to generalize, the registration requirements for software inventions are generally assessed as follows: “Japan is relatively lenient, while Europe (EPO) has the strictest requirements regarding technical contribution.”While the U.S. standards became stricter following the Alice decision, the July 2024 AI Examination Guidelines (Examples 47–49) have increased predictability in practice. In any of these countries, explicitly demonstrating that “a technical problem is solved by technical means” is key.
Q. Are there any well-known examples of AI agent patents?
A. Representative examples include Google’s patent (US10,452,978) for “Transformer,” which forms the foundation of modern LLMs,OpenAI’s patent on external API integration (US11,922,144), which corresponds to ChatGPT’s plugins and function calls; and DeepMind’s patent on external memory-extended neural networks (European EP3398117).In Japan, patents such as Mitsubishi Electric’s patent on neural network compression (JP7282070) and Komatsu’s patent on pre-trained models for construction machinery (JP7177608) have also been registered.
Q. Can I file a patent application even after publishing the technology in a paper or on GitHub?
A. As a general rule, once a technology is made public, it loses novelty and cannot be patented. Japan has an “exception to loss of novelty” (Article 30) system that provides relief under certain conditions, but the requirements vary by country, so it is not a universal solution. Given the rapid pace of disclosure in the AI industry, it is extremely important to complete the patent application before publishing a paper, releasing open-source software, or issuing a press release.
Q. Should I consult a patent attorney regarding inventions involving AI agents?
A. We strongly recommend it. Patenting AI agents requires a high level of expertise to extract “patentable technology” from systems that tend to be black boxes and to design claims (scope of protection) in accordance with the examination standards of each country. By consulting with a patent attorney well-versed in the IT and software fields, you can build a strong patent portfolio that is difficult for competitors to circumvent.