Improving the quality of our IP services with AI

1. Introduction

From July 1, 2026, IPX Patent Partners (“we” or “our firm”) will move the AI framework we have developed and tested in-house into full operation, including filing work. This document was prepared to explain our AI governance honestly, so that clients can feel secure as we make this move.

In bringing AI into our work, we chose to state our AI governance clearly to clients. We did so because (1) we value the trust of our clients above all, (2) we want to be able to ask clients in advance how they would like AI to be used, and above all (3) we do not want any ambiguity about who is responsible for the quality of our work.

2. IPX so far: a firm that has invested in automation

Before explaining how we use AI, here is a brief account of the stance our firm has taken so far.

2.1 Our own Word macros, built since the firm was founded

The repetitive work that comes with drafting a specification, such as format checks, layout, drafting support examples, and generating and updating re-drafted claims and abstracts, has been automated with our own set of Word macros. This has been a continuous effort since our founding and forms the core of our internal know-how. Across our proprietary technology as a whole, including these Word macros, we hold roughly 30 patents.

2.2 A management policy of investing to cut repetitive work

Patent firms normally spend a great deal of labor and cost on repetitive work. To cut that cost and drive down human error, we have invested heavily in training, workflow and systems. As a result, our members can spend more time on what only people can do: grasping the essence of an invention, judging patentability, and designing the scope of the right. This is the foundation that has supported our quality and speed.

2.3 A culture of sharing what we build across the whole firm

Long before AI appeared, we had a culture of sharing the functions we built and the practical know-how we gained with everyone in the firm, rather than leaving them with a few people, and of building together a system in which anyone can produce output at or above our standard. The scope of tool use and the rules for operating them, that is, governance, were also set out clearly, shared with everyone, and reviewed regularly.

To a consistent policy of using IT to raise quality and speed, and of setting rules for it while sharing tools and know-how across the organization and improving them continuously, we have added a new tool: AI.

3. How we came to use AI, and where we stand

3.1 Careful use of AI where unpublished information is involved

For pre-filing matters, which involve unpublished information, we have managed AI use carefully. We limited AI to uses that need no access to unpublished information, such as documenting general matters and searching, summarizing and organizing prior art.

3.2 Continuous research and validation toward full deployment

As a firm that has invested in IT, we began researching how to integrate generative AI into practice as soon as it appeared. In that research we took the greatest care with confidentiality: rather than pre-filing matters entrusted to us by clients, we mainly used matters already filed and applications in our own name to compare quality and efficiency with and without AI. For prosecution responses, we built up evidence through actual cases.

Through this research we developed our own skills, which make AI reproduce our practice procedures exactly; agents, which combine several skills to carry work forward on their own; a harness, which gives those agents firm footing; and AI access rights to our internal systems that reflect each person's permissions. We also studied context, which determines what information and how much of it produces good output when given to AI. To prevent information from being cross-contaminated through memory, we built memory management that gives each client a separate work area and keeps memory completely divided by client. If someone tries to store one client's information in memory belonging to another client, the AI warns them, so human error is less likely. We also keep MCP connections to a minimum.

All of these are proprietary intangible assets that reflect our knowledge and experience deeply. None of them are sold on the market.

3.3 Tuning general-purpose AI rather than IP-specific AI

AI services built specifically for intellectual property have appeared in recent years. They are easy to adopt, but they are hard to customize for individual circumstances, and almost everything about the setup, including the AI model running underneath, depends on the vendor.

We therefore chose general-purpose AI, which gives access to the most advanced models available to any company, rather than AI built for the IP industry, as the base of our AI governance. On top of it, drawing on the IT skills and automation know-how we have built since founding, we tuned that general-purpose AI thoroughly to our own practice and created an AI environment made for IPX.

General-purpose AI evolves quickly, and it has the large advantage of giving access to the world's most advanced frontier models. By combining such high-performance models with our AI governance, we have built a system optimized for our practice that no off-the-shelf specialist AI can reproduce.

3.4 Organizing IP practice into layers

We organize our methods for drafting specifications and responding in prosecution into several layers. Layer 1 is the common rules that every member must follow; layer 2 is practice procedures standardized by client or by field; layer 3 is the individual refinements each attorney in charge has built up.

Layer 1 prevents variation in quality on a foundation common to the firm, layer 2 applies settings customized by client or by field, and layer 3 lets each attorney reflect their own refinements and preferences. Layer 1 is strictly controlled by administrators and cannot be changed by members on their own. Layers 2 and 3, by contrast, can be updated flexibly by designated people through regular exchanges with the members in charge.

3.5 Sharing configuration files and skills across the firm

Configuration files and skills set up in detail by administrators, and skills that individual members develop and improve themselves, can both be shared with everyone in the firm through simple operations, without GitHub. It is a way to pass settings and skills around internally with ease.

One person's know-how or improvement is immediately shared as a practical skill for the whole organization, and it can be handed over between members when needed. This way of sharing results with everyone follows the culture we had before AI.

These are the reasons we concluded that the conditions were in place to enter a full deployment phase for AI.

4. IPX from here: our structure for the AI era

Building on this research, from July 1, 2026 we will use AI in practice under the following structure.

Generative AI is good at turning a direction into a finished document once it is told which way to go. Identifying the point of an invention, setting the approach to claims and drawings, and picturing how a right will be used still require a patent attorney's expertise and experience.

Unless a client asks otherwise, a patent attorney first sets the skeleton and approach of the specification, including the approach to claims and drawings, and AI then writes the embodiments along that approach. The attorney in charge examines and checks the draft AI produces, and completes the final work product.We never deliver what AI produced as it is. Responsibility for final quality rests with the patent attorney, as it always has.

At the same time, we will go on organizing and accumulating the IP know-how we have built, as skills. We will use them to deliver IP services that reflect what only IPX knows, at a higher quality than before.

5. The AI models we use, and safety

5.1 Client information is not used for training

Every AI we use is on a corporate contract, that is, an enterprise plan. Use of data for training is automatically opted out at the time of contract, and opting in is not possible. Client information sent to AI is therefore never used to train AI models.

We also clearly prohibit the use of free or personal AI accounts for work, that is, shadow AI. Shadow AI usually appears when people cannot use AI at work, or when the AI they can use is not the one they want. We provide such a wide range of AI models that there is no need to bring in shadow AI. This structurally removes and reduces the risk of client information leaking through an unexpected route.

5.2 What we pass to AI is judged against internal standards

As noted above, every AI we use is an enterprise version with training opted out, and client information is not used for training. On top of that, we judge carefully, against internal standards, what information may be passed to AI, according to the confidentiality and nature of the matter. If a client has a particular request, such as handling a certain matter without AI or using a specific AI, we will discuss it individually and accommodate it.

5.3 We review our rules and guidelines regularly

The technology, law and society around AI are changing fast. We do not treat our rules and guidelines for AI use as fixed; we review them regularly and revise them as needed. When a new risk or a better way of working comes to light, we reflect it in our structure promptly. If there is a material change to the policy we have shown clients, we will inform you again.

5.4 Preparing to introduce local LLMs

To be ready for measures such as bans on the use of certain AI models arising from geopolitical factors, we are preparing to introduce local LLMs, which complete all processing inside our own environment without going through outside servers. We aim for a structure that uses closed models, including frontier models, together with local LLMs based on open models, choosing between them as the situation requires.

6. People and training

It is not true that anyone can produce the same result simply by using AI. Using AI output as it is and using AI as a tool to raise quality are different things. The quality of the final output depends on whether you understand the limits and characteristics of AI and can apply it to the context of real practice.

6.1 Hiring that puts weight on expertise and literacy

We have hired only members whose IT literacy is above a certain level, in every role: patent attorneys, translators, drawing tracers, searchers and others. Our organization was built on a high level of IT literacy long before we used AI.

6.2 Frequent information sharing

At least once a week, everyone in the firm shares the latest examples and failures in AI use, along with newly built skills. We also keep up with the latest information on AI, such as model advances, new features and new risks, and pass it quickly to the members who need it. For that latest information, we have built a system that uses AI to add new stories to a database automatically.

AI moves extremely fast, and it is not unusual for a best practice from a few weeks ago to be out of date already. What we learn is reflected in our internal rules and skills as we go, so the organization always keeps the newest and best way of working.

6.3 Support for personal AI research

We also have a support program for members who pursue AI research outside work. The AI environment used for personal research is clearly separated from the one used for filing work, and the personal research setup is never used for filing work.

7. Delivering further value

7.1 Not only efficiency, but value that only AI makes possible

Using AI has made our existing work more efficient. In back-office work in particular, we have cut a great deal of time. For clients, we will go beyond efficiency and deliver value that only people and AI working together can produce.

7.2 AI is a tool that extends a patent attorney's expertise

Our use of AI does not replace a patent attorney's expertise. It is positioned as a tool for raising the quality of that professional judgment. Understanding the invention, judging patentability, designing the scope of the right, and preparing drawings that are necessary and sufficient remain the expertise of the attorney; on that basis, we aim for still higher quality and speed with AI.

7.3 Greater value delivered to clients

With AI, the density of our quality checks, the coverage of our searches and our response speed all improve beyond what they were. Higher quality, delivered faster. That is our promise.

8. Our promise to clients

Finally, here is what this document has said, in four points.

① Enterprise AI only	We use only enterprise AI with training opted out, and we protect client information.
② Delivering further value	The purpose is not cost cutting through efficiency, but raising quality and speed further.
③ The attorney's responsibility is unchanged	A patent attorney always examines and checks AI output and guarantees the final quality.
④ We keep learning as an organization	Through information-sharing sessions, in-house development of skills and our research support program, everyone in the firm keeps improving how we use AI.

If you have questions or concerns, or a specific request such as “please handle this matter without AI,” please tell us at any time.

We will continue to work toward services of even higher quality. We look forward to your continued trust in IPX Patent Partners.

Everyone at IPX Patent Partners