Beyond the blank page: Rethinking workflows when drafting patent applications with AI-assistance

Drafting software tools are one of the most ready AI use cases for patent attorneys

Search tools leveraging Artificial Intelligence (AI) for prior art documents have seen widespread use for several years now. With the advent of Large Language Models (LLMs) about three years ago, another application of AI became available in the intellectual property field.

For many patent attorneysFor simplicity, the term “patent attorney” is used throughout this article, but the content equally applies to patent engineers and similar roles involved in drafting, and to all genders., using off-the-shelf LLMs such as OpenAI’s ChatGPT became a useful addition to their workflows when uploading and analysing documents was introduced mid 2023. ChatGPT allowed for summaries of prior art documents or comparisons of the technical content of documents. While results were impressive, these early applications also highlighted accuracy problems such as hallucinations (the LLM lacks context to answer a question but nevertheless presents a wrong answer as factually correct) and limits of reasoning and technical “understanding” the current technology offers. Knowing how to instruct an LLM, providing its role and the data to analyse, in short “prompting”, became an essential skill when using LLMs.

In 2023 first dedicated software products for patent attorneys were launched, initially mostly for patent application drafting, but today basically for any reoccurring task in the IP field, from responses to office actions to licensing.

From testing many tools himself and keeping up with the latest benchmark studies, the author is under the impression that today the use case of Invention Disclosure support tools for inventors is ready for widespread use in companies, given that confidentiality requirements are met. Here, the LLM assistance can for example include writing aid for inventors that are not native speakers of the language required for the invention disclosure or asking inventors clarifying questions to enhance the quality right when the invention disclosure is put together. It remains a legal requirement though that the technical solution section must be conceived solely by a human contributor, as some jurisdictions have already clarified that an AI tool cannot be an inventor of a patent.

For responses to office actions, freedom-to-operate studies and oppositions, a different skill set than generating supporting disclosure is demanded from LLMs. Here, the software tools must attempt a feature-by-feature analysis of various documents. AI-based tools for these tasks are useful additions to any patent attorney’s workflow, but one can often recognize that limitations of the technical “understanding” of the AI tools are more obvious than in drafting.

For patent application drafting the interested reader can find several dozen software solutions in the market as of summer 2025. When generating text for the general as well as figure description of a patent application draft, the software tools have proven to be very helpful in allowing for more technical depth and time-efficient drafting by experienced users. Drafting therefore remains the prime example of LLM usage by patent attorneys. Therefore, this article focusses on providing an overview and insights into AI-assisted drafting of patent applications.

Approaches to AI assisted drafting

When utilizing LLM technology for drafting, users can apply off-the-shelf LLMs available on the internet, often even for free. However, confidentiality of a client’s invention might not be sufficiently secured in such an environment even if users select options that uploaded information will not be used for training purposes for next generation models. Therefore, most patent attorneys tend to use such tools solely for unproblematic requests such as providing text suggestions for a technical background section or definitions of technical terms. This way the full extent of the invention and its context is not uploaded.

More tech-savvy patent attorneys have turned to local installations of LLMs after Meta’s Llama models became available for free, thereby circumventing any confidentiality issues as no client data is uploaded anywhere. Due to the required knowledge and time for setting up and regularly updating such a system, as well as working with it in command line mode, this approach might however not be feasible for most IP specialists.

Therefore, users today often turn to dedicated software tools, often termed “AI Co-Pilot” (highlighting that the patent attorney is still in the pilot seat). Such tools typically provide easy-to-use graphical user interfaces and guarantee, to various extents, that confidentiality concerns are met.

A structured conceptual overview of the current Co-Pilot landscape

Before choosing a software provider, patent attorneys should comprehend the various concepts of current AI Co-Pilots. They can best be understood by categorizing them in two dimensions: User Interface and Type of LLM that’s used.

User Interface Concepts

Turning first to the type of user interface employed, three different approaches can be distinguished: Microsoft Word Plug-Ins, Custom Websites and Full Editors.

The Word Plug-in model integrates directly into a familiar and widely used software environment. For patent attorneys, this minimizes the learning curve and allows seamless adoption within existing workflows. Such plug-ins typically lack advanced drafting features, such as editing of figures or automated generation of flowcharts.

Custom Websites offer a leaner interface that is still easy to use, with browser-based access. Like the plug-in, they maintain a straightforward user experience but usually don’t allow for figure creation.

The Full Editor represents the most comprehensive and capable interface. It supports the entire drafting workflow from uploading invention disclosure and prior art input, creation of text and figures within the tool and finally export of a finalized patent application. Full editors can include automatically drafting flowcharts and simple system diagrams based on claimed features (extracted from the draft), automated generation of figure descriptions via computer-vision as well as manual figure editing plus managing of element labels.

LLM installation

The second aspect to consider is where the LLM that the AI Co-Pilot uses is installed.

Locally hosted LLM tools offer the highest level of data security, as all information including confidential invention disclosures remains offline. Additionally, local installations can technically support offline use, although in practice this is rarely required. However, these systems are often complex to install and maintain. Users require above-average IT skills. Updating to a newer model requires careful revalidation of prompts and workflows to ensure consistent output quality. Performance can also be a limitation, since hardware constraints of typical office laptops may result in quite slow response times of the LLM.

Agnostic systems offer a middle ground where the LLM may be cloud-based or locally hosted. For the local installation, the setup is essentially the same as described beforehand. When the user instead uses API access (a predefined interface) to a cloud-based LLM service, typically a “off-the-shelf” model is connected that is not specifically adapted to patent drafting. This user selection of the model to be connected has sometimes been adapted by providers of existing Drafting Tools that have enhanced their Drafting-specific editors with AI support. The text generation speed of cloud-based LLMs can be close to real-time, allowing for a very interactive mode of iterative generation and human revision.

Cloud-based Co-Pilots have a built-in connection with one or several LLMs that often have been finetuned (a pre-trained LLM is further trained for a specific task) to the patenting use case by the software provider. All technical issues such as updates and integrating new functionality are managed.

While concerns around confidentiality for uploaded client data have to be taken seriously by patent attorneys, many organizations are already accustomed to working in cloud environments and rely on tools such as Microsoft Office365, Teams, and Azure, without ever questioning data security. Software providers of AI Co-Pilots hence have the same responsibility for data security as Microsoft does for their cloud services. They must ensure strong encryption, data segregation between cases and clients as well as a zero-data-retention policy. All of these requirements can be checked by independent IT Audits initiated by the provider, offering patent attorneys a solid basis for trusting the services.

Considering ease-of-use, text output quality and text generation speed, cloud-based Full Editors currently appear to be the best choice for most organizations and therefore continue to get a lot of attention. Such tools are often amongst the top-rated Co-Pilots in benchmarking studies when it comes to the quality of automated text generation, because the software providers significantly adapt the used models to the patent drafting use case.

Before selecting an AI Co-Pilot, patent attorneys should also consult legal requirements and professional guidelines.

The EPI guidelines on use of Generative AI in the work of patent attorneys

Concerns regarding client data confidentiality have already been addressed by the EPI, providing general guidance for European Patent Attorneys. The guidelines of November 2024 do not favour any mode of operation (local or cloud-based) but clarify on a more abstract level what to consider when drafting with AI. The guidelines essentially state that the patent attorney signing off on work products such as a patent application retains full responsibility for quality and factual correctness, no matter how a document was created. This was true before AI when trainee patent attorneys and paralegals might have been involved, and it is true now when parts of the text generation are automated with AI Co-Pilots. Patent attorneys must make sure that clients are informed on the intended use of AI tools in their case, as well as that the chosen AI tools guarantee data confidentiality and that uploaded data is not used for training purposes. Furthermore, the usage of AI tools doesn’t have to be disclosed to patent offices, unless regulation other than the epi guidelines demand it.

Rethinking Drafting Workflows when integrating an AI Co-Pilot

The Introduction of AI Co-Pilots presents an excellent opportunity to rethink established workflows and processes. Instead of starting with a blank page in a word processor, patent attorneys can adapt their drafting workflow to leverage the AI capabilities of the tools. The author presents an exemplary workflow in the following paragraphs to highlight practical considerations when using a cloud-based Full Editor in the patent drafting process.

Building context

The first step is to build context for the LLM to “grasp” the technology. This includes uploading the invention disclosure and prior art as well as existing figure sketches (if available), and then completion of figures and figure descriptions with AI support. Once this context has been built, text generation quality is in general good. How context is built will be explained in more detail in the following paragraphs.

After uploading the invention disclosure, clarifying questions can be generated with the LLM. The questions are then either asked to inventors, or the patent attorney fills the gaps in the disclosure to the best of his or her knowledge. Additionally, the Co-Pilots can automatically add disclosure for alternative applications and embodiments based on the original disclosure, enhancing it further. Next, known state of the art documents can be uploaded or searched with an AI search tool. This adds further to the context that’s available for the AI.

After that, figure drafts can be created in the tool by either uploading existing figures, manually sketching them in the editor or automatically generating them with AI. The latter has seen major improvements in the past months, today allowing for not only generating flowcharts and simple system diagrams in simple “boxes and arrows” style, but even converting any colour picture of a machine to a simplified black and white line drawing in 3D. While not being perfect, the results are often impressive and serve as a solid starting point for quick and easy improvements by the human in the loop. Finally, reference numerals should be added to the figures and the final figures can then be analysed with the computer-vision functionality available in many current multi-modal LLMs. Making use of the context that’s already there, the tool can provide an initial text description of the depicted features. The quality of these automated interpretations varies greatly, depending on the available depth of the context and the complexity of the figures themselves. Therefore, it is advisable to critically review the figure descriptions and correct and expand were necessary. Once the figure description is correct and complete, the context has been sufficiently built.

Claim generation

With this information being available, claims can be generated for several claim categories. The author observed in his own practice that after the previous steps had been carefully taken, technical accuracy of the automatically generated claims is usually quite good, e.g. both for mechanical engineering and computer-implemented inventions. That means that logic and patent language are reasonably good, there are no clarity issues, and some generalizations are introduced automatically (e.g. from “car” in the invention disclosure documents to “vehicle”). Interestingly, also concepts like thresholds for parameters are often introduced in a meaningful way despite not being explicitly disclosed in the uploaded documents. On the negative side, the wording appears to be not as “on point” compared to an experienced human drafter. Another interesting observation is that all tools known to the author tend to introduce too many non-essential features into the independent claims, explaining too much and therefore potentially rendering the scope of protection too narrow. While this flaw might be overcome by appropriate prompting to some extent, it remains that an experienced patent attorney might chose a claim structure based on experience in prosecution, oppositions and enforcement as well as discussions with the client regarding competitor products and strategies. Such considerations today often can’t be easily verbalized, if at all, and added to the context for the AI Co-Pilot. Consequently, currently it is best practice to take inspiration from automatically generated claims but essentially the patent attorney should heavily rephrase and restructure the claim set to assure the highest quality.

Description generation & templates

For the various sections of a patent application, such as technical background, problem statement, advantages of the claimed solution, support for the dependant claims (see Deep Dive 1) and figure descriptions, specific prompts can be engineered to steer the AI Co-Pilot towards the desired output. In some AI Co-Pilots, these instructions are stored in prompt libraries while others combine various prompts for several sections in so-called patent application templates; both approaches amount to essentially the same benefit of customized text generation.

This customization is further enhanced if the patent attorney adds text passages from existing, manually crafted patent applications, to each specific prompt. The examples guide the AI on the level of detail, style, wording, structure and reasoning expected in the output. With such detailed instructions, a patent attorney might construct sets of prompts (or a template) that is customized for one client based on text examples taken from previous patents of this client. As a result, the work product will be close in style to what the client is used to, presumably making manual quality checking easier. Another noteworthy effect can be that teams of patent attorneys serving one client can produce patent applications that are more uniform in style, adding to the effect of easier quality checks. Similar benefits are achievable when prompts or templates are designed not only for specific clients, but also for technical fields, e.g. differentiating between mechanical engineering and computer-implemented inventions.

Quality & Proofreading

In the author’s experience, overall initial quality of a section generated by an appropriately prompted tool (that is given the context as explained above) can be in the ballpark of 80% directly useable text, while the remaining 20% must be revised with human supervision.

Once the application text body is generated, a review has to take place to check for clarity, factual correctness, completeness and a logical “storyline” progressing from embodiment to embodiment. With cloud-based Full Editors, the speed of text generation is so fast that a back-forth-interactive review is possible. The patent attorney may for example identify a certain text passage to be incomplete and hence prompt the AI to add a few sentences, e.g. on how exactly the system could be set up technically if a data analytics step was done in a cloud environment instead of a local device. As a result, an iterative improvement of the disclosure can be performed until the quality reaches the level that the human expects. Alternatively, patent attorneys can of course rewrite the description manually.

A quite interesting functionality in some AI Co-Pilots is the capacity to perform a wide range of quality checks automatically. This can include highlighting technical terms that might not be sufficiently defined or profanity that should be rephrased. Again, these functions can be customized by the user to facilitate proofreading. Alternatively, one might perform similar checks with any properly prompted LLM installation instead of an AI Co-Pilot. The author suspects that context-sensitive automated proofreading is a very valuable workflow improvement even for patent attorneys that do not want to generate patent applications with AI support (yet?). It goes beyond the ubiquitous spelling and grammar checks performed by word-processors that have become a standard in our profession. Given the ease of use of spell-checking features, today’s patent attorneys would never send a draft with many misspelled words to a client. Starting now, patent attorneys might similarly avoid issues such as terms that are not defined properly and insufficient support for an embodiment, all of which could have easily been spotted automatically with LLM support before filing.

Advanced workflow improvements

Once a patent attorney has mastered drafting with an AI Co-Pilot, how can he or she leverage his or her new knowledge further? The following paragraphs suggest two improved workflow ideas: Drafting in three sessions and Collaborative Drafting.

Drafting in three sessions

A common problem for many patent attorneys, both in private practice and inhouse, comes with the time requirements for manual drafting. It is often difficult to carve out several consecutive half-days, let alone full days of focus time in one’s calendar. The resulting overall preparation time for a draft tends to stretch out in many short sessions over weeks, which is inefficient because the patent attorney will need significant time to reread and freshen up his or hers understanding of the technology and the draft structure each time a new drafting session is commenced. With the efficiency gains made possible with automated text generation, this process can be simplified and sped up.

In a first session, typically a few hours long, the patent attorney can read the invention disclosure, the state of the art and then build the context for the AI Co-Pilot as described above. Uploading documents, converting images and placing reference numerals are steps that can even be performed by appropriately trained paralegals. Once the figures, initial figure descriptions and the claim set are ready, the first session is completed.

At the beginning of the second session, the patent attorney discusses the claim set and the figures with the inventors. After receiving feedback, claims and figures can be revised if necessary. Then all other sections of the patent application draft can be automatically generated while the human makes sure that all instructions are completed correctly. In addition, first corrections and additions can be made directly (manually or prompted), e.g. to add a further definition of a technical term or to provide more technical background for an embodiment. With that, the second session is completed.

In the third session, the slow but rewarding manual review of the complete application must be performed as described beforehand. After a few hours, the draft should be ready for submission to the client for feedback or direct approval.

In this workflow, review of the completed application documents is facilitated by two aspects. First, the client had already been educated on the structure and the wording of the claims in the second session. Second, usage of client- and technology-specific prompts (or templates) assured that the structure and style of the generated text body is similar to previous applications known to the client.

Collaborative drafting with AI-assistance

Larger companies with inhouse patent attorneys can suffer a conflict of objectives. On the one hand it makes sense that all high-value patent applications are drafted by inhouse patent attorneys who have expert knowledge on the technology of the company, making sure that the best innovations of the company are best protected. On the other hand, inhouse patent attorneys are often very busy with the operational day-to-day work of managing large IP portfolios. This can result in long-delayed completion of patent application drafts or simply in that some high-value patent applications can’t be drafted inhouse, underutilizing the inhouse patent attorney’s expertise.

With the advent of AI Co-Pilots, inhouse patent attorneys could try a new approach. They can manually prepare a claim set, figure sketches and reference numerals in close alignment with the inventors. Then this scaffold application draft can be handed over to a patent attorney in private practice, who additionally reads the invention disclosure and the state of the art documents and uses an AI Co-Pilot to complete the application within a few work hours. With this workflow, the inhouse attorney brings his or her knowledge to the table but is less burdened with the time-consuming later steps of drafting. The application draft can be completed for the highest quality of disclosure in a time- and cost-efficient manner.

Where to start if you want to get into AI assisted drafting

Various tests of AI Co-Pilots for drafting patent applications are available on the Internete.g. “GenAI patent drafting tools under scrutiny” by Sébastien Ragot and Lionel Vial, published 02/25 on https://www.iam-media.com. and on LinkedIn. The top tools in user satisfaction appear to be quite stable in the rankings, indicating them to be a good starting point for patent attorneys who want to start drafting with AI support. Given the breadth of different concepts available in the market, it might make sense for a patent attorney to demo several tools before deciding which one to license and work with. Patent attorneys should select only software providers that have been independently audited and provide appropriate security measures as stated before.

Most software providers today charge flat fees in the range of 200€ - 800€ per month per patent attorney for unlimited use of an AI Co-Pilot for drafting, which is the equivalent of a few billed attorney hours per month. In other words, costs for the tools can be easily recovered if a patent attorney drafts several patent applications per month with such a tool and presumably saves hours of work on each one. However, if a patent attorney does not engage in drafting every month, costs could be prohibitive.

Adopting AI-based workflows is not just about licensing the right tool, but also about law firms and industrial IP departments acquiring new knowledge and a new skillset. Hence, the next two sections will cover upskilling of patent attorneys and training of candidate patent attorneys.

Upskilling of patent attorneys

While it is often propagated that any professional should become an expert in prompting and customizing LLMs, the continuous amount of work time (and associated opportunity cost) required to keep up to date with the technological developments might be prohibitive for individuals and organizations alike. It seems reasonable that not every patent attorney needs to become a prompt engineer and customize their own prompt libraries and templates. Most of the time, getting training through online courses on the underlying principles and limits of current LLMs plus training for a specific tool will suffice.

A best practice solution for teams of patent professionals would be to have a “multiplier”. This person becomes an expert user and helps others to setup their instructions, e.g. for different sections of a patent application, using advanced prompts and individual text examples from own hand-made applications of the respective colleagues.

Training of patent attorney candidates

Patent attorneys will need to sign off on any work product created, no matter if hand-crafted or AI-generated. In the author’s opinion, future patent attorneys will most likely only be able to be efficient quality gates for AI generated text passages, if they learned the skillset for drafting in the traditional manual way first.

Therefore, candidates should early on in their training learn to develop a mental model for what’s important in each case, including trial and error in iterative discussions with an experienced patent attorney. This way, one can learn the craft including how to present an argument well in writing and orally. Once the manual approach has been internalized, candidates are ready to enhance their skills with AI-support.

Conclusion and Outlook

AI-assisted drafting of patent applications is a use case of AI that’s ready for broad implementation by patent attorneys. With the advent of cloud-based Full Editors, systems that are able to also assist with figure creation, the user can build broad technical context for an LLM and automatically generate large parts of the description with a high initial quality. However, formulating claims at the level of an experienced patent attorney appears to be elusive for today’s AI Co-Pilots. Taken together, drafting a patent application in part changes to a review und improvement role. In the end, the patent attorney always remains responsible for the content of a patent application draft and has to assure accuracy, legal correctness and completeness.

With updated workflows that incorporate an AI Co-Pilot, patent attorneys can either save a significant amount of time when drafting or increase the depth of the technical disclosure in the description at a given time budget.

Currently, the advancement of reasoning capabilities in AI systems remains impressive, but no one – not even the developers of the AI systems themselves – knows how far the abilities of AI systems can be improved in the next 5-10 years. We live in a world now that not so long ago could have been the setting of a science fiction novel.

Expert predictions range from a steady and relatively slow incremental increase in AI capabilities, technological breakthrough scenarios where artificial general intelligence (AGI) disrupts pretty much the whole labour market, to runaway self-improvement scenarios where artificial superintelligence (ASI) emerges. Therefore, it seems impossible to make trustworthy predictions for what lies ahead for patent attorneys in terms of available AI assistance, but surely the current systems will only improve over time.

If patent attorneys keep adapting to the challenges, but also to the novel business opportunities that AI-based workflows provide, our profession will prevail and continue to provide valuable legal guidance for clients in protecting their best innovations.

Prompt, revise, repeat: the future is now!

Connect with the author on LinkedIn:

Comments