The Hidden Work Behind Patent Prosecution
When people think about patent attorneys, they often imagine the process of drafting brand-new patent applications: taking an invention, translating it into precise legal and technical language, and filing it with the patent office.
But in practice, a large part of patent prosecution happens after the initial application is filed.
One of the most demanding and recurring parts of a patent attorney’s work is responding to claim rejections issued by patent examiners. These rejections usually arrive in the form of Office Actions, where the examiner explains why certain claims are not currently allowable. The attorney must then review the examiner’s reasoning, analyze the cited prior art, compare it against the claim language, and decide how to respond.
This process is highly technical, highly detailed, and often repetitive.
For each rejected claim, the attorney may need to answer questions such as: Does the cited prior art actually disclose the claimed feature? Is the examiner interpreting the claim language correctly? Is there support in the original specification for a possible amendment? Should the response argue against the rejection, amend the claim, cancel the claim, or combine several strategies?
The challenge is not simply writing a response. The challenge is gathering the right information from multiple long documents, understanding the relationship between the claims and the prior art, and turning that analysis into a clear, defensible legal argument.
That was the core problem behind this project: how can we help patent attorneys respond to claim rejections faster, while keeping legal judgment and final decision-making firmly in human hands?
The Business Problem
Responding to an Office Action is a document-heavy and decision-heavy workflow.
A single response may require the attorney to move between the original patent application, the current claim set, the Office Action, cited prior art references, patent rules, MPEP guidance, and relevant case law. Before drafting even begins, the attorney must locate the relevant passages, understand the examiner’s argument, identify which claims are affected, and determine what response strategy is most appropriate.
This creates several bottlenecks.
First, the relevant information is scattered across multiple long and technical documents. The attorney has to manually connect claim language, examiner reasoning, prior art passages, and legal support.
Second, the work has to be done claim by claim. A general summary of the Office Action is not enough. Each independent claim rejection may require a different argument, amendment, or legal strategy.
Third, the work is repetitive but still high-stakes. Many steps follow a predictable pattern, but the final legal judgment requires professional expertise. This makes the process difficult to automate fully, but very suitable for AI assistance.
The goal of this project was to reduce the manual workload involved in responding to Office Actions. Based on the workflow comparison, the system is estimated to save roughly 80–90% of the time typically spent preparing a response to a claim rejection.
Project Overview
This project was built as an end-to-end AI-assisted workflow for helping patent attorneys respond to claim rejections.
The goal was not to build a generic legal chatbot. The goal was to design a structured system that follows the actual working process of a patent attorney: uploading documents, extracting the relevant information, retrieving supporting material, generating possible response strategies, and letting the attorney review, select, and edit the final response.
Before the main workflow was built, I created dedicated vector databases to support the retrieval layer. One database was built from patent rules, laws, and the MPEP. This allows the system to retrieve relevant legal and procedural guidance during analysis. A second database was created from important patent law cases, giving the system access to precedent-based reasoning when constructing or supporting a response.
The workflow begins in the user interface, where the attorney uploads the relevant documents: the patent application as filed, the Office Action, an optional document with the most recent claims, and optional prior art documents.
Once the attorney starts the process, the system extracts the key information from the documents, fetches referenced patents where needed, analyzes each independent claim rejection, retrieves relevant context from the knowledge bases, generates possible responses, and finally compiles the selected responses into a formal draft.
From Document Upload to Structured Extraction
The first step in the workflow is document ingestion.
The attorney uploads the required prosecution documents through the UI. At minimum, the system needs the patent application as filed and the Office Action. If the case is not at the first Office Action stage, the attorney can also upload the most recent claims. Prior art documents can also be uploaded when available.
After the upload, the system validates the inputs and begins extracting the information required for the rest of the process. This includes the claims, claim rejections, core invention details, referenced patents, and relevant technical context.
This extraction stage is important because the quality of the downstream reasoning depends on the quality of the structured data. The system needs to understand what was rejected, which claims were affected, what references were cited, and what technical features are central to the invention.
Instead of asking the AI to respond to an entire Office Action in one broad prompt, the workflow breaks the problem into smaller, more controlled steps.
Retrieval-Augmented Legal and Technical Analysis
Once the relevant information has been extracted, the system moves into retrieval and analysis.
If the Office Action references prior patents, the system checks whether those patents already exist in the internal patent database. If not, it attempts to fetch them using a search API and makes them available for analysis. This ensures that the system can compare the rejected claims against the same type of material the examiner relied on.
For each independent claim rejection, the system generates a targeted search query and retrieves the most relevant chunks from the available knowledge bases. These include the patent rules and MPEP database, the case law database, and the patent or prior art database.
This retrieval-augmented generation approach is central to the system. Legal and patent work should not rely only on a model’s general knowledge. The response needs to be grounded in relevant legal rules, procedural guidance, precedent, claim language, and prior art.
By retrieving focused context for each rejection, the system gives the AI a more reliable foundation for generating useful response options.
Generating Response Strategies
After the relevant context is retrieved, the AI generates several possible response strategies for each independent claim rejection.
These responses may include disputing the examiner’s interpretation, proposing an amendment to the claim, or using another strategy depending on the nature of the rejection. The goal is not to produce a single answer and present it as final. The goal is to provide the attorney with several well-reasoned options.
Each response option includes an estimated probability of acceptance, together with the reasoning behind that estimate. This helps the attorney compare the available strategies and decide which response is most appropriate.
This is where the system becomes more than a drafting assistant. It does not simply generate legal-sounding text. It organizes the rejection, retrieves supporting material, compares possible response paths, and presents the attorney with structured choices.
Human-in-the-Loop Review
A core design principle of the project was that the attorney must remain in control.
For each independent claim rejection, the attorney reviews the response options inside the UI and selects the preferred one using a radio widget. The AI supports the analysis and drafting process, but it does not make the final legal decision.
This human-in-the-loop structure is essential for legal workflows. The system can reduce manual effort, speed up review, and generate strong first drafts, but the attorney still provides the professional judgment, strategic choice, and final approval.
After the attorney has selected the preferred response for each rejection, the system compiles those selections into a formal response document. The draft is then displayed in the UI, where the attorney can continue editing manually or use AI assistance to refine the language and structure.
Results and Takeaways
The final result is a guided workflow that takes the attorney from document upload to a structured draft response.
Instead of starting from a blank page, the attorney receives a draft built from extracted rejection data, retrieved legal context, prior art analysis, and selected response strategies. This can reduce the time spent preparing a response to an Office Action by an estimated 80–90%.
The main lesson from this project is that legal AI should not be designed as a generic chatbot. The real value comes from building systems around the actual professional workflow.
For patent claim rejections, that means breaking the process into clear stages: document ingestion, structured extraction, patent retrieval, claim-by-claim analysis, legal and technical retrieval, response generation, attorney review, and draft production.
AI is most useful here not because it replaces the attorney, but because it removes a large portion of the repetitive preparation work. That gives the attorney more time to focus on what matters most: legal strategy, technical reasoning, and final judgment.