How to Enable an LLM Workflow Patent Application
An LLM workflow patent application is enabled when it teaches a competent engineer to build and run everything the claims cover without turning the project into a research program. What needs teaching is the workflow around the language model: what goes into the model, how that input is assembled, what the model is asked to do, how its output is checked, and what happens next. Applications that describe all of that in concrete terms tend to be enabled. One that says "a large language model generates a compliant summary" and stops there is usually an expensive, well-formatted wish.
What enablement means for an LLM patent
Patent lawyers have a habit of using ordinary words in unordinary ways, so some vocabulary first. "Enablement" means the written application, read by someone with ordinary skill in the field, lets that person make and use the claimed invention without undue experimentation. It is a cousin of "written description," which asks whether the application shows you actually had the invention when you filed. The two travel together, and an LLM application that fails one usually has trouble with the other. Neither can be fixed later, because nothing new can be added to an application after filing, so whatever an examiner, a court or a competitor's lawyer will someday need to see has to be there on day one.
How Amgen v. Sanofi applies to AI claims
The rule that matters most comes from the Supreme Court's 2023 decision in Amgen v. Sanofi, which involved antibodies but reads as if it were written with AI claims in mind. Amgen claimed every antibody that performed two functions, identified 26 of them, and offered methods for finding the rest. The Court held that wasn't enough, explaining that "the more a party claims, the broader the monopoly it demands, the more it must enable," and describing Amgen's discovery methods as "little more than two research assignments" (Supreme Court opinion). LLM claims are often written in the same functional style: a model "configured to" produce some result. If the application describes the result and leaves the reader to figure out how to get there by trial and error, it has the Amgen problem, whether the subject is proteins or prompts.
The Court also recognized that a specification may call for a reasonable amount of experimentation, and that what is reasonable depends on the nature of the invention and the underlying art. An invention at the convergence of AI and biotechnology needs far more enabling disclosure than a better mousetrap. Nobody expects you to disclose the weights of a foundation model or explain transformer architecture from first principles, though. A skilled engineer in this field already knows how to call a commercial model, set up a vector database and parse JSON, and the application can lean on that common knowledge. What it can't lean on is whatever makes your workflow better than the obvious version, since that is, almost by definition, not common knowledge. Those details have to be written down.
What an LLM workflow application should disclose
We look for six kinds of detail. Start with the structure of the workflow itself: the components, the order in which they run, and the data that passes between them, ideally with a diagram a reader could implement. Prompts come next. The application doesn't necessarily need the exact production prompt, which you may prefer to keep confidential, but it should give the prompt structure, the roles and instructions it contains, how examples are selected and inserted, and at least one representative prompt written out in full. Context assembly matters whenever the workflow retrieves documents or records before calling the model; explain how they are split, indexed, ranked and filtered, and how much of them reaches the model.
For model requirements, say what class of models works, which capabilities matter (context length, tool use, structured output), and whether any fine-tuning was done. If it was, describe the training data's characteristics: where it came from, how it was labeled and what made it suitable. Output handling covers how the workflow validates, scores, corrects or rejects what the model produces, and how it copes with the model's unfortunate tendency to give different answers to the same question on different days. Evidence comes last, meaning at least one worked example with real inputs, real outputs and a measured result compared against a baseline.
A worked example does double duty. It shows the reader that the workflow can actually be built, and it gives you something concrete to point to when someone argues that the claims cover more than you taught. At bottom, the Amgen Court's complaint was that the patent described a destination and handed the reader a map with no roads on it, and a worked example is a road. We ask inventors for that evidence. Where they don't have it yet, we draft prophetic examples that draw on the state of the art in the public domain and give a skilled engineer enough guidance to make and use the invention without undue experimentation.
How broad can LLM workflow claims be?
Breadth is where most of the fighting happens. Claims that cover any large language model need an explanation of why the workflow doesn't depend on a particular one, and results on more than one model family help. Claims that reach any document type, any industry or any downstream action call for an explanation of what makes the approach general, or enough varied examples that a reader can see the pattern. If the workflow only performs well with your in-house model, the claims and the disclosure need to be honest about that. A specification that quietly relies on a proprietary system the reader can't access invites the argument that only you could practice the invention, which is the opposite of what a patent is supposed to provide.
This area is genuinely unsettled. We're not aware of any court that has squarely decided how much an LLM application must disclose to be enabled, so the safest assumption is that the old rules apply with full force to the new technology.
Why detail also helps with patent eligibility
In the U.S., the bigger near-term threat to an LLM application is usually patent eligibility, the doctrine that keeps abstract ideas out of the patent system, and the same disclosure that enables the invention is often what saves it there too. In Recentive Analytics v. Fox, the Federal Circuit held that claims applying generic machine learning to a new data environment, without disclosing improvements to the models themselves, were ineligible, and the Supreme Court declined to review that decision in December 2025 (Supreme Court docket). On the other side of the ledger, the USPTO's own appeals panel found machine-learning training claims eligible because the specification explained how the invention improved the way the model itself operated, and the USPTO designated that decision precedential, which means examiners have to follow it (USPTO).
Examiners raise eligibility against AI workflow applications regularly. We describe and claim these inventions as systems, a combination of hardware and software, which typically keeps them from being treated as abstract during examination. We also explain in detail why the invention can't be performed in the human mind, which gives the examiner one more reason not to call it abstract. Neither step works on its own if the claims describe generic computing, which is why the technical-improvement explanation matters as much as the claim format.
The application should also say what technical problem the workflow solves (latency, accuracy, hallucination rate, token cost, memory use) and how the specific design solves it. "It does the job faster than a paralegal" is a business benefit, while "it cuts model calls by 60 percent by routing only ambiguous records to the larger model" is a technical one.
Enablement for AI inventions in Europe
Europe takes a similar view with a different vocabulary. The European Patent Office calls enablement "sufficiency of disclosure," and its examination guidelines say that when an invention's technical effect depends on particular characteristics of the training data, those characteristics must be disclosed unless a skilled person could work them out from common general knowledge, though the dataset itself generally need not be (EPO Guidelines). The EPO's appeal boards have refused a neural-network application that said only that the training data should cover a wide range of patients, because that told the reader nothing about what data would actually work (EPO Case Law of the Boards of Appeal). More recent board decisions have stressed that there are no special rules for machine learning and that sufficiency is decided case by case, which is reassuring until you realize it means nobody can promise you a safe harbor. The EPO also asks whether the claimed features contribute to a technical solution of a technical problem, so the advice about stating the technical problem applies there too.
Disclosure rules for AI patents in China
China has become the most explicit of the major offices. Revised examination guidelines from the China National Intellectual Property Administration took effect on January 1, 2026 (CNIPA, in Chinese). For inventions that build or train a model, they say the specification generally needs to describe the model's necessary modules, layers or connections and the specific training steps and parameters. Most LLM workflows fall in the other category, inventions that apply a model in a particular field, and there the specification generally needs to explain how the model is combined with that field and how its inputs and outputs are set up to show their inherent connection. The same revision tells examiners that simply moving an existing model to a new scenario, without technical changes, generally won't be inventive. An application written to the Chinese standard will often cover much of what the U.S. and European offices look for as well.
Patents, trade secrets and the details you'd rather keep
All of this creates a real tension with trade secrets. The prompt structure, retrieval tuning and validation logic that make a workflow enabled are exactly the details many companies would rather not publish. Keeping them back is a legitimate business choice, but it has to be made deliberately and before filing, because an application that holds back the essential details may produce an invalid patent while still publishing everything else it describes. Sometimes the workflow can be split, with what competitors could reverse engineer from the product going into the patent and what stays behind the server going into the trade secret program. Which parts go where depends on the product, the competition and how the company plans to make money, and it's worth deciding with someone who has seen both kinds of protection fail.
Inventor evidence to gather before filing
Engineers who build LLM workflows iterate constantly, and the version that works best is often three versions after the one described in the invention disclosure. Before filing, ask the team what they changed since the disclosure was written and why. The answers are frequently the most valuable paragraphs in the application.
Two other things often go missing. One is the prompts themselves. An invention has two parts, conception and reduction to practice, and because an AI can't be an inventor, the human inventors' prompts can be important evidence that they conceived the invention. Using AI tools doesn't change who the inventors are. The USPTO's current guidance treats AI as a tool, like any other software, and applies the same inventorship standard as for any invention: a single inventor must have conceived it, and each of several joint inventors must have contributed in a significant manner (USPTO guidance).
The other is proof of concept, such as a working prototype. When the inventors don't have one yet, we draft prophetic, best-educated-guess examples into the provisional application so those gaps can be filled at the non-provisional filing. A provisional application needs enabling examples too. Later claims get the provisional's filing date only if the provisional itself describes and enables them, so the Amgen standard matters from the first filing.
If you're preparing an application on an LLM workflow and want a second set of eyes on whether it teaches enough, contact us.
This FAQ provides general information, not legal advice, and does not create an attorney-client relationship. Outcomes depend on specific facts and jurisdictions. See our Legal Notices.
