Choosing An AI Patent Attorney: Everything You Need To Know
As an AI Patent Attorney, it’s clear that artificial intelligence (AI) is permeating almost every sector, from ChatGPT to smart indoor farming to speech recognition…
[SERVICE]AI patents
AI patent attorneys for models, pipelines and AI products
We patent the technical improvement inside your AI, in how the model is built, trained, fed or run, and claim it so the patent survives the eligibility test and reads on a competitor's system. Registered patent attorneys with software and AI backgrounds, and one flat fee for each step, quoted before any work begins.
Your attorneys
Andrew RapackeManaging Partner · Registered Patent AttorneyElectrical and computer engineering. Former Naval Engineering Officer.
Chris WattRegistered Patent Attorney · Former USPTO ExaminerSoftware, AI and machine learning, and MedTech. A full-stack software engineer.
Ken BrooksRegistered Patent AttorneySoftware, AI and machine learning, electrical and mechanical systems.
[01]Who it's for
For founders whose product runs on a model.
A competitor can call the same foundation model you do. What sets you apart is how you train it, feed it, run it or put it to work, and that is what a patent can protect.
AI-native startups
Teams whose product is the model, the agent or the pipeline around it, filing before the first launch, paper or investor demo.
Software companies adding AI
AI features built into a product you already sell, where the way the model is wired into the system is the invention.
Hardware and edge AI
Accelerators, on-device inference and ways to run models with less memory, power or bandwidth.
Applied AI in regulated fields
Models that read scans, signals, transactions or sensor data in health care, finance and vehicles.
[02]What's patentable
Patent the technical improvement, not the math.
An AI patent covers a technical method or system. A formula, a model described only as math, or a business result reached with AI is an abstract idea on its own.
The method also has to be new and non-obvious. The $999 search tells you where you stand before you spend on filing.
Can be patented
- Model architecturesNew layer arrangements, attention mechanisms or structures that make a model more accurate or efficient.
- Training methodsNew ways to train, fine-tune or update a model, such as learning a new task without losing an old one.
- Data pipelinesPreparing, labeling, cleaning or generating training data in a new technical way.
- Applied AI in a fieldA model put to work on a specific technical task, with its output used to control or change a system.
- Hardware accelerationCircuits, memory layouts and scheduling that run models faster or on less power.
- Inference and deploymentServing, compressing or distilling models so they run at lower latency or cost.
Not on its own
- The math aloneAn equation, a loss function or an algorithm claimed in the abstract is a mathematical concept.
- A known model on new dataApplying generic machine learning to a new kind of data, without improving the model, is not enough.
- AI added to a business processAutomating a known business practice with a model adds nothing patentable by itself.
- Training data by itselfA data set is information, not an invention. Secrecy and contracts protect it.
- The model weightsLearned parameter values are data too. Keep them as a trade secret.
- Steps a person could do mentallyObservations, evaluations and judgments a person could make are treated as abstract, even when a computer makes them.
[03]The work
What does an AI patent attorney do?
An AI patent attorney finds what is technically new in your AI system, decides what to patent and what to keep secret, writes claims that can pass the eligibility test, and answers the USPTO until the patent is granted.
- Find the invention. A discovery call with your engineers on the architecture, the training pipeline and the data, to separate what is new from what is a standard building block.
- Search before you spend. A patentability search of US and foreign patents and published applications, with a written report on what is likely to be allowed.
- Draft for eligibility and for enforcement. A specification that explains the technical improvement in enough detail, and claims written so a competitor's product infringes them.
- Answer the examiner. Responses to rejections for eligibility, prior art and lack of detail, and interviews with the examiner when a call will move faster than a letter.
- Build the portfolio. New filings as the model and the product change, foreign filings where you sell, and enforcement if someone copies what you patented.
Apart from inventors filing for themselves, only attorneys and agents registered with the USPTO can prosecute patent applications for others. Every application here is drafted by a registered patent attorney.
Comparing firms? Read choosing an AI patent attorney and what to look for in an artificial intelligence law firm.
[04]Where it lives
Where AI inventions usually are.
Look past the prediction. The part of an AI product worth patenting is usually what your team had to change to make the model work in the real world.
- The model itself. A new architecture, a new way to combine models, or a change that cuts memory or compute for the same accuracy.
- Training. Curriculum, continual learning, fine-tuning with less data, synthetic data generation, or a way to train across sites without moving the data.
- Data preparation. Labeling, cleaning, feature extraction or a new way to turn raw signals into training examples.
- Inference and serving. Caching, batching, quantization, distillation or routing requests between models to cut latency and cost.
- Agents and wrappers. How an agent plans, calls tools, checks its own output or keeps memory, and how a product built on someone else's model grounds, filters or verifies what it returns.
- Applied AI. Computer vision, speech, anomaly detection or control systems where the model's output changes what a machine does.
- Hardware. Chips, accelerators and on-device inference that run models faster or on less power.
A useful test is to ask what your engineers argued about. The problem that took three attempts to solve is often the invention.
More on whether an AI wrapper is patentable, whether AI agents are patentable and AI patent examples from well-known companies.
[05]Eligibility
Will your AI invention pass the Alice test?
It can, when the claims show a specific improvement to how the model or the computer works. Every AI claim is screened under 35 U.S.C. 101 with the two-step test from the Supreme Court's Alice and Mayo decisions, and the USPTO has applied it to AI in detail since 2024.
What the USPTO looks for
The USPTO's current eligibility guidance is in sections 2103 to 2106.07 of its examining manual, the MPEP, with a July 2024 update on artificial intelligence. One of its AI examples shows the line. A claim that trained a neural network and used it to detect network anomalies was not eligible. A claim that also dropped the malicious packets and blocked the traffic in real time was eligible, because it improved network security. A claim to a chip built to run the network was eligible too.
In August 2025, the USPTO reminded examiners in the software and AI art units to read claims as a whole, not to treat steps a person could not practically perform in their head as mental steps, and to reject only when a claim is more likely than not ineligible.
What changed in 2025
In September 2025, an Appeals Review Panel led by the USPTO Director held in Ex parte Desjardins that a method of training a machine learning model to learn a new task while protecting what it learned earlier improved how the model itself works. The decision was made precedential on November 4, 2025, and added to the MPEP in December 2025. The specification has to describe the improvement in enough detail that an engineer would recognize it. A bare statement that the invention is better is not enough.
Since December 2025, the USPTO has also pointed applicants to subject matter eligibility declarations, sworn evidence from an engineer or scientist about the technical improvement, filed under its existing Rule 132 practice.
What the courts held
Courts are not bound by USPTO decisions. In April 2025, the Federal Circuit held in Recentive Analytics v. Fox that patents claiming generic machine learning applied to a new data environment, without an improvement to the model, were not eligible. A patent that is granted still has to hold up in court, so we draft for both.
The two-step test and the cases behind it are on our software patent attorney page. For more on AI specifically, read patent eligibility for AI inventions and can you patent an algorithm.
[06]How we draft
Claims written around the technical improvement.
We start from your system, not from a description of what the model predicts. What goes in, what happens to it at each step, what comes out, and what is different from the approaches you tried first.
The improvement in the claim, the parts that deliver it
We write the step that produces the technical result into the claim, so the claim as a whole reflects the improvement and not only the idea. Then we pair claims written around what the system does with claims that name the parts that do it, so the patent reaches a competitor who builds the same result a different way.
Training claims and inference claims
A training claim is performed once, by whoever trains the model. An inference claim is performed every time someone runs it. We usually claim both, plus the system and the stored instructions, so the patent reaches a company that licenses a trained model as well as one that trains its own.
Enough detail to be enabled
A patent must describe the invention well enough for a skilled engineer to make and use it. For machine learning that means the architecture, the training procedure, the kind of data and how it is prepared, and the result it produces, with ranges and alternatives. Broad claims with thin detail invite rejections for lack of written description and enablement, and claims that only combine a known model with a known kind of data invite obviousness rejections.
What to bring to your first call
- An architecture or data-flow diagram, even a whiteboard photo
- How the model is trained and on what kind of data, at a high level
- Benchmarks or test results that show the improvement
- What is already public, including papers, preprints, repositories, demos and launch dates
- Models, products and open source tools that do something similar
No finished code or trained model is needed. Your attorney turns the call into the description the application needs, and no confidential data has to be shared.
Beyond the filing
The same team handles what comes after. Patentability and freedom-to-operate opinions, office action responses, post-grant proceedings at the USPTO, patent litigation in federal court, and licensing agreements when a partner wants to use your model or method. For AI features inside a cloud platform, see SaaS patents, and for the field as a whole, machine learning patent attorneys for tech startups.
[07]Inventorship
Can you patent an invention made with AI?
Yes, when a person conceived what is claimed. An AI system cannot be named as an inventor, and the USPTO treats it as a tool the inventor uses, like lab equipment or software.
In Thaler v. Vidal (2022), the Federal Circuit held that the Patent Act requires inventors to be natural persons, so an application naming an AI system as the only inventor could not proceed. The Supreme Court declined to review the case.
In November 2025, the USPTO replaced its earlier guidance on AI-assisted inventions. The same inventorship standard now applies to every invention. Conception, the formation in a person's mind of the complete invention, is the test. The joint inventorship factors apply only when more than one person is involved, because an AI system cannot be a joint inventor. The same rule applies to design patents.
Keep a record of who conceived what
- Who framed the problem and defined what the system had to do
- Which outputs from AI tools were used, and what a person chose, changed or tested
- Dated notes, commits and test results that show the idea taking shape
- Signed invention assignments from every founder, employee and contractor who contributed
More on whether AI can be an inventor.
[08]Patent or secret
Patent the model, or keep it secret?
Patent what a competitor can see or work out from your product. Keep secret what never leaves your servers, like weights, training data and tuning. Most AI companies need both.
| Patent | Trade secret | |
|---|---|---|
| Use it for | Architectures, methods and features a competitor could learn from your product, API or papers | Model weights, training data, prompts and tuning that never leave your servers |
| How you get it | Apply to the USPTO and pass examination | Take reasonable measures to keep it secret, such as access limits, NDAs and employee agreements |
| How long it lasts | Generally 20 years from filing the full application | As long as it stays secret |
| Stops someone who builds it on their own | Yes | No. Independent development and reverse engineering are lawful |
| Made public | Yes. Applications are usually published 18 months after the earliest filing date | Never. Disclosure ends the protection |
| Proving a copy | Easier when the claimed step shows in a product, an API or an output | Needs proof the secret was taken or misused |
A patent is only as useful as your ability to see infringement. A training method that runs inside a competitor's data center can be hard to prove, so we favor claims on what shows in the product, the API or the output, and leave the rest to secrecy.
For copyright in the same comparison, see patent, copyright or trade secret. For more on the secret side, read aspects of AI that fit trade secret protection.
[09]The flat-fee path
Three flat fees to get on file.
Every step is quoted before any work begins, and the search fee is credited when you file. Most AI founders file a provisional before a launch, a demo or a paper, then the full application within 12 months.
01Start here
Patentability search
$999
Patentable or your money back. Credited when you file
- An invention discovery call with your attorney
- A search of US and foreign patents and published applications
- A written patentability report and a review call
Report in about 5 days
Not patentable, full refund*02Then
Provisional patent
$2.5k–3.5k
Software flat fee
- A strategy call on your architecture, training and data
- Drafted around the technical improvement by a registered patent attorney
- You review and approve it before filing
Patent pending the day it is filed
Patentability issues, full refund*03Within 12 months
Non-provisional patent
From$11,000
Drawings and USPTO filing fees included
- Training, inference, system and medium claims written for examination
- Formal drawings and USPTO filing fees
- Improvements made since the provisional, folded in
Filed before your 12-month deadline
Flat fee, not covered by the guarantee
*Refunds subject to the RLG Guarantee Policy in your engagement letter. Utility (non-provisional) patents are quoted flat and are not covered by the guarantee. How The RLG Guarantee works
[10]The process
From first call to granted patent.
What happens at each step and who does the work. Timing shown is typical, never a promised date.
AI patentSearch to grant
- Day one01Day oneDiscovery callWalk your attorney through the system, the training and the data.
- $99902$999Patentability searchUS and foreign patents and published applications.
- About 5 days03About 5 daysReport and reviewA written report, then a call on what to claim and what to keep secret.
- $2.5k–3.5k04$2.5k–3.5kProvisional filedDrafted around the technical improvement, approved by you, filed.
- Months 1 to 1205Months 1 to 12Model improvements addedNew training methods and features go into another provisional or the full application.
- Month 1206Month 12Full application filedTraining, inference, system and medium claims, before the deadline.
- 07Examination and grantWe answer the examiner. Grant often comes 2 to 3 years after the full filing.
[11]Startup IP strategy
An IP plan for an AI startup.
File a provisional before anything goes public, build on it for 12 months, keep the rest secret and register the name. Most AI startups we work with follow some version of this.
- Find what is new. Start with the discovery call and the search, so you file on the part of the system that is both new and worth protecting.
- File a provisional before it goes public. A paper or preprint, a public code repository, a conference talk, a demo day or a launch can make the invention public. The US gives a one-year grace period after your own disclosure. Many countries give none.
- File again as the model changes. A provisional covers only what it describes. The full application, due within 12 months, can claim the dates of more than one provisional, so new training methods and features can be added along the way.
- Decide on other countries by month 12. A PCT application keeps them open. Eligibility rules for AI differ abroad, so read AI patents by country before you choose.
- Protect the name. Clear and register the product and model name before it goes on your site, your app listing and your API docs. A trademark is a $2,000 flat fee with us.
- Own what you build on. Written assignments from everyone who contributed, and a record of where your training data and open source models came from and the licenses on each.
More on AI training data and IP and patents on open source AI. Not sure what to file first? The call is free.
[12]Questions
AI patent questions.
Still deciding? Your first call is free and comes with a straight answer on whether to file.
Book A Free Strategy CallWhat does an AI patent attorney do?
Finds what is technically new in your AI system, runs a patentability search, drafts claims around the technical improvement so they can pass the eligibility test, answers USPTO rejections and builds a portfolio as the product grows. A registered patent attorney with a software background also tells you what to keep as a trade secret instead.
Can you patent an AI model?
You can patent a new architecture, training method or way of running a model when the claims show a technical improvement, such as less memory, faster training or higher accuracy on a defined task. A model described only as math, or a known model applied to a new kind of data, usually is not eligible. The learned weights are data and are better kept secret.
Can you patent an algorithm?
Not as an abstract formula. You can patent the specific steps a system performs with it when they improve how a computer or another technology works. Claims that tie the algorithm to a technical result, and say how the result is reached, are what the USPTO looks for.
Can AI be listed as an inventor?
No. US law requires inventors to be natural persons, as the Federal Circuit held in Thaler v. Vidal in 2022. The USPTO's November 2025 guidance treats AI systems as tools, so the person who conceived the claimed invention is the inventor.
Can I patent an invention I made with AI tools?
Yes, if a person conceived what is claimed. Using a generative model, a code assistant or a design tool does not stop you from patenting the result, but the inventor has to be a person who formed the complete idea. Keep records of who framed the problem and what people chose, changed and tested.
Should I patent my model or keep it a trade secret?
Patent what a competitor could learn from your product, your API or a paper. Keep weights, training data and tuning secret if they never leave your servers. A secret lasts as long as it stays secret, but it does not stop someone who builds the same thing on their own. Many AI companies use both.
How much does an AI patent cost?
With us, three flat fees to get on file. A $999 patentability search, a $2,500 to $3,500 provisional for software and AI, and a full application at $11,000 with drawings and USPTO filing fees included. Office action responses are quoted flat before any work begins. USPTO issue and maintenance fees come later, and our software patent page lists them.
How long does it take to get an AI patent?
A provisional makes you patent pending the day it is filed. The full application is due within 12 months, and grant often comes 2 to 3 years after it is filed, depending on the art unit and how many rounds with the examiner it takes. That is typical timing, not a promised date.
Should I file before I publish a paper or open-source the model?
Yes. A preprint, a conference paper, a public repository or a live demo can make the invention public. The US gives a one-year grace period after your own disclosure, but many countries give none. File at least a provisional first.
Can I patent an AI wrapper built on someone else's model?
Not the underlying model, and not the idea of using it for a task. You may be able to patent what your product adds, such as how it grounds, filters, verifies or routes the model's output, when that solves a technical problem in a new way.
Does a US patent protect my AI abroad?
No. Patents are territorial. You generally have 12 months from your first filing to file in other countries or file a PCT application, and the eligibility rules for AI and software differ, especially in Europe, where a further technical effect is required.
What does The RLG Guarantee cover for AI patents?
The patentability search and the provisional application. If the search finds your invention is not novel, you get a full refund or another search. If patentability issues come up on your provisional, you get a full refund or more searches. Non-provisional applications are not covered. Refunds are subject to the RLG Guarantee Policy in your engagement letter.
[13]Learning Center
Go deeper on AI patents.
Artificial Intelligence Law Firm: What to Look for When Patenting Your AI
Choosing an artificial intelligence law firm decides whether your AI patent survives Alice/Mayo. Learn what to vet, then book a free IP strategy call.
Patent Eligibility for AI Inventions: What Inventors Need to Know
Learn the rules for patent eligibility for AI inventions, the USPTO’s 2024 guidance, and how to draft claims that survive § 101. Book a free IP strategy call.
Can You Patent an Algorithm and What Does It Actually Take to Qualify
Algorithms are generally understood and broadly defined by Merriam-Webster’s dictionary as a “step-by-step procedures for solving problems or accomplishing some ends.” In…
Can AI Be an Inventor? Intellectual Property in the AI Era
Can AI be an inventor? As everyone around the world gets excited about the new age of artificial intelligence (AI), innovation around AI has exploded. According to IBM…
Strategies for Claiming AI Inventions in Patent Applications
Inventions based in artificial intelligence (AI) are very interesting because these inventions often incorporate a variety of inventive technologies. There is usually a…
[14]Client reviews
Founders who filed with us.
4.7 across 614 public reviews on Google, Trustpilot and Avvo.
#1Top reviewed IP law firm*
Read client reviews“Andrew Rapacke and his team are exceptional. They did our provisional and non provisional patents. The entire team is professional and an ease to work with.”
Von DorjevskiFounder, Inguardia Health · Client review “I can safely trust them with scientific and legal integrity regarding ownership and accountability of inventions.”
Hussain AtherPatent · Client review “Their writing allowed for protection of my invention's most important aspects, as well as giving me room to modify and adapt my product.”
Andrew WebberProvisional patent · Client review
Reviews reflect individual client experiences. Quotes are excerpts. Prior results do not guarantee a similar outcome. *#1 top reviewed means the most public client reviews on Google, Trustpilot and Avvo of any US patent or trademark firm.
Protect the model before you publish it.
Walk through your AI system with a registered patent attorney. The call is free, with no obligation.