An AI agency in Atlanta built a retrieval-augmented generation system for a legal services client in 2025. The system worked well, the client was happy, and the agency moved on to other projects. Six months later, the agency received a letter from a patent holding company asserting that the RAG architecture used in the system infringed on three of their patents. The patents covered specific methods for combining retrieved documents with language model prompts, chunking strategies for document retrieval, and hybrid search techniques. The agency had never considered that their implementation might infringe on existing patents. The licensing demand was $250,000. The alternative was litigation, which would cost more regardless of the outcome. The agency settled for $175,000—more than their entire profit from the engagement.
The AI patent landscape is one of the fastest-growing and most contentious areas of intellectual property law. Major technology companies, universities, patent holding companies, and individual inventors have filed thousands of patents covering AI techniques, architectures, training methods, and applications. Any agency building AI systems is potentially implementing patented methods without knowing it.
This post covers the AI patent landscape your agency needs to understand, strategies for managing patent risk, and how to think about patents as both a defensive and offensive asset for your business.
The Current AI Patent Landscape
Volume and Velocity
AI-related patent filings have exploded over the past five years. The USPTO, EPO, and other patent offices worldwide are processing tens of thousands of AI-related applications annually. As of 2026, hundreds of thousands of active AI patents exist globally, covering everything from fundamental machine learning techniques to specific application methods.
Who is filing:
- Big tech companies (Google, Microsoft, IBM, Amazon, Meta, Apple) hold the largest portfolios, with thousands of AI patents each
- Chinese companies (Baidu, Tencent, Alibaba, Huawei) have been filing aggressively, particularly in computer vision, NLP, and autonomous systems
- Universities and research institutions hold significant patents in fundamental AI techniques
- Patent holding companies (NPEs) have been acquiring AI patents and asserting them against operating companies
- Startups and smaller companies hold patents in niche areas and specific applications
What Is Being Patented
Foundational techniques: Patents exist on various aspects of neural network architectures, training algorithms, optimization methods, and inference techniques. While some fundamental techniques are in the public domain, many specific implementations and improvements are patented.
Application methods: How AI is applied in specific domains—medical diagnosis, financial analysis, content recommendation, autonomous systems—is extensively patented. These application patents are often the most relevant to AI agencies because they cover specific use cases.
Data processing: Methods for preparing, augmenting, and managing training data are patented. This includes data labeling techniques, synthetic data generation, data augmentation methods, and dataset curation approaches.
Architecture innovations: Specific neural network architectures, attention mechanisms, memory systems, and model composition techniques are patented. The transformer architecture itself is not patented (it was published as open research), but many variations and improvements are.
RAG and retrieval methods: Given the popularity of retrieval-augmented generation, there is a growing body of patents covering retrieval methods, chunking strategies, re-ranking approaches, and hybrid retrieval techniques.
Fine-tuning and adaptation: Methods for fine-tuning, few-shot learning, prompt engineering, and model adaptation are increasingly patented.
Evaluation and monitoring: Methods for evaluating model performance, detecting bias, monitoring for drift, and ensuring quality are patented.
Patent Quality Issues
Many AI patents are of questionable quality—they cover techniques that may be obvious to practitioners, or they have prior art that was not identified during prosecution. However, even low-quality patents create risk because defending against patent assertions is expensive regardless of the merits of the patent.
The practical reality: A patent does not need to be valid to be expensive. Defending against a patent infringement claim costs $2-5 million on average through trial. Even early-stage defense (pre-discovery) typically costs $200,000-500,000. This cost asymmetry means that patent holders can extract settlements well below litigation costs even when their patents are weak.
Patent Risk for AI Agencies
Types of Patent Risk
Direct infringement: Your agency builds an AI system that implements a patented method. The patent holder can assert the patent against you.
Contributory infringement: Your agency provides AI components or services that enable your clients to infringe patents. The patent holder can assert contributory infringement against you.
Induced infringement: Your agency designs an AI system and instructs your client to use it in a way that infringes a patent. The patent holder can assert induced infringement.
Client indemnification risk: Your client agreements may include intellectual property indemnification clauses that require you to defend and indemnify your clients if the AI systems you build for them infringe third-party patents.
High-Risk Technical Areas
Some technical areas carry more patent risk than others because they are heavily patented and commonly implemented by agencies.
Recommendation systems: Collaborative filtering, content-based filtering, and hybrid recommendation methods are extensively patented. If you build recommendation AI, patent risk is significant.
Natural language processing: Specific NLP techniques, sentiment analysis methods, entity extraction approaches, and text classification methods are patented.
Computer vision: Object detection, image classification, facial recognition, and visual search methods are heavily patented.
Retrieval-augmented generation: As RAG has become the dominant architecture for enterprise AI, patents covering RAG techniques are being actively asserted.
Conversational AI: Dialog management, intent recognition, context handling, and response generation methods are patented.
Predictive analytics: Specific methods for time series forecasting, anomaly detection, and predictive modeling are patented across many industries.
Managing Patent Risk
Freedom to Operate Analysis
A freedom to operate (FTO) analysis evaluates whether a specific product or technology infringes existing patents.
When to conduct FTO analysis:
- Before launching a new AI product or service
- Before entering a new technical area or application domain
- When a client engagement involves novel technical approaches
- When you receive a patent assertion or licensing demand
How FTO analysis works:
- Identify the key technical features of your AI system
- Search patent databases for patents that cover those features
- Analyze the claims of potentially relevant patents against your implementation
- Assess the risk of infringement for each identified patent
- Recommend design changes, licensing, or risk acceptance based on the analysis
FTO analysis is expensive ($15,000-50,000 for a thorough analysis) and time-consuming. Most agencies cannot afford FTO analysis for every engagement. Prioritize FTO analysis for products you plan to market broadly, high-value client engagements, and technical areas you know to be heavily patented.
Design-Around Strategies
When FTO analysis identifies potential infringement, design-around strategies modify your implementation to avoid the patented claims.
Approaches:
- Study the patent claims carefully to identify the specific elements that define the protected method
- Modify your implementation to avoid one or more claim elements
- Document the design-around and the reasoning behind it
- Have patent counsel review the design-around to confirm it is effective
Limitations: Design-arounds are not always possible. Some patents are broadly written, and avoiding them may require fundamentally different approaches. Also, design-arounds that are too close to the patent claims may still be found infringing under the doctrine of equivalents.
Defensive Publications
If your agency develops novel AI techniques, consider defensive publications—publishing your innovations to create prior art that prevents others from patenting similar techniques.
How defensive publications work:
- Document your innovation in sufficient technical detail that it would qualify as prior art
- Publish the documentation in a format that patent examiners will find (technical journals, the Defensive Patent License database, or IP.com)
- The publication creates prior art that prevents anyone (including you) from patenting the disclosed technique
When to use defensive publications:
- When you develop a novel technique that you want to be free to use but do not want to invest in patenting
- When you want to prevent competitors or patent trolls from patenting techniques you have developed
- As a complement to your patent portfolio—patent your most valuable innovations, defensively publish the rest
Patent Licensing
Sometimes the most practical approach to patent risk is licensing.
Inbound licensing: If your AI system implements a patented technique that you cannot design around, licensing the patent may be more cost-effective than redesigning or litigating.
- Negotiate license terms carefully—scope, territory, duration, and royalty structure all matter
- Seek non-exclusive licenses that do not restrict your ability to work with competitors of the patent holder
- Include sublicensing rights if your clients need to operate the AI system independently
Cross-licensing: If your agency holds patents, cross-licensing arrangements with other patent holders can provide mutual freedom to operate.
Patent pools: In some technical areas, patent pools aggregate patents from multiple holders and offer licenses on standardized terms. Check whether relevant patent pools exist in your technical areas.
Building a Patent Portfolio
For larger agencies, building a patent portfolio provides defensive value and potential revenue.
Defensive value: If you hold patents, potential patent asserters face the risk that you will counter-assert your patents against them. This creates mutual deterrence that can prevent patent disputes from escalating.
Licensing revenue: Patents covering in-demand AI techniques can generate licensing revenue from other companies that implement those techniques.
What to patent:
- Novel AI architectures or methods that you have developed
- Specific application methods that are unique to your domain expertise
- Performance optimization techniques that provide measurable advantages
- Novel evaluation or monitoring methods
Patent filing considerations:
- Utility patents provide the strongest protection but take 2-3 years to issue and cost $15,000-30,000 per patent through prosecution
- Provisional patent applications provide a 12-month priority date at lower cost ($2,000-5,000), giving you time to assess the value of the invention before committing to full prosecution
- International patents (PCT applications, individual country filings) are expensive but necessary if you operate internationally
AI-Specific Patent Issues
Patentability of AI Inventions
The patentability of AI-related inventions is an evolving area of law.
Abstract idea rejections: Patent examiners frequently reject AI patent applications under 35 USC 101 as being directed to abstract ideas. Mathematical algorithms, mental processes, and methods of organizing human activity are not patentable subject matter. AI inventions must demonstrate a practical application or technical improvement beyond the abstract idea.
AI as inventor: Courts and patent offices have generally held that AI cannot be listed as an inventor on a patent. Human inventors must be identified, which can be complicated when AI plays a significant role in the inventive process.
Enablement and written description: AI patents must provide sufficient detail for a person skilled in the art to reproduce the invention. This is challenging for AI systems that rely on complex training processes—describing the architecture alone may not be sufficient without describing the training data, hyperparameters, and training methodology.
Open Source and Patent Interactions
Many AI frameworks, models, and tools are released under open source licenses. Some open source licenses include patent grants or patent retaliation clauses.
Apache License 2.0: Includes an express patent grant from contributors. Also includes a patent retaliation clause that terminates the patent grant if you initiate patent litigation against any contributor.
MIT License: Does not explicitly address patents, creating ambiguity about whether an implied patent license exists.
GPL family: The GPLv3 includes patent grants; GPLv2 does not explicitly address patents but may create an implied license.
Implications for agencies: If you use open source AI tools, understand the patent implications of the licenses. If you build on Apache-licensed tools, asserting patents against the contributors could terminate your right to use those tools.
Responding to Patent Assertions
If your agency receives a patent assertion (demand letter, licensing request, or lawsuit), respond strategically.
Do not ignore it. Ignoring a patent assertion can lead to enhanced damages (up to triple damages) if the patent holder can show willful infringement.
Engage patent counsel immediately. Patent litigation is specialized. Do not attempt to respond without experienced patent counsel.
Assess the assertion:
- Is the patent valid? Look for prior art, obviousness arguments, and eligibility challenges.
- Does your technology actually infringe? Analyze the patent claims against your specific implementation.
- What is the exposure? Calculate the potential damages and compare to the cost of licensing and the cost of litigation.
Consider your options:
- License: If the patent is valid and you infringe, licensing may be the most cost-effective option
- Design around: If you can modify your technology to avoid the patent claims, this eliminates ongoing liability
- Challenge validity: If the patent is weak, challenging its validity (through IPR proceedings at the Patent Trial and Appeal Board) can be effective and less expensive than full litigation
- Litigate: If the assertion lacks merit and settlement is unreasonable, litigation may be necessary
Your Next Step
Conduct a basic patent risk assessment for your agency. Identify the technical methods you use most frequently in client engagements. Search patent databases (Google Patents, USPTO PAIR) for patents covering those methods. You do not need a full FTO analysis at this stage—just develop awareness of the patent landscape in your technical areas.
Then review your client contracts for intellectual property indemnification provisions. Understand what you have agreed to regarding patent infringement claims related to your work. If your contracts include broad IP indemnification, consider whether your insurance covers patent infringement defense costs.
The agency that understands the patent landscape makes better technical decisions, negotiates better contracts, and avoids the costly surprise of an unexpected patent assertion. Patent risk does not go away because you ignore it—it accumulates until someone decides to collect.