Patent protection is pursued for all types of technologies. Why should anything be different just because the technology is based on artificial intelligence (AI)?
Nothing is different when reduced to first principles. But in practice, old habits are easier to follow than first principles. Those habits are ripe for challenge in an AI era. Consider the following example.
A company designs a commercially valuable mechanical safety lock to prevent a child from accessing a cabinet. Should they patent it? Start with a patentability search because the art of mechanical safety locks is well trodden. Assuming novelty, a patent can be drafted that attempts to protect the lock broadly using functional language to capture the means achieved, rather than the specific implementation. The value proposition for obtaining a patent is clear. Other market players can reverse engineer the lock, but the patent will prevent them from manufacturing or selling the lock. The patent aims to protect useful variations of the lock falling within the inventive concept. This sets the company up for a monopoly in the marketplace. The company’s revenue stream is protected.
Now consider another company that also designs products for child safety. This company uses AI to analyse images to determine whether the images are suitable for viewing by a child. The AI can traverse a database storing millions of images and filter out images that are inappropriate. Should they patent it?
First, AI quickly advances and is not bound by the laws of classical mechanics. The invention is based on a trained neural network that receives an image and outputs a probability that the image is inappropriate. Within two years, the approach is obsolete and the company pivots because large language models (LLMs) emerge and are superior. The image is instead tokenized and input into a multimodal LLM that is prompted to classify the image as either appropriate or inappropriate. The focus is on fine-tuning and prompt engineering. Two years later, the approach is obsolete and the company pivots because agentic AI emerges and is superior. An orchestrator agent determines that parental control should be turned on and generates an action plan. The plan includes retrieving images via a tool call and instructing a specialized agent utilizing a low-latency LLM to classify each image. Another specialized agent judges the classification and, only when confidence in the classification is low, re-evaluates using a large and powerful LLM. The combination results in lower overall latency and superior image filtering. Will the company pivot again in two years? What’s next?
Second, a patentability search has little value because the technology is emerging too quickly. Everything within the last 1 ½ years, i.e., the most relevant, is not yet even published in the patent databases. The inventors presumably already know about everything relevant that is easily findable by internet searching.
Third, the invention is hidden. How would the competitor copy? How would the company know who is infringing?
Fourth, patent offices around the world cast a wary eye on software-based inventions. A difficult road may lie in front of the patent office with little to show at the end.
Fifth, the race to market for applying AI technologies is on. The company does not want to be slowed down by the financial and human resource cost of preparing a patent application and doing it before the company discloses the invention. Besides, the main inventor is a proponent of open source and does not see the value in patents.
The list could go on.
Is the answer “do not patent AI”? Sometimes that is the answer. But sometimes the answer is the opposite: you should patent it, but with the right focus to do it better.
Three hypothetical case studies involving companies with AI inventions are presented. In each, the conclusion is to pursue patent protection, but it is only done right when you shift the focus.
Case study #1 – SmallCo in food distribution
This case study takes inspiration from Zest Labs, Inc (“Zest”) and its trade secret dispute with Walmart. A few comments on Zest later.
SmallCo has created technology to mitigate the problem of fresh food waste. As produce is sorted for distribution, different characteristics of the produce are measured, such as its moisture content, physical damage, mold, etc., in part by analyzing an image of the produce. A trained machine-learning model utilizing a neural network receives the measured characteristics and predicts shelf life. Produce with longer shelf life is sorted for shipping a longer distance and vice versa.
Through their R&D efforts, SmallCo discovered the specific combination of characteristics to measure for best performance, trained the machine-learning model using supervised learning based on experiments they conducted linking produce characteristics to actual shelf life and wrote the code for deployment of the AI on the food distribution equipment owned by their customers.
SmallCo’s next step is customer collaboration: present the solution to a key customer under a confidentiality agreement and partner with that customer to test it out.
Should SmallCo obtain patent protection?
There is not much value in patenting the AI. The training method is only occasionally performed and hidden. The AI method applied is also hidden and will likely change in ways not yet imagined, like the way traditional machine-learning classification is moving to LLMs, agentic AI and beyond.
So, forgo the patent? No – shift the focus. The situation at hand is not about AI. Deploying the AI is the implementation, the sweat of the inventor that is proudly shown to the patent agent. But the point is not to award the inventor with a patent for their effort. Patent protection should only be pursued if it aligns with and protects the business opportunity.
The business opportunity is the deployment of the code on the customer’s food distribution equipment. Collaboration with a customer is about to begin under a confidentiality agreement. A confidentiality agreement offers limited protection. SmallCo’s best move is to try to control access to what is market essential – the minimum viable product, which is operation of food distribution equipment to sort produce based on measured characteristics indicative of shelf life. One could use AI to perform the sorting, but maybe not. One could measure the specific characteristics that SmallCo measures, but maybe not.
The business opportunity is lost if the collaborating customer does not adopt SmallCo’s system but instead develops their own in-house solution. In such a situation, SmallCo would be well served to have more in their back pocket than a complaint seeking relief based on a confidentiality agreement. A patent would bolster SmallCo’s position and help defend the business opportunity. A patent that aims to control access to the minimum viable product. For example, a patent claim to a “food distribution system” that “estimates physical properties of produce, determines an indicator related to shelf-life based on the estimate and sorts each piece of produce based on the indicator”. Do not get hung-up on patenting the cool tech, i.e., the AI. Go broad, which is more important. Then see how patent examination unfolds.
Returning to Zest, Walmart showed an interest in Zest’s AI-based food distribution technology. Zest shared trade secrets, proprietary algorithms, etc. under a confidentiality agreement. Walmart subsequently ended their engagement as a potential customer and launched a competing platform. In May 2025, a jury found Walmart had willfully and maliciously stolen trade secrets, awarding $222.7 million to Zest. But Zest’s win is not a success story. The business opportunity had much more potential and was lost.
Hindsight is 20/20. However, in the hypothetical fact pattern at hand, SmallCo should be careful in their collaboration with a customer. There is more potential risk in the collaboration than from an unsuspecting competitor. Disclose the minimum amount of technical detail necessary for the collaboration, even under a confidentiality agreement. And, yes, SmallCo should pursue a patent. However, the patent should try to control access to the technology broadly, not the AI component.
Case study #2 – BigCo in software
BigCo is a dominant market player in its niche software. Patents played no role in BigCo’s success. If anything, patents are a threat to its dominant market position. BigCo understandably has no tolerance for patent trolls, but BigCo also wishes it could do away with the patents owned by competing market players.
BigCo’s flagship software incorporates an LLM-powered chatbot. BigCo has determined that users who have the most engaging interactions with their chatbot are more likely to convert into customers. The developers at BigCo have enhanced the chatbot to increase user engagement. A repository of the most engaging chatbot interactions with previous users is embedded and stored as vector entries in a vector database. Then, when a new user interacts with the chatbot, the reply generated by the LLM powering the chatbot is embedded and compared to the embeddings in the repository. If the reply generated by the LLM is not close enough to an engaging interaction in the repository, the LLM is prompted to reply differently to nudge the interaction in a more engaging direction.
Should BigCo pursue patent protection for their chatbot enhancement?
It is easy to say “no”. BigCo is not a proponent of patents. The chatbot enhancement is not a market-essential feature. The enhancement is hidden, and its implementation could change in ways unimagined in the coming years as AI develops. More importantly, the developers are busy and laser-focused on software delivery. Patents are the last thing on their minds.
However, BigCo operates in a market in which patents exist. BigCo is no longer a scrappy startup, but a sophisticated entity, a target and a leader in the technology. BigCo recognizes that budget and thought need to be applied not only to grow a patent portfolio for defensive purposes, but also to help show their market dominance and leadership in the technology. It is not about raising barriers – they can make a patent pledge like Tesla to show they are progressive, collaborative and altruistic. It’s about demonstrating that they are a leader, that they own the space and that they will always reward behaviour that builds the ecosystem in which they play.
The best defensive patents for BigCo are the offensive ones – the commercially relevant ones that purposely target the technology adopted by the other market players. However, the chatbot enhancement is not one of those. Nonetheless, BigCo understands that it is a numbers game to an extent. Size and breadth of the patent portfolio matter, even if it is cluttered with patents whose future value is unknown and questionable. Chatbot enhancement is one of those.
Chatbot enhancement seems worth adding to their patent portfolio, but the patent is of mediocre value. Should BigCo call upon the mediocre patent agent that is less expensive? After all, apply an effort commensurate with the perceived value, right? No – shift the focus. What is most valuable is not the patent, but the developer’s time. BigCo wants a patent involving as little of the developer’s time as possible and that can be obtained with as little friction as possible. Bring the right person to the table, the one who uses as little of the developer’s time as possible, who does not need a written invention disclosure from the developer but just enough of the developer’s time for an oral explanation and who aims up front to make prosecution as smooth as possible by framing the invention as a type that can be patent-eligible. A patent office will cast a wary eye on an invention described as solving the problem of chatbot user engagement. A thoughtful and creative framing will increase the chance of smooth prosecution. For example, framing the invention not as enhancing user engagement with a chatbot (as described by the developer), but as a technical solution that mitigates the problem of LLMs producing irrelevant content or hallucination, e.g. due to insufficient training.
BigCo should not take a page from the playbook of classic invention harvesting, the playbook that requires inventors to submit comprehensive written invention disclosures, followed by internal patent review board meetings, followed by patentability search and followed by farming out to an external patent agent for drafting in collaboration with the inventor. BigCo should shift the focus to an agile route. A process that can move quickly and easily for all involved, especially for the developers.
Case study #3 – Startup in investment advising
Startup InvestAI intends to disrupt the investment advising industry with its AI advisor platform. The platform is under development and soon to be launched. Budget is tight, but runway is in place for achieving product-market fit within a year of launch, followed by Series A funding.
InvestAI utilizes AI to make it easy for unsophisticated consumers to learn about, buy, and sell investments. The underlying technology uses an LLM that powers a chatbot. The LLM has a large context window that includes the definition of various tools and when to call them. A variety of tools are developed and more are being added, including tools to perform internet and database searching, maintain user profiles, and interface with automated broker platforms.
InvestAI is iterating on the technical underpinnings of their platform but has arrived at something fit to launch. They can squeeze one patent out of their budget and wonder whether filing a patent makes sense. The patent agent tells them that VC investors consider patents important (see page 1307), and they should file a patent before launching to market because once you disclose the invention you are barred from obtaining a patent in certain countries. The mantra in the patent world is “file early and file often”. The patent agent warns them that the general concept is not patent-eligible, but they can fall back as needed to a technical detail, e.g. details related to the specific tools and tool calling.
Should InvestAI pull the trigger on the patent?
No, not yet – shift the focus. InvestAI is in a precarious position. The platform has not yet launched and there is no product-market fit. Budget is constrained to one patent filing. The general concept is not patent-eligible, and so the technical underpinnings need to be relied upon, but such details are mostly hidden and in flux. InvestAI admits that they could easily pivot on the UX and technical underpinnings. A premature patent filing may end up focusing on technical details that looked good at launch but are of no commercial interest after a pivot towards the product-market-fit offering. InvestAI cannot afford to risk their “one patent” budget on hidden technical details of no commercial interest.
InvestAI should launch the product and find product-market fit. Then, within a year of launch, file for patent protection on a market-essential technical underpinning that is part of the product-market-fit offering and that is discoverable. By waiting on filing a patent application, InvestAI may very well be barred from obtaining a patent in some countries on anything they publicly disclosed at initial launch, but InvestAI can refrain from publicly disclosing technical details. Moreover, some important countries like the U.S. have a one-year grace period that may be relied upon to protect details publicly disclosed.
The patent application can still try to protect the broadest concept possible related to the market-essential technical underpinning, subject to previous disclosure. Such subject matter may face a hurdle in examination, but there is value in having such a patent application pending, waiting for examination, even if the broader concept ends up being rejected by the patent examiner. Patent pendency can be drawn out as long as possible if desired, e.g. by filing a PCT application to delay examination in the national patent offices. When it is time for substantive examination to begin, the patent application can be considered anew to confirm that what is being pursued still aligns with the market. By then, InvestAI will hopefully have secured their Series A funding and be well on their way.
Conclusion
What can we take from the case studies above? First, not every technology has the same patent trajectory. Second, different circumstances require different approaches. Third, market position of the company matters. Fourth, in each circumstance, shift the focus to the right spot, apply first principles, and the rest should take care of itself. Sometimes it means patenting less AI, sometimes not. Sometimes it means patenting early and often, sometimes not. Sometimes it means pursuing patent applications that have questionable patent eligibility, sometimes not, etc. In any case, if the patent agent has the right mindset backed by the hard skills, you will be in good hands.
How about the original example? The company focused on child safety, that uses AI to determine whether images are suitable for viewing by a child. Should they file for patent protection? I’ll let you know once I speak with them and we shift the focus to the right spot.
Stay tuned
Stay up to date on IP law developments in Canada in 2026 through our IP Insights and Rx IP Updates. Subscribe to receive these in your inbox throughout the year.
The preceding is intended as a timely update on Canadian intellectual property and technology law. The content is informational only and does not constitute legal or professional advice. To obtain such advice, please communicate with our offices directly.