Your product team just shipped a new artificial-intelligence feature. Customers are using it. Internally, everyone is proud of it. And if you are like many founders I talk to, nobody has asked the next question yet: Is any part of this ours to protect, or did we build a useful product around tools anyone could buy?
That question gets skipped because “AI feature” sounds like one thing. It is not. It is a stack of different components, and they do not all present the same legal or business considerations.
An AI Product Has Layers
When I look at a software or artificial-intelligence product with a founder, I find it useful to separate the components instead of treating “the AI” as a single block.
The data layer includes how information is collected, cleaned, labeled, structured, and prepared before it reaches a model. Two companies using the same model may produce different results because their data pipelines and preparation methods differ. Whether those differences support legal protection depends on what the team actually devised and whether the applicable requirements are met.
The model layer is the part many people assume must be the valuable one. Sometimes it is. The relevant questions include whether the team built or materially modified the model and, if so, what technical contribution those changes represent. Calling an existing model through an application programming interface creates a different starting point, but it does not end the inquiry.
The workflow layer includes how the system handles the model’s output, what rules operate around it, when the system requests human review, when it acts automatically, and how it combines the output with other information. This layer may contain important engineering and business decisions even when the underlying model came from another source.
The application layer is what the user experiences. The useful question is whether that experience depends on a specific technical implementation or primarily reflects a desired result that others could reproduce through conventional tools.
None of these layers is automatically patentable, and none is automatically excluded. The analysis turns on the claimed invention, including whether it is eligible subject matter, new, and nonobvious.
Using AI Does Not Answer the Patent Question
Incorporating artificial intelligence into a product does not, by itself, make the product patentable or unpatentable.
“We used AI” does not identify a technical contribution. Patent eligibility also requires more than attaching generic computer functions to an abstract idea. Conventional computer components performing their ordinary functions do not necessarily supply the additional inventive concept required for eligibility.Papst Licensing GMBH & Co. v. Xilinx Inc., 193 F. Supp. 3d 1069, 1083 (N.D. Cal. 2016).
The inquiry becomes more useful when it focuses on what the team actually devised. That may be a particular method of structuring data, a specific architecture, a technical control over model behavior, or a validation process that changes how the system operates.
Software can qualify as a nonabstract improvement when the focus of the patent claims is a specific improvement in computer capabilities rather than an abstract process for which a computer is merely being used as a tool.BSG Tech LLC v. Buyseasons, Inc., 899 F.3d 1281, 1285 (Fed. Cir. 2018).
That does not mean every detailed workflow is eligible or patentable. It means the analysis should focus on the mechanics and the proposed patent claims, not the appearance of the term “AI” in the product description.
What This Means in Practice
In my experience, this discipline applies well beyond artificial intelligence. I have evaluated inventions involving mechanical and electromechanical systems, casino operations, medical accounting platforms, educational technology, and computer networks. The useful question remains: What did the team build, and what is technically meaningful about the difference?
For an artificial-intelligence feature, the answer may involve a method of preparing training data, a workflow that combines model output with other systems, or a process for validating or correcting results before they reach the user. These examples identify areas to investigate; they do not establish patentability.
Whether a particular component should be evaluated for a patent, maintained as a trade secret, protected through copyright, addressed through contracts, or simply documented internally depends on what the team built and how the company uses and safeguards it.
Patent and trade-secret protection also involve different strategic choices. A patent application generally requires disclosure of the invention, while trade-secret value depends on information remaining secret and being handled accordingly. The right choice may differ from one layer of the system to another.
Please note that whether the company may rely on trade-secret protection, access controls, confidentiality practices, and contractual restrictions should be considered before the information is shared more broadly. If patent protection may be appropriate, public disclosure and filing timing should be evaluated before release.
Why the Review Matters Even If You Never File
Not every component of an artificial-intelligence system belongs in a patent application. Some information may be better handled as a trade secret. Some material may be protected by copyright or contract. Some features may not support meaningful exclusivity at all.
Knowing that is still valuable. It allows the company to understand where its business advantage may come from and whether its internal practices match the protection strategy it intends to use.
This review is easier before an investor, acquirer, or competitor forces the question. A founder facing due diligence may be asked to explain exactly what the company owns, what makes the technology difficult to reproduce, and what measures the company has taken to protect it. Mapping the system in advance makes that conversation more productive.
A Measured Next Step
If your company has released an artificial-intelligence feature within the past year and has not examined its separate layers, consider doing that before the next product release or financing discussion.
Sit down with the people who built the feature and ask what technical decisions they made at the data, model, workflow, and application layers. Ask which parts came from existing tools and which parts required the team to devise a particular solution.
If the answer is that the team combined available tools in a conventional way, that is useful information. If the discussion reveals a specific technical approach that appears meaningfully different, it may justify a closer review of patentability and other protection options.
The Garcia-Zamor Law Firm serves as fractional general counsel for innovators, helping protect business operations and intellectual property. Ruy Garcia-Zamor leads business-growth strategy as a registered United States patent attorney. Elliott Alderman handles intellectual-property matters, Claudia Castillo focuses on employment law, and registered patent attorney Amulya Annasamudram focuses on patents and intellectual property. Contact the firm through its website or at (410) 531-9853.
