Where in the Stack?

There is nothing wrong with the question. "Should we patent this?" reflects an awareness that IP matters, that something novel has been built, and that protecting it is worth thinking about. For most of the history of technology company formation, it was also roughly the right question. That is mostly because the thing being built and the IP strategy required to protect it were close enough together that one question could do most of the work.

AI has pulled them apart.

The question carries two embedded assumptions that made sense in a slower-moving technical environment and are leading somewhere less useful than it appears. The first is that the thing being evaluated is the right thing to evaluate. Filing a well-drafted patent on the wrong thing, at the wrong level of abstraction, is an expensive way to feel like you handled something without having handled it. The second is that there is time to think about this properly when the moment arrives. In AI, the moment the question arrives is almost always the wrong moment to be building the framework for answering it. There is a paper about to be published, a product about to launch, a competitor who may be filing on something adjacent. Urgency in IP strategy is the enemy of the upstream analysis that makes the downstream decision a good one rather than just a timely one. A patent application filed under time pressure, without the prior question answered, is often a patent application filed on the right invention at the wrong layer for the wrong business reasons. And by the time that becomes clear, the priority date is set, the claims are drafted, and the disclosure is permanent.

The question that produces a better answer is upstream of "should we patent this." is: where in the stack is our actual differentiation, how durable is it, can it be detected from the outside, and what do those answers tell us about the right instrument and the right timing?

The Stack is not Flat

Different layers of the technical stack have fundamentally different IP characteristics. What is worth patenting at one layer may not be worth patenting at another. What is protectable as a trade secret at one layer can be fragile at another. There are three layers worth understanding.

This is where the underlying models and architectures live. The training methods, model architectures, and pretraining approaches that produce general-purpose AI capability. Most founders think offensively here: can we patent this, can we own a piece of this space? In most technical spaces it would be the right one to follow.

The problem is that the foundation layer has been the subject of aggressive patent filings by Google, Meta, Microsoft, IBM, and others for years. The prior art base is enormous. Many of the techniques AI startups are building on top of are already claimed, in whole or in part, by someone with a much larger IP budget and a much longer head start.

The more useful question may be: given what the incumbents already own at this layer, can we build and ship our product without infringing something? That is the freedom to operate question, which is an intelligence exercise before it is a filing exercise.

This is where most serious AI startups are actually competing. Proprietary fine-tuning methods that produce measurably better results in a specific domain. Novel approaches to retrieval augmented generation that reduce hallucination rates in ways competitors have not replicated. Specific implementations of general AI capability that work differently, and often better, for a particular problem than anything else available.

This is also where the patent versus trade secret decision is most consequential and most often made badly. The technique is usually novel enough to be worth protecting. The competitive half-life is usually long enough to justify a patent prosecution timeline. The domain-specific application often provides the practical application hook that makes the §101 analysis more survivable than founders expect. And yet the filing decision here requires an analysis that most IP conversations never reach.

Two things sit at the customer-facing end of the stack. Both create some of the longest-lived competitive advantages in AI. Neither receives particularly coherent IP treatment.

The data layer holds the proprietary datasets, curation methodologies, synthetic data generation techniques, and training signal quality that determine what a model can do. This layer is almost never patented, since doing so would require disclosing the very thing that makes it valuable. But it is frequently under-protected as a trade secret. The governance burden is high, the reasonable steps standard is demanding, and as we discussed in the last issue, the embedded AI environment has made maintaining that protection more complicated than most organizations realize. The data layer's competitive half-life is often longer than any other layer in the stack. A dataset assembled and refined over five years is not something a competitor can replicate quickly regardless of how capable their models are. That durability makes it worth protecting rigorously. This means trade secret governance, not patent filings.

The integration layer focuses on how AI capability is embedded into workflows, user interfaces, and customer processes. This layer moves the fastest and is the hardest to protect with patents alone. By the time a patent on a specific workflow implementation issues, the implementation has changed several times. The moat here is organizational: speed, network effects, switching costs, and the accumulated understanding of what customers actually need. IP plays a supporting role rather than a central one.

The question is not: what did we build? It is: at which of these layers does our company's real competitive edge live, and specifically, the layer where, if a well-resourced competitor could freely replicate what we do, we would lose the thing that makes us hard to displace? That is the layer that deserves the most deliberate IP strategy. And the strategy it deserves depends on where that layer sits in the stack.

Three Variables That Determine the Right Instrument

Knowing which layer you compete at narrows the decision. Within any layer, there are multiple possible instruments (patent, trade secret, or some combination) and choosing the wrong one for the right layer is nearly as costly as choosing the right one for the wrong layer.

A technique that leaves enough of a trace in the outputs that a sophisticated competitor could reverse engineer the method from the results will not remain a secret regardless of how carefully you guard it internally. The protection will erode from the outside in. When a technique is detectable, patent protection is almost always the better choice, because the exclusivity a patent provides will outlast the secrecy that trade secret law requires.

A technique that produces a result but leaves no recoverable signal about the method that produced it is a different situation. Patenting requires permanently disclosing what is currently invisible to everyone outside the company. The question becomes whether the exclusivity that disclosure buys is worth more than the secrecy it destroys. Often it is not.

Detectability is one of the few IP variables that genuinely requires both technical and legal judgment to evaluate well. What the outputs reveal about the method is a technical question. What that means for the choice between patent and trade secret is a legal one. The two questions belong in the same conversation.

A patent takes two to three years from filing to grant in most AI-related technology areas. The question is whether what you are protecting will still matter when the protection arrives.

A technique with a competitive half-life of five years or more is worth patenting even accounting for prosecution timelines. A technique likely to be superseded within eighteen months may not justify the same investment. Sometimes the priority date itself has value independent of whether the specific claims remain commercially relevant. Establishing prior art position can matter at the foundation layer regardless of what the claims ultimately cover. But for most startups at the core technology layer, the simpler question applies: will this still matter when the patent arrives?

The data layer sits at the far end of the durability spectrum. A proprietary dataset assembled and refined over years is more durable than almost any model architecture or training method. The longer the competitive half-life, the more the cost of maintaining adequate protection is worth bearing.

Patenting requires telling the world exactly how your invention works. That disclosure is permanent, public, and irrevocable.

Consider a novel inference optimization method that reduces latency at deployment that produces a measurable performance advantage. If that method is visible in the product's response characteristics, if a sophisticated competitor running benchmarks could reasonably reverse engineer the approach, and if the underlying technique is likely to be independently developed by others working on the same problem, then the disclosure a patent requires costs very little extra. The information will find its way into the competitive landscape regardless. The patent simply ensures that when it does, the company that developed it first has something to show for the effort.

Now consider a company whose competitive advantage lives in how it curates and weights training data, a proprietary process developed over years of iteration that produces measurably better model behavior in a specific domain. The process is not visible in the outputs. Competitors cannot infer it from the product. Nobody outside the company knows it exists in the form it does. Patenting it would require describing it in enabling detail in a public specification, the exact curation logic, the weighting methodology, the iterative refinement process that took years to develop. The exclusivity a patent provides may be worth considerably less than the secrecy being surrendered to obtain it. A trade secret that competitors do not know exists is often more valuable than a patent that tells them exactly what to build around.

The honest version of this variable requires asking: if we filed and the specification published tomorrow, what would our most capable competitor do with it? If the answer is "not much that they couldn't do already," the disclosure tradeoff favors patenting. If the answer is "they would have a detailed description of exactly what we spent two years figuring out," the tradeoff deserves more scrutiny.

How the Variables Work Together

The three variables interact and the interaction is the analysis, with no single variable being determinative.

A technique that is highly detectable, highly durable, and where disclosure adds little to what competitors could learn anyway is a strong patent candidate. File, draft the claims carefully to maximize scope while surviving examination, and build a continuation strategy that tracks the product as it evolves. A technique that is genuinely invisible in its outputs, likely to be superseded within eighteen months, and where a detailed specification would materially benefit competitors is a strong trade secret candidate. Govern it rigorously under the reasonable steps standard, limit access carefully, and accept that the protection is fragile but currently adequate.

Most real situations fall somewhere between those poles, which is why the framework requires evaluating all three variables together rather than leading with the one that points toward a preferred answer.

Filing Strategy Follows Stack Position

The three variables tell you which instrument fits the differentiation you have. Deploying it well depends on which layer you are working from.

The IP posture at the foundation layer for most startups is an intelligence program. Understanding what the incumbents have claimed, how broadly those claims read, and where your product sits relative to the landscape is a defensive necessity.  This is because the decisions made about product architecture, about which open source foundations to build on, and about which third-party APIs to integrate are all implicitly IP decisions. Making them without understanding the landscape is how companies build products they cannot commercialize without a license they did not know they needed.

A freedom to operate analysis is the foundation layer's IP investment. It is the prior work that makes everything above it strategically sound.

This is where the patent investment can produce the strongest return for most AI startups and where the filing strategy requires the most deliberate design.

File earlier than feels necessary. The urgency that compresses the upstream analysis should not compress the filing decision. However, the prior art base is expanding every month and the window to establish meaningful priority at the core technology and application layer narrows with every paper published and every open source release. A provisional application filed before public disclosure preserves your priority date at relatively low cost and buys twelve months to develop the full strategy. Use that time.

File at the right level of abstraction. The most common filing mistake at this layer is claiming too narrowly, i.e., covering the specific implementation rather than the underlying method. A patent that covers exactly how you did something today may not cover how you do it eighteen months from now, and it certainly will not cover how a competitor does something functionally equivalent through a different implementation. The claims should be as broad as the prior art and §101 analysis permit, with dependent claims providing fallback positions if the broad claims face challenges.

Build a continuation strategy from the get-go. This is the mechanism that makes a portfolio relevant in a fast-moving space and it is the most consistently underused tool in the startup IP toolkit. The initial application establishes the priority date and the foundational disclosure. Continuations allow you to pursue new claims covering specific product implementations, specific use cases, specific technical improvements that are supported by the original specification but were not the focus of the original application. A continuation family that grows with the product provides coverage that remains current as the technology evolves. The portfolio that survives a fast-moving space is one that was designed to evolve.

The application layer deserves specific mention here because it is more patentable than most founders believe. A specific technical implementation of general AI capability in a specific domain, such as medical imaging, legal document analysis, financial risk modeling, materials discovery, etc., often provides the practical application framing that the current USPTO guidance requires for §101 survivability. The combination of a general method and domain-specific technical knowledge produces something that is neither the general method alone nor the domain knowledge alone, and that combination is frequently more defensible than founders expect.

The data layer's IP strategy is often a governance question, similar to what we examined in the last issue. The competitive advantage at this layer lives in the combination of what the data contains, how it was assembled, and the organizational knowledge required to use it effectively. Patenting would require disclosing all of that. The instrument is trade secret protection, maintained through the reasonable steps standard.

The integration layer's IP strategy is primarily speed and organizational capability, supported at the margins by defensive patent filings. File enough to deter and to have something to cross-license. Do not expect patents to provide the primary protection at a layer where the product evolves faster than the patent system moves.

The sequence is not complicated. Locate the layer where your company would lose something irreplaceable if a well-resourced competitor could freely replicate what you do. Ask the three variables for that layer. Then ask what the answers tell you about the instrument and the timing. In a space moving as fast as AI, the window to establish a meaningful position narrows every month. The prior art accumulates, the incumbents file, and the continuation opportunities that were available from a well-timed foundational application become unavailable once the technology has moved past the point the original specification describes.

The right question is not should we patent this, but where does our real differentiation live, and what does that tell us about how to protect it.