Patent Strategy for AI Companies

AI companies that come to us with a "we need patents" question almost never actually need that conversation. The question that drives strategy is different: what does a defensible portfolio look like for a company whose product architecture is going to change three times before Series B, whose competitors are also iterating, and whose investors care more about IP than they did even three years ago. That question is harder to answer for AI companies than for most other software businesses. Eligibility under 35 U.S.C. 101 is the central battleground for AI claims in a way it simply isn't for most mechanical or chemical inventions. Model architectures change faster than prosecution timelines. The line between platform claims and product claims hits differently when both the platform and the products are moving. And open-source components in modern AI stacks create freedom-to-operate questions that didn't apply to earlier software waves. This is what AI patent strategy looks like in 2026: less about whether to file, more about which surfaces to cover, how broadly to scope claims, and how to coordinate filings with architecture changes and funding milestones.

Why AI patent strategy is structurally harder than other software patent strategy

Three structural factors make AI patent work harder than typical software patent work, and each shapes how the portfolio gets built. Eligibility is the distinctive hurdle — but the trend is misunderstood. It is widely assumed that eligibility scrutiny for AI has been steadily tightening. At the USPTO examination level, the opposite has been true since Alice Corp. v. CLS Bank (2014) caused the initial spike in Section 101 rejections for software and AI. The USPTO's "Adjusting to Alice" study found that the 2019 Revised Patent Subject Matter Eligibility Guidance (2019 PEG) produced roughly a 25% decrease in the likelihood that Alice-affected technologies — including AI — would receive a first office action with a 101 rejection, along with a 44% drop in eligibility-determination uncertainty. The 2024 Patent Subject Matter Eligibility Guidance Update, Including on Artificial Intelligence (the "2024 AI SME Update," 89 FR 58128, effective July 17, 2024) went further, supplying worked examples of AI claims that do not even recite an abstract idea. A follow-on examiner memo (August 4, 2025) directed the software art units to stop over-applying eligibility rejections, and a precedential decision, Ex Parte Desjardins (2025), expressly upheld a machine-learning training claim. The eligibility risk hasn't disappeared — it has migrated. The residual uncertainty now lives more at the Federal Circuit than at the examiner's desk: in Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), claims that applied generic machine learning to automate an abstract idea in a new field were held ineligible. So the practical point for founders is not "the door is closing." It is that eligibility for AI turns on a specific, learnable distinction — and the cost of getting it wrong is concentrated in claims that read as generic AI pointed at a problem, rather than as a particular technological improvement. Architectures change faster than prosecution timelines. A patent application commonly takes on the order of 18–36 months from filing to allowance, and the software-heavy art units where AI is examined (Technology Centers 2100, 2600, and 3600) tend toward the longer end of that range. For an AI company, that is enough time for the core model architecture to change once or twice. A claim drafted around a 2024 transformer implementation may describe technology that has been deprecated by the time of allowance. Provisional filings and continuation practice give companies optionality, but the coordination between engineering roadmap and prosecution timing matters more here than in slower-moving fields. The platform-versus-product question is sharper. Most AI companies have both: a core platform — the model architecture, training methodology, infrastructure — and one or more deployed products built on that platform. Patents can cover the platform broadly, the specific products narrowly, or both. The right weighting depends on competitive position, funding stage, and what competitors are likely to copy. Platform-level claims are harder to draft and prosecute but carry more strategic value. Product-level claims are easier to obtain but cover less. Open-source components complicate freedom-to-operate. Modern AI stacks heavily incorporate open-source components — base models, training frameworks, tokenizers, evaluation libraries. Each potentially carries patent exposure that the deploying company inherits. FTO analysis for an AI product means looking at a much larger surface area than for a typical software product.

What the eligibility test actually asks

Because eligibility is the controlling issue, it pays to be precise about what the analysis is — not "measurable performance," which is not the standard, but the two-pronged framework the USPTO actually applies under Step 2A of its guidance. Step 2A, Prong One asks whether the claim recites a judicial exception. For AI, that almost always means one of three enumerated groupings of abstract ideas: mathematical concepts, certain methods of organizing human activity, and mental processes. Two distinctions decide most AI cases here:

• Reciting versus merely involving. A claim that merely involves or is based on a mathematical concept does not recite one. The USPTO's own examples illustrate the line: claiming "training the neural network" merely involves math and does not recite an abstract idea, whereas a claim that names the calculation — for example, "a backpropagation algorithm and a gradient descent algorithm" — recites a mathematical concept and triggers further analysis. The drafting lesson is direct: naming the specific math in the claim can pull an otherwise-eligible limitation into Prong One.
• The mental-process limit. The "mental processes" grouping covers only what can practically be performed in the human mind. Claim limitations that encompass AI in a way the human mind is not equipped to perform — analyzing network packets in real time, for instance (SRI Int'l v. Cisco) — do not fall within the grouping at all. Step 2A, Prong Two asks whether the claim as a whole integrates the recited exception into a practical application. For AI claims, the dominant route to eligibility is showing that the claim reflects an improvement in the functioning of a computer or to another technology or technical field — the search for a technological solution to a technological problem (McRO; Enfish). The pivotal distinction, and the one that sinks most AI claims, comes from the Stanford cases: improving the abstract idea itself (e.g., a more accurate statistical or mathematical prediction) is not a technological improvement and remains ineligible, while improving how the computer or model operates is. The improvement must be disclosed in the specification with enough detail that a person of ordinary skill would recognize it — not asserted in a conclusory way — and the claim must actually reflect it. Ex Parte Desjardins (PTAB Appeals Review Panel, Sept. 26, 2025; designated precedential and incorporated into the MPEP in December 2025) is the cleanest current illustration on the eligible side. The claim covered training a machine-learning model across successive tasks. The panel agreed the claim recited a mathematical concept at Prong One, but found it eligible at Prong Two because the specification described how the model was trained to learn new tasks while protecting prior knowledge — solving "catastrophic forgetting" in continual-learning systems, and yielding reduced storage and system complexity. Those were improvements to how the model itself functions, not merely to the underlying math. Recentive (Fed. Cir. 2025) is the mirror image: generic ML deployed to automate a process in a new field, with no claimed improvement to the technology, was ineligible. Most AI eligibility work is, in practice, the work of landing on the Desjardins side of that line. One more practical point founders rarely hear: examiners are now instructed that on a genuine close call they should reject only when it is more likely than not — over 50% — that a claim is ineligible, and not merely because they are uncertain (August 4, 2025 memo; MPEP 706's preponderance standard). That shifts the benefit of a well-drafted close call toward the applicant.

The three orientations we see in AI patent portfolios

When Lynch LLP works on an AI company's patent strategy, the conversation usually surfaces one of three orientations — sometimes a combination — depending on what the company is trying to defend. Platform-genus orientation. The company wants protection for the core architecture itself: model design, training methodology, inference infrastructure. Claims target the technology that is reused across multiple products. This orientation makes sense for platform companies whose competitive advantage is the technology stack rather than a specific application, and for companies that anticipate licensing the underlying technology. Lead-asset species orientation. The company wants protection for a specific deployed product. Claims target the particular implementation — the way the model integrates with a specific data pipeline, a specific user workflow, a specific output format. This makes sense where the competitive advantage is in the application layer, where the product's value comes from how the technology solves a specific customer problem. Defensive blocking orientation. The company wants claims that prevent competitors from approaching the same problem space. Claims may be drafted with intentional breadth to cover variants the company doesn't itself ship. This makes sense where the competitive risk is later entrants, where freedom-to-operate matters more than direct enforcement, and where standards-setting or M&A optionality is part of the strategy. Most mature portfolios use some combination of all three. The question usually isn't which orientation to choose, but how to weight them. A Series A AI company with a single product and a small engineering team might lean 70% lead-asset, 20% platform, 10% defensive. A platform company at Series B with multiple downstream products and licensing ambitions might invert that ratio. The right weighting changes as the company scales. The orientation discussion is the first thing Lynch LLP works through with founders and CTOs, because every claim-drafting decision downstream gets framed by it.

Claim scope trade-offs: platform genus vs. lead-asset species

The platform-versus-product framing surfaces a concrete drafting question on every claim: how broad to scope. Broader claims cover more variants — including variants the company hasn't built yet, and variants competitors might build. Broader claims also face more eligibility scrutiny and a higher rejection rate. Narrower claims describing specific implementations tend to face less eligibility risk but cover less of the strategic surface. Neither answer is right in isolation. The trade-off varies by what's being claimed. For a genuinely novel model architecture, broad claims may be defensible if the architecture is new and the specification makes the technological improvement concrete and apparent to a skilled artisan. For an application of a known model architecture to a new problem domain, broader claims often run into eligibility rejections because the application reads as a routine deployment of existing technology — the Recentive problem. The trade-off also varies by company stage. A pre-product-market-fit company has less certainty about which surfaces matter, so broader claims hedge against being narrow on the wrong dimension. A scaled company with a stable product has clearer visibility into what matters and can prosecute more focused claims efficiently. Lynch LLP frames this trade-off explicitly with clients at the drafting stage. The conversation is usually less "how broad can we go" and more "what coverage profile does this filing buy us, and what coverage do we expect from later filings." Single-filing strategies that try to cover everything in one application tend to face the toughest eligibility battles. Multi-filing strategies that stage claim breadth across an initial filing and one or more continuations spread the risk and build coverage incrementally.

Timing considerations

Patent prosecution timing for AI companies sits at the intersection of three different clocks: the engineering roadmap, the funding cycle, and the public-disclosure pipeline. Engineering roadmap. Major architecture changes — model swaps, training-methodology shifts, infrastructure overhauls — affect what's worth claiming and when. Filing on an architecture the company is about to deprecate burns money. Filing too late, after a new architecture has been publicly demonstrated, can foreclose patent rights entirely. Coordination between engineering leads and patent counsel matters here, particularly around when an architecture is stable enough to commit prosecution resources to. Funding cycle. Most AI companies face their first serious IP review at Series A diligence. Investors increasingly evaluate IP defensibility as part of technical due diligence, and gaps surface late in the process if filings haven't been planned ahead. The timing question isn't usually "before or after Series A" — most companies file before — but "what coverage will the Series A reader expect to see." Lead-asset claims and at least one platform-level filing are typical expectations for AI companies entering Series A. Public-disclosure pipeline. Conference talks, blog posts, GitHub releases, academic publications, customer demos — all of these are public disclosures with patent-timing implications. AI companies tend to disclose more publicly and earlier than companies in other fields, both because the technical community moves through publication and because customer demonstrations are part of how AI products are sold. Coordinating filing dates with the disclosure pipeline is more intricate for AI companies than for, say, hardware companies that ship products in finished form. (The U.S. one-year grace period under 35 U.S.C. 102(b)(1) is a backstop, not a plan — and it does not exist in most foreign jurisdictions, so a public disclosure can forfeit foreign rights on day one.) Lynch LLP coordinates these three clocks as part of patent strategy work — mapping the engineering roadmap against the funding cycle against the disclosure pipeline so that filings land in the right windows.

The practical questions Engineering and Legal navigate together

In our experience, the most productive AI patent work happens when patent counsel sits directly with engineering leads — not just at the inventor-interview stage, but earlier, during architecture planning. Inventor interviews structured around architecture diagrams. For AI work, the best interviews work from the actual system architecture: model card, data-pipeline diagram, training-loop pseudocode, evaluation methodology. Because eligibility turns on a disclosed technological improvement, the interview's real job is to surface what specifically gets better and why — the equivalent of the "catastrophic forgetting" story in Desjardins. Claims that emerge from architecture-grounded interviews tend to be far more defensible than claims that emerge from abstract problem statements. Provisional-to-non-provisional planning around architecture changes. Provisional applications give the company a year of optionality. For AI companies, that year often includes architectural changes. Planning the non-provisional content around the most-likely-stable elements of the architecture — and explicitly anticipating which elements may change — produces better continuation pathways than treating the provisional as a frozen draft of the eventual non-provisional. Continuation strategy for evolving product surfaces. Continuations let companies file additional claims off the same initial disclosure as the product evolves. For AI companies, continuation practice is often the primary mechanism for keeping coverage current as the product surface changes. The first filing is rarely the whole strategy; the continuation tree built off the initial disclosure usually matters more. Trade secrets versus patents for model weights and training methods. Some elements of AI technology are best protected as patents — architectures, training methodologies, system designs that can be claimed as technological improvements. Others are best protected as trade secrets — specific model weights, fine-tuning data, evaluation benchmarks. The allocation depends on what is enforceable through patent claims, what is discoverable through reverse engineering, and what the company can keep confidential operationally. Lynch LLP works through this allocation with founders rather than defaulting to either category. A note on AI-assisted inventions. A separate question from eligibility is inventorship when AI tools contribute to the invention. The USPTO has been explicit that whether an invention was created with the assistance of AI is not relevant to the Section 101 eligibility analysis. But its February 2024 Inventorship Guidance requires that a natural person make a significant contribution to each claimed invention; AI systems cannot be named as inventors. For teams that lean heavily on AI tooling in R&D, getting the named inventors and the contribution record right is now part of doing AI patent work properly.

A note on capability

Lynch LLP has worked on AI-related patent prosecution across the company stages described here — from pre-seed companies filing their first provisional through scaled platform companies managing continuation trees on multiple product lines. Our case study on AI document processing — a USPTO-allowed claim set after extensive eligibility scrutiny — is one example of the work in this space. We also handle eligibility-heavy claim drafting, FTO analyses for AI products incorporating open-source components, and coordination with venture-side IP due diligence. The AI patent landscape is moving fast enough that this article will likely need an update before the end of 2027. The fundamental strategic questions — claim scope, timing, platform-versus-product weighting — are likely to remain stable. The specific examiner trends, court decisions, and USPTO guidance will keep shifting; the through-line, for now, is that examination has grown more favorable to AI while the harder questions move to the Federal Circuit.

Talk to counsel

If you're building an AI company and the patent-strategy questions in this article are live for you, a Patent Consultation is the right next step. Thirty minutes, free, with an attorney who'll discuss the specific surfaces of your technology and walk through what a defensible portfolio could look like for your company's stage and competitive position. Schedule a Patent Consultation →

Frequently asked questions

Why are AI patents harder to get than other software patents? Eligibility under 35 U.S.C. 101 is the distinctive hurdle: AI claims frequently recite an abstract idea (usually a mathematical concept or a mental process), which triggers a closer look. But "harder" doesn't mean "tightening." At the examination level, the USPTO has steadily made AI eligibility more favorable and more predictable — through the 2019 PEG, the 2024 AI SME Update, an August 2025 memo reining in over-rejection, and the precedential Ex Parte Desjardins decision upholding a machine-learning claim. The claims that fare best tie the AI to a specific technological improvement — an improvement in how the computer or model functions, not merely a better mathematical result. The residual unpredictability lives mostly at the Federal Circuit (see Recentive Analytics v. Fox, 2025). What's the difference between platform claims and product claims for AI? Platform claims target the underlying technology — model architecture, training methodology, inference infrastructure — reused across multiple products. Product claims target a specific deployed application: how the model integrates with a particular data pipeline, workflow, or output format. Platform claims are harder to draft and prosecute but cover more competitive surface. Product claims are easier to obtain but cover less. Most mature portfolios include both. What makes an AI claim eligible under Section 101? The claim must integrate any recited abstract idea into a practical application — most often by reflecting an improvement to the functioning of a computer or another technology or technical field, disclosed in the specification and reflected in the claim. Improving the abstract idea itself (a more accurate prediction or calculation) is not enough; improving how the system operates is. Drafting matters at the margin: naming a specific algorithm (e.g., backpropagation, gradient descent) in a claim can pull it into the abstract-idea analysis, whereas describing what the trained system does often does not. What timing considerations matter most for AI patent strategy? Three: the engineering roadmap (filing windows around major architecture changes), the funding cycle (Series A diligence is typically the first serious IP review investors conduct), and the public-disclosure pipeline (conference talks, blog posts, GitHub releases, and customer demonstrations all carry timing implications, and the U.S. grace period generally has no foreign equivalent). Coordinating these three timelines is one of the harder parts of AI patent work. Should AI companies use trade secrets, patents, or both? Typically both. Patents work well for architectures, training methodologies, and system designs that can be claimed as technological improvements. Trade secrets work well for specific model weights, fine-tuning data, and proprietary evaluation benchmarks that can be kept confidential operationally. The allocation depends on what is enforceable as a claim, what is discoverable from a deployed product, and what the company can credibly keep confidential. How do continuations affect AI patent strategy? Continuations let companies file additional claims off the same initial disclosure as the product evolves. For AI companies whose product surface changes frequently, continuation practice is often the primary mechanism for keeping coverage current. The first filing is rarely the whole strategy; the continuation tree built off the initial disclosure usually matters more in the long run. Can we patent an invention our team built with heavy AI assistance? Yes — using AI tools does not, by itself, affect eligibility. But at least one natural person must have made a significant contribution to each claimed invention, and AI systems cannot be named as inventors (USPTO Inventorship Guidance, February 2024). Teams that rely heavily on AI in R&D should be deliberate about documenting human contribution and naming inventors correctly. What does a Series A investor expect to see in an AI company's patent portfolio? Most Series A investors expect at least one lead-asset filing (covering the specific deployed product) and at least one platform-level filing (covering the underlying technology). The portfolio doesn't need to be deep — most Series A AI companies have between one and five active applications. What matters more is that the filings demonstrate coordinated strategy: clear scope choices, evidence that engineering and legal have been talking, and a plausible continuation roadmap for the next 12–24 months.