Imagine spending three years and $50,000 prosecuting a patent application, only to watch the resulting patent die on a motion to dismiss before the defendant has even filed an answer. No claim construction. No discovery. No expert reports. No trial. Just a judge reading your claims, comparing them to a handful of Federal Circuit precedents, and concluding that your invention is nothing more than an "abstract idea" implemented on a generic computer. Case dismissed, often with a one-line invocation of Alice.

This is not a hypothetical. It is the reality facing software and business method patent holders every day in the decade-plus since the Supreme Court's unanimous 2014 decision in Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014). The numbers are stark. In recent years the Federal Circuit has decided roughly two dozen patent appeals annually on substantive Section 101 grounds and found claims eligible in only a small handful—an invalidity rate well above ninety percent. At the Patent Trial and Appeal Board (PTAB), examiners' Section 101 rejections are affirmed the overwhelming majority of the time. Empirical scholarship on "quick decisions" in patent cases has documented that federal district courts have invalidated patents under Section 101 at rates exceeding sixty percent in most years since Alice was decided, frequently on the pleadings. (Because aggregate litigation statistics shift with each new docket and each new study, treat the precise percentages here as well-supported approximations rather than fixed figures.)

And yet, patents on software and business method innovations continue to be granted, enforced, and upheld. The path through the Alice framework is narrow, but it exists—and the practitioners who understand it are securing meaningful protection for their clients in artificial intelligence, fintech, cybersecurity, cloud computing, and dozens of other technology-driven fields. The difference between the patents that survive and the patents that die almost always comes down to how the invention is framed, how the claims are drafted, and how the prosecution is conducted. Eligibility, in other words, is far less about what you invented than about how you said it.

Before we dive in, a word for every reader who is not a patent lawyer. The term you will see most often is Section 101—shorthand for 35 U.S.C. § 101, the statute that defines what kinds of inventions can be patented at all. It is the very first gate an invention must pass through, before anyone even asks whether the invention is new (Section 102), nonobvious (Section 103), or adequately described and enabled (Section 112). A judicial exception is a category the courts have carved out of that statute by judge-made rule—laws of nature, natural phenomena, and abstract ideas—that cannot be patented no matter how clever or useful. The whole Alice drama is a fight over one of those exceptions: the abstract idea. We will define each term of art as we go, so that a judge, a seasoned patent prosecutor, and a curious founder can all follow along.

This guide provides a comprehensive map of the post-Alice landscape. We examine the two-step eligibility framework and its statutory and historical roots, trace the Federal Circuit decisions that define its contours, analyze the USPTO's evolving examination guidance (including the 2019 Patent Eligibility Guidance, the 2024 AI update, and the landmark 2025 Desjardins decision), and detail the practical strategies for claim drafting, prosecution, and portfolio management that separate successful applicants from those left holding unenforceable paper. This is the first of two companion pieces on the topic; for an alternative pass at the same terrain with additional litigation-side framing, see Patent Eligibility After Alice (part two). For a broader survey of how patents fit alongside copyright, trade secrets, and contracts in a software business, pair this guide with our overview of the legal protection of software. For companies and inventors building the next generation of software-driven products, this is the playbook for protecting what you build.

The Alice/Mayo Framework: Two Steps, One Minefield

A Short History: How We Got Here

To understand Alice, it helps to know that software was not always under siege, and that Section 101 itself is famously generous on its face. The statute makes eligible any "new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof." The Supreme Court has read that language expansively, observing that Congress intended patent protection to extend to "anything under the sun that is made by man" (Diamond v. Chakrabarty, 447 U.S. 303, 309 (1980)). The trouble has always come from the implied exceptions. As far back as Diamond v. Diehr, 450 U.S. 175 (1981), the Court confirmed both that laws of nature, natural phenomena, and abstract ideas lie outside the statute, and that a claim does not become ineligible merely because it employs a mathematical formula—there, the famous Arrhenius equation used in a process for curing rubber. Diehr upheld the claims because they applied the equation to a concrete industrial process; the formula was a means to an end, not the end itself. That distinction—between claiming a building block and claiming a specific application of it—is the seed from which the entire Alice doctrine grows.

For a stretch in the late 1990s and 2000s, the gates swung wide open. The Federal Circuit's decision in State Street Bank & Trust Co. v. Signature Financial Group, 149 F.3d 1368 (Fed. Cir. 1998), held that a method producing a "useful, concrete, and tangible result"—there, a data-processing system for managing mutual funds—was patentable, and the business-method patent boom was on. The Supreme Court began tightening the screws in Bilski v. Kappos, 561 U.S. 593 (2010), which rejected a patent on a method of hedging commodity-price risk as an unpatentable abstract idea. Bilski did two enduring things: it held that the "machine-or-transformation" test (does the claim tie the process to a particular machine, or transform an article into a different state?) is a "useful and important clue" but not the exclusive test for eligibility, and it confirmed that "fundamental economic practices long prevalent in our system of commerce" are paradigmatic abstract ideas. Then came Mayo, then Alice, and the pendulum swung hard the other way.

How the Test Works

The Alice decision did not invent the concept of patent-ineligible subject matter. What it did was formalize the two-step analytical framework first articulated in Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012)—a life-sciences case about correlating drug-metabolite levels with dosage—and apply it with devastating effect to software patents. The animating concern in both cases is preemption: the worry that a patent on a building block of human ingenuity would tie up the basic tools of scientific and technological work and "impede innovation more than it would tend to promote it" (Mayo, 566 U.S. at 71; Alice, 573 U.S. at 216). Keep preemption in mind. It is the policy engine under the hood of every Section 101 fight, and it is the most persuasive theme an applicant can sound when arguing that narrow, specific claims leave the field open for others.

The framework is deceptively simple in theory and maddeningly unpredictable in application.

Step One asks whether the claims are "directed to" a judicial exception—an abstract idea, law of nature, or natural phenomenon. If the claims focus on a patent-eligible application of such a concept rather than on the concept itself, the inquiry ends and the claims are eligible. If the claims are directed to a judicial exception, the analysis proceeds to Step Two.

Step Two asks whether the claim elements, considered both individually and as an ordered combination, contain an "inventive concept" sufficient to "transform the nature of the claim" from a patent on an abstract idea into a patent on a specific application of that idea (Alice, 573 U.S. at 217–18). The inventive concept must amount to "significantly more" than the judicial exception itself. The phrase "significantly more" is doing enormous work here, and—as we will see—nobody has ever defined it with precision.

In Alice itself, the Court applied this framework to patents claiming computer-implemented methods for mitigating settlement risk in financial transactions. The claims were directed to the abstract idea of "intermediated settlement"—a fundamental economic practice as old as banking itself—and the addition of generic computer components performing generic computer functions did not supply the necessary inventive concept. Every claim—method, system, and computer-readable medium—was invalidated, and the Court refused to let the system and media claims survive on the theory that they recited tangible hardware: a "draftsman's art" of restyling a method as a machine cannot save an otherwise ineligible idea (Alice, 573 U.S. at 226). The Court drew an explicit line back to Bilski: just as hedging is a fundamental economic practice, so is using a third-party intermediary to mitigate settlement risk (Alice, 573 U.S. at 220–21).

Why "Abstract Idea" Remains Undefined

The most consequential sentence in the Alice opinion may be this one: "we need not labor to delimit the precise contours of the 'abstract ideas' category in this case" (Alice, 573 U.S. at 221). With that single disclaimer, Justice Thomas declined to provide the definition that every patent examiner, practitioner, and judge has needed ever since.

The Court offered only examples: fundamental economic practices, certain methods of organizing human activity, mathematical relationships and formulas, and "an idea 'of itself.'" The Federal Circuit has since developed what it candidly acknowledges is a "common law approach"—deciding eligibility by comparing the claims at hand to claims already adjudicated in prior cases, rather than applying a principled, freestanding legal test. As Judge Linn observed in his concurrence in Ultramercial, Inc. v. Hulu, LLC, 772 F.3d 709, 731 (Fed. Cir. 2014), "there is no such single, succinct, usable definition or test" for abstract ideas. The court has even split internally over whether the test is administrable at all; the denial of en banc review in Athena Diagnostics, Inc. v. Mayo Collaborative Services, 927 F.3d 1333 (Fed. Cir. 2019), produced eight separate opinions, several of them open pleas for the Supreme Court or Congress to intervene.

The practical consequence is profound unpredictability. Two experienced patent attorneys can look at the same claim set and reach opposite conclusions about eligibility. A patent that survives a district court challenge may fall at the Federal Circuit. And the absence of Supreme Court guidance since Alice—the Court has denied certiorari in dozens of Section 101 cases, including some where the Solicitor General urged review—means the Federal Circuit's approach, with all its inconsistencies, remains the final word.

For a framework intended to distinguish the truly innovative from the merely abstract, this is an extraordinary amount of uncertainty. It is also the environment in which every software patent must be prosecuted, maintained, and enforced. The good news, as the next section shows, is that the chaos has a shape. The cases rhyme. And once you learn the melody, you can write claims that sing on the right side of the line.

The Federal Circuit Scorecard: Where Claims Survive and Where They Die

The Narrow Path to Eligibility

Despite the forbidding statistics, the Federal Circuit has identified categories of software inventions that satisfy the Alice framework. These decisions function as the closest thing practitioners have to a roadmap, and understanding their reasoning is essential for drafting claims that survive. Treat the cases below not as trivia but as templates: each one is a pattern an examiner and a judge will recognize, and the drafting goal is to make your claims look like these and not like the casualties that follow.

Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016). If there is a single post-Alice decision that software patent practitioners should commit to memory, it is Enfish. The court held that claims to a "self-referential" database table were patent-eligible at Step One—they never even reached Step Two—because the claims were "directed to a specific improvement to the way computers operate," not to an abstract idea (Enfish, 822 F.3d at 1336). The court drew a critical distinction: claims that "purport to improve the functioning of the computer itself" are fundamentally different from claims that merely "invoke computers as a tool" to implement an abstract idea. The self-referential table was a specific data structure that provided concrete technical advantages over conventional relational databases—faster searching, more flexible data modeling, and a smaller memory footprint. Those advantages made the claims about computer improvement, not about an abstract concept. The court put the principle memorably: "much of the advancement made in computer technology consists of improvements to software that, by their very nature, may not be defined by particular physical features but rather by logical structures and processes" (Enfish, 822 F.3d at 1339). The lesson for applicants is to frame every software invention as an improvement to how the computer operates, and to make sure the claims capture that improvement in specific technical terms.

DDR Holdings, LLC v. Hotels.com, L.P., 773 F.3d 1245 (Fed. Cir. 2014). Decided just months after Alice, DDR was the first Federal Circuit decision to uphold software claims under the new framework, and it remains a workhorse citation. The claims solved a problem "necessarily rooted in computer technology"—keeping a website's visitors on a host site rather than losing them to a third-party merchant's page when they clicked an advertisement—by generating a hybrid web page that combined the look and feel of the host with the merchant's product information. The court held the claims were not directed to a pre-internet business practice merely dressed up in computer clothing, because the asserted solution was "rooted in computer technology in order to overcome a problem specifically arising in the realm of computer networks" (DDR, 773 F.3d at 1257). DDR gave us the durable idea that a claim solving an inherently technological problem in an unconventional way can survive even when it touches commerce. It also illustrates the limit: the court was careful to say not every claim that recites the internet is eligible—only those that do not merely recite a longstanding business practice performed online.

McRO, Inc. v. Bandai Namco Games America Inc., 837 F.3d 1299 (Fed. Cir. 2016). The court found eligible claims reciting specific rules for automated lip synchronization in computer animation. The claims were not directed to the abstract idea of "using rules" but to "a patentable, technological improvement over the existing, manual 3-D animation techniques" (McRO, 837 F.3d at 1314). The specific rules—morph weight sets applied to subsequences of phonemes—were the key: they replaced subjective human judgment with a defined, automated process that produced results the old method could not. The court emphasized two things worth absorbing. First, the rules were particular, not a claim to all ways of automating the task, so they did not preempt the field. Second, the automated process was not simply the computerization of what animators had always done by hand; it used a different, rules-based technique. McRO is the case to cite when your invention encodes human know-how into a particular, non-preemptive rule set.

BASCOM Global Internet Services, Inc. v. AT&T Mobility LLC, 827 F.3d 1341 (Fed. Cir. 2016). BASCOM is the indispensable Step Two case. The claims covered a system for filtering internet content—an idea the court was willing to call abstract at Step One—but the court held the claims survived at Step Two because the ordered combination of limitations supplied an inventive concept. The patent placed the filtering tool at a specific location, a remote ISP server, and associated each network account with its own customizable filtering scheme. Individually, every piece was conventional; arranged this particular way, the claims described a "non-conventional and non-generic arrangement of known, conventional pieces" that gave users individualized filtering with the administrative ease of a central server (BASCOM, 827 F.3d at 1350). The lesson is that even when Step One goes against you, a genuinely unconventional architecture—how the known pieces are combined—can rescue the claims. BASCOM is your fallback when Enfish is unavailable.

Core Wireless Licensing v. LG Electronics, 880 F.3d 1356 (Fed. Cir. 2018). Claims to an improved user interface for computing devices survived because they recited "a specific improvement over prior systems" that produced "an improved user interface for electronic devices." The key was the specificity of the claimed interface—an application summary window displayed on a limited portion of the screen, reachable while the application is in an unlaunched state, with links to underlying data—rather than a generic recitation of displaying information on a screen. Core Wireless teaches that even a UI claim, normally vulnerable, can survive when it recites a particular structured display solving a particular problem (here, the inefficiency of navigating menus on small-screen devices).

Finjan, Inc. v. Blue Coat Systems, 879 F.3d 1299 (Fed. Cir. 2018). Claims directed to a "behavior-based" virus-scanning system employing "a new kind of file that enables a computer security system to do things it could not do before" were found patent-eligible. The invention was a security profile representing a new approach to malware detection—proactive analysis of code behavior rather than reactive comparison against known signatures—and the claims captured this technical advance with sufficient specificity. Finjan is the cybersecurity practitioner's anchor case.

These cases share a common DNA. Each involves a specific technical improvement to how a computer system functions, or a specific unconventional arrangement of components. Each describes the improvement through concrete claim limitations—particular data structures, defined algorithmic rules, specific interface designs, new analytical approaches, non-generic architectures—rather than through result-oriented language. Each is supported by a specification that explains the technical problem, describes the prior art's limitations, and identifies how the claimed invention provides a measurable advantage. And in each, the claims are narrow enough that they do not threaten to preempt an entire field.

The Recurring Patterns of Ineligibility

The far more numerous decisions finding ineligibility reveal patterns that practitioners must recognize and avoid.

The information collection/analysis/display trap. In Electric Power Group, LLC v. Alstom S.A., 830 F.3d 1350 (Fed. Cir. 2016), the court held that claims directed to "collecting information, analyzing it, and displaying certain results" are directed to an abstract idea because "information as such is an intangible." Claims that follow the pattern of "gather data, process data, display results"—without specifying a particular technical method of gathering, processing, or displaying—are almost certainly doomed. This is the most common trap in software patent claims and the easiest to fall into; Electric Power Group appears in Section 101 rejections more than perhaps any other case.

The "specification-claim disconnect." In Intellectual Ventures I LLC v. Symantec Corp., 838 F.3d 1307 (Fed. Cir. 2016), the specification described genuine improvements to email-filtering technology, but the claims failed to capture those improvements as claim limitations. The specification's account of how the invention improved computer functionality was legally irrelevant because it did not appear in the claims themselves. The lesson is unforgiving: if the technical improvement is not in the claim language, it does not count.

The "do it on a computer" trap. In Intellectual Ventures I LLC v. Capital One Bank (USA), 792 F.3d 1363 (Fed. Cir. 2015), claims to tracking financial transactions to spot budget limits, and to customizing web-page content based on user information, were both held abstract because they recited longstanding human practices implemented with generic computing. This is the direct descendant of Alice: take something people have always done, add "with a computer," and you have not added an inventive concept.

The "targeted advertising" and "data organization" categories. In Broadband iTV, Inc. v. Amazon.com, Inc., 113 F.4th 1359 (Fed. Cir. 2024), claims requiring the collection, organization, and display of viewing information for electronic program guides were held abstract, and claims using viewing history for recommendations were characterized as "targeted advertising"—a category the Federal Circuit has repeatedly found abstract. More broadly, claims directed to organizing, categorizing, or displaying information according to user preferences tend to be treated as abstract unless the claims specify a novel technical mechanism for performing those operations.

The landmark AI decision: Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. 2025). This decision—the most significant post-Alice ruling for AI patents—addressed a question of first impression: whether "claims that do no more than apply established methods of machine learning to a new data environment are patent eligible." The court held they are not. The patents claimed methods for generating optimized broadcast schedules and network maps using machine learning, but the court found they relied on "conventional" machine-learning techniques running on "generic computing equipment." The court rejected the argument that iterative training or dynamic adjustment transformed the claims, calling those features "incident to the very nature of machine learning."

Recentive draws a line every company developing AI-driven products must understand. Merely applying off-the-shelf machine learning to a new problem domain—however valuable the resulting application—does not satisfy Alice. To survive, AI patent claims must describe specific improvements to the machine-learning technology itself: novel model architectures, improved training methodologies, more efficient parameter optimization, reduced computational overhead, or enhanced generalization. "We used neural networks to solve scheduling" is abstract. "We modified the neural network architecture in the following specific way to reduce training time by forty percent while maintaining accuracy" is a technical improvement that stands a fighting chance.

For companies navigating the intersection of AI and patent rights, our article on AI-generated inventions and ownership explores the additional question of who qualifies as the "inventor" when machines contribute to the creative process, and our overview of artificial intelligence key legal issues maps the wider terrain. Eligibility and inventorship are distinct doctrines, but for AI builders they arrive at the same desk on the same morning.

USPTO Guidance: The Examiner's Lens

Federal Circuit precedent controls the law, but USPTO guidance controls the examination—and for applicants prosecuting patents, examination is where the battle is most often won or lost. A patent examiner does not perform a freewheeling read of Supreme Court opinions; the examiner applies the agency's own structured procedure. Understanding that procedure is half of practice. One caveat to hold onto throughout: agency guidance does not override the courts. The Federal Circuit said so flatly in In re Rudy, 956 F.3d 1379 (Fed. Cir. 2020), holding that where USPTO guidance "contradicts or does not fully accord with" Supreme Court or Federal Circuit precedent, the case law controls. Guidance is a powerful tool before an examiner and a near-irrelevance before a district judge. Smart drafting satisfies both.

The 2019 Revised Patent Eligibility Guidance (the "2019 PEG")

In January 2019, the USPTO issued its Revised Patent Subject Matter Eligibility Guidance—commonly called the 2019 PEG—restructuring the Alice framework into a more detailed analytical procedure designed to reduce inconsistency across thousands of examiners (84 Fed. Reg. 50 (Jan. 7, 2019)). The 2019 PEG remains the backbone of examination today and has been folded into the Manual of Patent Examining Procedure (MPEP § 2106).

The 2019 PEG reorganized the analysis into a set of clearly labeled steps. Step 1 simply asks whether the claim falls within one of the statutory categories (process, machine, manufacture, or composition of matter). Step 2A, Prong One asks whether the claim recites a judicial exception, and—crucially—it confines abstract ideas to three enumerated groupings:

  • Mathematical concepts (mathematical relationships, formulas, equations, and calculations);
  • Certain methods of organizing human activity (fundamental economic practices; commercial or legal interactions; and managing personal behavior, relationships, or interactions between people); and
  • Mental processes (concepts that can be performed in the human mind, including with pen and paper).

Step 2A, Prong Two asks whether the claim, if it recites a judicial exception, integrates that exception into a practical application that imposes a meaningful limit beyond the exception itself. This "practical application" inquiry was the 2019 PEG's most important innovation: it pulled forward, into Step One territory, the question of whether the claim does something real with the abstract idea, so that many claims could be found eligible without ever reaching the conventionality analysis. The guidance lists recognized signs of integration—improving the functioning of a computer or other technology, applying the exception with a particular machine, transforming an article to a different state, or otherwise applying the exception in a meaningful way beyond merely linking it to a technological environment. Step 2B asks the remaining claims whether they recite elements that, individually or in combination, amount to "significantly more" than the exception—the classic Alice Step Two inventive-concept question.

The 2019 PEG was widely perceived as lowering the Section 101 barrier, particularly for computer-related inventions, and empirical studies confirmed that allowance rates for AI- and software-related applications rose after its issuance. But two caveats matter. First, the guidance binds only examiners—not the courts—so an examiner's finding of eligibility under the 2019 PEG does not guarantee that the resulting patent will survive a Federal Circuit challenge (In re Rudy, again). Second, by confining abstract ideas to the three enumerated groupings, the guidance handed applicants a powerful argument: if the examiner cannot fit the claim into mathematical concepts, methods of organizing human activity, or mental processes, the claim does not recite an abstract idea at all under the agency's own rules. Force the examiner to name the grouping. Many cannot, cleanly, and the rejection wobbles.

The July 2024 AI-Specific Update

In response to the federal government's broader push on artificial intelligence policy, the USPTO issued an update in July 2024 specifically addressing AI inventions (89 Fed. Reg. 58128 (July 17, 2024)). The 2024 update introduced new worked examples (Examples 47–49) demonstrating how the Alice/Mayo framework applies to AI claims—including claims involving an artificial neural network for anomaly detection in network traffic, AI-based analysis of speech signals, and related scenarios.

The update's central message is that AI inventions are more likely to be deemed eligible when the claims articulate how the AI achieves a particular technical improvement, not merely that AI is used. A claim reciting "a machine learning model that optimizes scheduling" is result-oriented and vulnerable. A claim reciting "a neural network with a specific attention mechanism that reduces training convergence time by selectively weighting temporal features" describes a technical improvement to the model itself. The update also reinforced that a specific hardware implementation, details about training methodology, and integration into a broader technological process all support eligibility—and it cautioned examiners against over-reading the "mental process" grouping to swallow operations that no human could practically perform.

The 2025 Examiner Guidance on Mental Processes

The USPTO issued additional internal guidance in 2025 addressing several recurring problems in Section 101 examination, particularly in software-related art units. The guidance addressed the mental-process grouping head-on, reminding examiners that they must evaluate whether claimed processes can practically be performed in the human mind—not whether they are merely theoretically conceivable as mental operations. A claim to a specific computational process operating on millions of data points in real time should not be rejected as a "mental process" merely because a human could, given unlimited time and paper, in principle perform the same calculation.

The guidance also reinforced the important distinction between claims that recite a judicial exception and claims that merely involve one. A claim that uses a mathematical equation as one step of a larger technical process does not necessarily "recite" a mathematical concept in the relevant sense—it may involve the concept without being directed to it. This is Diehr's lesson, expressed in examination language: the rubber-curing process used the Arrhenius equation without being about it. Applied consistently, this distinction meaningfully reduces the number of first-action Section 101 rejections.

Ex parte Desjardins: The Game Changer

The single most important development in USPTO examination practice since the 2019 PEG is the 2025 decision in Ex parte Desjardins, issued by the USPTO's Appeals Review Panel (ARP) and designated precedential and incorporated into the MPEP. The decision vacated a PTAB affirmance of a Section 101 rejection of claims directed to training machine-learning models.

The claims at issue described a method for sequentially training a machine-learning model on multiple tasks while preserving performance on previously learned tasks—a problem known in the AI field as "catastrophic forgetting," where a neural network trained on a new task tends to overwrite what it learned on earlier ones. The ARP held that the claims integrated the judicial exception into a practical application because they improved "how the machine learning model itself operates"—specifically, by reducing storage requirements and preserving task performance across sequential training. The decision relied emphatically on Enfish, treating it as settled precedent that examiners and the PTAB must apply, and reaffirming that "software can make non-abstract improvements to computer technology just as hardware improvements can."

But the most consequential passage in Desjardins may be its closing guidance: in substance, that Sections 102, 103, and 112 are the traditional and appropriate tools for limiting patent protection to its proper scope, and that those statutory provisions—not Section 101—should be the focus of examination for claims that genuinely improve a technology. That language—from a precedential agency decision—signals a deliberate institutional pivot. The message to examiners is clear: stop using Section 101 as a first-line gatekeeping tool. If a claim recites a genuine technical improvement, test its novelty under Section 102, its nonobviousness under Section 103, and its disclosure under Section 112. Do not reject it as "abstract" merely because it involves software.

The USPTO has begun implementing Desjardins through MPEP updates instructing examiners to apply Enfish-style analysis to AI and machine-learning claims, to evaluate specifications for detailed disclosures of technological improvements, to confirm that the claim "as a whole" reflects the disclosed improvement, and to avoid dismissing limitations as "generic computer components" when they confer real technical benefits. For applicants, Desjardins supplies a powerful new argument in every Section 101 rejection response: the agency's own precedential authority says that claims directed to technical improvements in software and AI should be analyzed under traditional patentability requirements, not filtered out at the eligibility threshold. We discuss how to deploy that argument, alongside the rest of your toolkit, in our companion guide on responding to patent office actions.

The Art of Claim Drafting: What Separates the Living from the Dead

The difference between a patent that survives Alice and one that doesn't is almost never the underlying innovation. It is the way that innovation is described and claimed. Software engineers build remarkable things; the challenge is translating their work into claim language the law recognizes as directed to a technical improvement rather than an abstract idea. What follows are the principles that matter most. (A claim, for non-specialist readers, is a single numbered sentence at the end of a patent that defines the precise legal boundary of the invention—the "metes and bounds." Everything in Alice turns on what those sentences say.)

Frame Everything Around Technical Improvement

Every software patent claim should answer one question: What does this invention make the computer do better? Not "what problem does it solve for the user" or "what business outcome does it achieve," but what specific improvement does it provide to the functioning of the computer system, the efficiency of a computational process, or the capability of the underlying technology?

The specification—the lengthy written description preceding the claims—must explicitly identify the technical problem being solved, describe the limitations of prior approaches, and explain how the claimed invention delivers a concrete, measurable improvement. Terms like "faster processing," "reduced memory requirements," "improved accuracy," "enhanced security," or "lower latency" are valuable, but only when supported by a specific technical explanation of how the improvement is achieved. A specification that says "our system is faster" invites the question "faster at what, and why?" A specification that says "our system reduces query-response time by using a hierarchical index structure that eliminates redundant traversals of the search tree" answers the question before it is asked.

Critically, these improvements must be reflected in the claim language itself. The Federal Circuit has repeatedly emphasized that a specification describing impressive technical benefits cannot save claims that fail to capture those benefits as claim limitations. Intellectual Ventures v. Symantec is the cautionary tale: brilliant specification, abstract claims. Review every claim against the specification and ask: does this claim contain the technical improvement, or does it merely describe the result that the improvement achieves?

Avoid the Result-Oriented Trap

The single most common drafting mistake in software patent claims is describing what the invention achieves without specifying how it achieves it. Consider two ways to claim the same AI invention.

Hypothetical (illustrative only). Claim A: "A method comprising determining an optimal delivery schedule using a machine-learning model trained on historical delivery data." Claim B: "A method comprising training a recurrent neural network on sequential delivery-event data using a regularization term that penalizes changes to network parameters identified as important to previously learned routing tasks, the importance of each parameter computed from a Fisher information matrix, thereby reducing on-device memory required to retain prior routing competence by at least X percent."

Claim A describes a result and is almost certainly abstract under Recentive—it applies an established technique to a new data set. Claim B describes a method—a specific architectural and training mechanism—and, not coincidentally, resembles the kind of catastrophic-forgetting solution the agency upheld in Desjardins. It has a fighting chance.

This principle extends to every element of the claim. "Analyzing data" is abstract. "Applying a convolutional filter with a kernel size of 3×3 and a stride of 2 to a feature map derived from sensor input" is technical. "Displaying results" is abstract. "Rendering a summary view on a constrained display area using a priority queue that allocates screen real estate based on user-interaction frequency weighted by recency" is technical (and looks a great deal like Core Wireless). The specificity must be genuine—not technical jargon grafted onto an otherwise abstract concept, but a real description of how the invention works at the implementation level.

Mind Preemption

Because preemption is the doctrine's policy engine, drafting narrow, specific claims is not just defensive—it is affirmatively persuasive. A claim that recites a particular rule set, a particular data structure, or a particular architecture, and that plainly leaves competitors free to solve the same problem other ways, embodies exactly the value the Court worried Alice-ineligible claims threatened. McRO turned on this point. When you respond to a Section 101 rejection, do not only argue that your claim is technical; argue that it does not preempt the field, and if you can, identify the alternative approaches your specification acknowledges remain open to others.

Navigate the Mental Process Trap

Many software claims face rejection under the "mental process" grouping—the argument that if a process could theoretically be performed in the human mind, however impractically, it qualifies as a mental process and is therefore abstract. The 2024 and 2025 USPTO guidance pushed back on this, but examiners still invoke the grouping frequently.

The most effective response is to demonstrate practical impossibility. A claim that involves processing millions of data points in real time, iterating over complex neural-network architectures, or computing multi-dimensional optimizations cannot "practically" be performed in the human mind, regardless of theoretical conceivability. The specification should make this explicit: describe the scale of the data, the computational complexity, and the time constraints that necessitate computational implementation. The argument is not merely that computers are faster—it is that the claimed process is fundamentally incompatible with mental execution because of its technical requirements.

Special Considerations for AI and Machine Learning Claims

After Recentive, the central challenge for AI patents is distinguishing genuine technological advances from the mere application of known techniques to new data domains. The court's holding is clear: claims that "do no more than apply established methods of machine learning to a new data environment" are not patent-eligible. This means claims must articulate what is new about the AI technology itself, not just what is new about the problem domain to which it is applied.

The strongest AI patent claims focus on one or more of the following: improvements to the model architecture (a novel attention mechanism, a new type of layer, an innovative connectivity pattern); improvements to training methodology (a specific loss function that accelerates convergence, a curriculum-learning strategy that improves generalization, a regularization technique that prevents catastrophic forgetting); improvements to computational efficiency (reduced memory footprint, faster inference time, lower power consumption); or improvements to data processing (a novel feature-extraction technique, a preprocessing pipeline that improves model accuracy on noisy data).

Claims that merely recite "a neural network trained on [domain-specific data] to predict [domain-specific output]" will not survive. A worked example sharpens the line.

Hypothetical (illustrative only). Suppose Acme Robotics develops a warehouse-picking system. A claim to "using machine learning to optimize the order in which a robot retrieves items" is the Recentive death trap—an established technique pointed at a new domain. But a claim to "a reinforcement-learning policy network whose reward function incorporates a collision-risk penalty computed from a depth-sensor occupancy grid updated at 60 Hz, the network's action space constrained to a precomputed motion-primitive library to bound inference latency below 8 milliseconds" describes specific architectural and efficiency improvements. The first claim is about what Acme did; the second is about how. Only the second has a real path through Alice.

Prosecution Strategy: Winning the Examination Battle

Deconstructing the Examiner's Rejection

A Section 101 rejection is only as strong as the examiner's characterization of the abstract idea. Many rejections rest on overgeneralization—describing the claims at a high level of abstraction that strips away the very technical features that distinguish the invention from the prior art. The Federal Circuit has warned repeatedly against "describing the claims at such a high level of abstraction and untethered from the language of the claims" that virtually any invention could be characterized as abstract (Enfish, 822 F.3d at 1337).

Begin every response by holding the examiner to the prima facie burden. The MPEP requires the examiner to set out a prima facie case of ineligibility and to articulate the rejection's rationale with enough specificity that the applicant can respond. If the office action merely asserts that the claims are "directed to an abstract idea" without identifying which of the three enumerated groupings applies, or recites a boilerplate "generic computer" conclusion, the rejection is procedurally deficient—traverse it on that ground first.

Then scrutinize the examiner's Step 2A characterization. Has the examiner accurately described what the claims recite, or has the examiner abstracted away the specific technical elements? If the claims recite a specific data structure, algorithm, or computational process, and the examiner has characterized the claims as merely "organizing data" or "performing calculations," the response should demonstrate—limitation by limitation—how the examiner's characterization fails to account for the claim's technical specificity. Remember the 2019 PEG's grouping argument: if the alleged abstract idea does not fit cleanly within mathematical concepts, methods of organizing human activity, or mental processes, the agency's own rules say there is no abstract idea to begin with. And at Prong Two, marshal the practical-application signs the guidance itself lists—improvement to computer functionality, use of a particular machine, transformation—mapping each to a specific claim limitation.

At Step 2B, examine whether the examiner has provided adequate factual support for the finding that the additional elements are "well-understood, routine, and conventional." This is where Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018), changed the game. The Federal Circuit held that whether claim elements are well-understood, routine, and conventional is a question of fact—not something an examiner or court may simply assert—and a companion decision, Aatrix Software, Inc. v. Green Shades Software, Inc., 882 F.3d 1121 (Fed. Cir. 2018), confirmed that factual disputes on this point can defeat a motion to dismiss in litigation. The USPTO promptly issued the "Berkheimer memorandum" requiring examiners to support any conventionality finding with one of four specified kinds of evidence: an express statement in the specification or an applicant admission; a citation to a court decision noting the element's conventionality; a citation to a publication demonstrating it; or a statement of official notice. A conclusory "these are generic computer components" no longer suffices, and pointing that out is one of the most reliable ways to traverse a Step 2B rejection.

Leverage the USPTO's Own Guidance

The USPTO has published a substantial library of subject-matter-eligibility examples (now numbering roughly four dozen) demonstrating how the Alice/Mayo framework applies to various inventions. These examples, while not legally binding on courts, carry significant weight before examiners—many of whom were trained on them—and provide ready-made templates for successful arguments.

For any Section 101 response, identify the example most closely analogous to your claims that was found eligible. Walk through the example's reasoning and demonstrate that the same logic applies to your claims. Conversely, if the examiner has cited an ineligible example, distinguish it on technical grounds: explain what makes your claims technically different and why those differences place your claims on the eligible side of the line. After Desjardins, AI and machine-learning applicants have an additional weapon—a precedential agency decision holding that claims directed to improvements in how machine-learning models operate are patent-eligible under Enfish. Cite Desjardins in every AI-related Section 101 response.

Examiner Interviews: The Underutilized Tool

Patent prosecution interviews remain one of the most underutilized tools for overcoming Section 101 rejections. An interview lets the applicant understand the examiner's specific concerns, explore what claim amendments might resolve them, gauge whether the examiner is genuinely uncertain or simply entrenched, and build rapport that may influence future office actions. The conversation during an interview rarely creates problematic prosecution history, making it a relatively low-risk opportunity to gather intelligence and advance the case. For complex software and AI applications, where the technical merits may not be obvious from the claims alone, an interview that walks the examiner through the specification's disclosure of the technical improvement can be dispositive. (For the full menu of prosecution mechanics—after-final practice, RCEs, and PTAB appeals—see our guide to responding to patent office actions.)

Declaration Evidence

Applicants may submit declarations under 37 C.F.R. § 1.132 providing objective evidence on the technical improvements the claimed invention achieves, how the invention provides "significantly more" than the judicial exception, why the claimed processes cannot practically be performed mentally, and how a person of ordinary skill would understand the specification's disclosure of the technological improvement. Because Berkheimer made conventionality a fact question, a well-supported declaration can directly rebut a Step 2B finding. Declarations from inventors or independent technical experts are especially valuable when the innovation operates in a specialized field where the examiner may lack deep domain expertise—a common situation with AI, cryptography, and other advanced software technologies. Take care, though, that a declaration does not inadvertently admit that surrounding elements are conventional; that admission becomes Berkheimer evidence against the patent later.

The Appeal Calculus

If prosecution before the examiner is exhausted, an appeal to the PTAB may be warranted—though the statistics counsel realistic expectations, since Section 101 rejections are affirmed at high rates. Applicants should keep continuation applications pending to preserve the ability to update claims as the law evolves; claims rejected today may be allowable under tomorrow's guidance or precedent. It is also worth researching the tendencies of potential PTAB panels, which vary in their approach to Section 101. For AI and machine-learning claims, Desjardins now gives appellants strong arguments that panels must apply Enfish-style analysis to claims directed to technological improvements. For the closely related challenge of beating prior-art rejections that often accompany a Section 101 fight, see our deep dive on overcoming obviousness rejections under Section 103.

Portfolio Strategy: Audit, Protect, Layer

Auditing Existing Patents for Eligibility Vulnerability

Companies with existing software or business-method patent portfolios should conduct eligibility audits to assess vulnerability to Section 101 challenges before they arise—whether in litigation, licensing negotiations, or due diligence. An audit should identify high-risk patents whose claims focus on organizing information, implementing fundamental business practices, or bolting generic computer components onto an otherwise abstract process. It should assess claim scope—broader claims face greater eligibility risk than narrower claims tied to specific implementations. And it should comb the prosecution history for damaging admissions: statements in office-action responses that characterized claim elements as "conventional," "generic," or "well-known" will be quoted back at the patent owner in litigation, often with devastating effect under Berkheimer.

The audit also matters because Section 101 is the cheapest, earliest weapon an accused infringer has. Eligibility is frequently resolved on a motion to dismiss or for judgment on the pleadings, before discovery costs accrue. A patent owner who understands this will think hard about which patents to assert and against whom. For the defensive playbook from the other side of the "v.," and how an early eligibility attack fits into a broader litigation strategy, see our overview of patent infringement litigation.

For important patents that exhibit eligibility vulnerabilities, consider reissue (broadening reissue is available only within two years of issuance) or continuation strategies to file new claims that incorporate Enfish- and BASCOM-style technical-improvement framing. Supplemental examination directed at Section 101 can be used to address substantial new questions of patentability before an adversary raises them. These tools are most valuable when deployed proactively, before litigation forces the issue under adversarial conditions.

Layered IP Protection: Patents Are Not the Only Tool

The post-Alice landscape has made one thing abundantly clear: patents alone are not sufficient to protect software innovations. A comprehensive IP strategy must layer multiple forms of protection, each covering ground the others cannot. Our overview of the legal protection of software treats this in depth; here is the short version.

Trade secrets protect the underlying algorithms, training data, model weights, and implementation details that make software valuable—and they do so without the public disclosure that a patent requires. For AI companies whose competitive advantage lies in proprietary training methodologies or curated datasets, trade-secret protection is often more valuable, in practical terms, than a patent. The catch is that you cannot patent and keep secret the same thing: a patent application publishes (generally eighteen months after filing), so the strategic choice between patenting and secreting must be made deliberately, claim by claim. Building the program correctly—reasonable secrecy measures, access controls, NDAs—is its own discipline; see building a trade secret protection program from scratch and trade secrets in the age of remote work and cloud computing.

Copyright protects the specific code that implements the invention—the literal source and object code, and certain non-literal elements such as a program's structure, sequence, and organization. Copyright does not protect functionality (that is the patent's job), but it does prevent competitors from copying the particular expression of a solution, and it arises automatically on fixation. For software specifically, our guide to copyright registration of computer programs explains how to register without surrendering trade-secret material in the deposit.

Contractual protections—software license agreements, SaaS terms, and confidentiality provisions—impose binding restrictions on how software is used, accessed, and disclosed. These layers operate independently of IP statutes and can restrict conduct that IP law alone might permit. For many software companies distributing through the cloud, the license agreement is the most practically significant form of protection; our piece on drafting software license agreements covers the key terms.

Design patents protect novel, non-obvious, ornamental aspects of user interfaces—icons, screen displays, animated transitions, and other GUI elements. Design patents do not protect functionality, but they can stop competitors from copying the visual design of an interface, and they are both relatively inexpensive to obtain and largely immune to Alice challenges. For an app company, a thin layer of design patents over the most distinctive screens can be a quiet but potent deterrent; we map the full menu in our guide to protecting your mobile app.

The following table summarizes how the layers compare:

Protection Type What It Covers Alice Vulnerability Duration Public Disclosure Required
Utility patent Functional inventive concepts High—must survive the two-step framework 20 years from filing Yes (application publishes)
Trade secret Confidential algorithms, data, processes None—no patent examination Indefinite (while secret) No—secrecy is essential
Copyright Specific code expression Low—protects expression, not function Life + 70 years, or 95/120 years for works made for hire Partial (deposit for registration)
License/contract Whatever the parties agree to restrict None—contractual, not statutory Contract term No
Design patent Ornamental UI/UX elements Very low—protects appearance 15 years from grant Yes

Strengthening New Filings

For new applications, eligibility should inform every stage of the process, beginning with the inventor interview. Do not accept surface-level descriptions of what the invention does. Elicit detailed technical information about how it works: the specific data structures, the algorithmic steps, the computational architecture, the measurable performance improvements. The more technical detail captured at the inventor interview, the richer the specification and the more defensible the claims. Our guide on how to prepare an invention disclosure for your patent attorney is a useful starting point to put in front of your engineers.

Draft specifications that read like technical white papers, not marketing brochures. Describe the technical problem in precise terms. Identify the limitations of prior approaches with specificity. Explain how the claimed invention overcomes those limitations through a described technical mechanism, and—because preemption matters—acknowledge that other approaches to the same problem exist and remain available. Include performance benchmarks, comparative data, and concrete examples wherever possible. This detail serves a dual purpose: it supports the eligibility argument that the claims are directed to a technical improvement, and it satisfies the enablement and written-description requirements of Section 112—the requirements that, per Desjardins, should be the primary focus of examination.

Finally, draft independent claims at varying levels of specificity. Broader claims capture competitive space and provide the strongest licensing leverage, but they face greater eligibility risk. Narrower claims tied to specific implementations provide fallback positions if broader claims face rejection or invalidation. A layered claim strategy—several independent claims at different abstraction levels—maximizes the probability that at least some claims survive both prosecution and litigation. Common questions about this and other basics are collected in our patent FAQs, and the fundamentals are covered in utility patent basics.

The Legislative and Judicial Horizon

The Patent Eligibility Restoration Act (PERA)

The most significant potential reform is the Patent Eligibility Restoration Act (PERA), introduced by Senators Thom Tillis and Chris Coons in 2023 and reintroduced in subsequent Congresses. PERA would fundamentally reshape Section 101 by eliminating all judicially created exceptions to patent eligibility and replacing them with a specific statutory list of exclusions. The enumerated exclusions in the bill's various versions include a mathematical formula apart from a useful invention; an unmodified human gene as it exists in the body; an unmodified natural material; certain purely mental processes; and a process that is "substantially economic, financial, business, social, cultural, or artistic" unless it cannot be performed without the use of a machine or manufacture. PERA would also direct courts to consider eligibility separately from novelty, obviousness, and enablement, and would remove the word "new" from Section 101 to prevent conflating eligibility with prior-art analysis—echoing the very point Desjardins made administratively and the point the Court itself once stressed in Diehr.

Supporters argue that PERA would restore U.S. competitiveness in AI, biotechnology, diagnostics, and other emerging technologies where current eligibility standards leave innovations unpatentable in the United States despite being protectable in Europe, China, Japan, and other major jurisdictions. The European Patent Office, for example, routinely grants patents on AI inventions that face uncertain prospects in the United States under Recentive—an asymmetry that places U.S. innovators at a structural disadvantage and complicates global filing strategy, a theme we take up in global patent litigation strategies. Critics counter that PERA would resurrect the patent-thicket problems of the State Street era and re-open the door to broad business-method and diagnostic-method patents.

PERA has not yet advanced to passage, but bipartisan sponsorship and sustained industry pressure keep legislative reform plausible. Companies should monitor developments and maintain prosecution flexibility—continuation and provisional filings preserve the ability to adapt claims to a reformed framework if and when it arrives.

Supreme Court Silence

Since Alice, the Supreme Court has declined to hear any Section 101 case on the merits, despite dozens of certiorari petitions—including several where the Solicitor General urged the Court to provide "much-needed clarification." The denials in Tropp v. Travel Sentry (2023), CareDx, Inc. v. Natera, Inc. (2023), and a string of software and AI petitions since have left the Federal Circuit's approach effectively unreviewable—and the AI community without definitive guidance on when machine-learning innovations cross the line from "abstract" to "technical improvement." The court's own fractured Athena denial shows that even the judges who must apply the test are divided about whether it can be applied coherently.

Until the Court acts or Congress passes PERA, the Federal Circuit's decisions—and their high invalidity rate—remain the law. Practitioners must plan accordingly, treating Supreme Court intervention as a possibility to monitor rather than a probability to rely upon.

Emerging Frontiers: AI, Blockchain, and Quantum Computing

The Alice framework was developed with financial-services software in mind, but it now governs eligibility for technologies its authors never contemplated. AI patent filings have surged—the USPTO's own research documents a sharp, sustained increase in AI-related patenting over the past decade (USPTO, Inventing AI: Tracing the Diffusion of Artificial Intelligence with U.S. Patents)—and the interplay of Alice, Recentive, and Desjardins will shape how these innovations are protected for years to come.

Blockchain and distributed-ledger technology face the same fork in the road. A claim directed to "recording transactions on a distributed ledger" risks being treated as an abstract method of organizing economic activity, while a claim describing a specific cryptographic mechanism, a novel consensus protocol, or a particular data-structure innovation may fare much better—the DDR/BASCOM lesson applied to a new substrate. Quantum computing raises a genuinely hard question that the courts have barely touched: are algorithms designed for quantum hardware "abstract" because they are, at bottom, mathematical, or "technical" because they exploit quantum-mechanical phenomena that classical computers cannot replicate? Diehr suggests an answer—a mathematical algorithm tied to a specific physical implementation should be eligible—but expect that question to land on the Federal Circuit's docket within this decade.

For companies developing frontier technologies, the strategic imperative is to file early, file specifically, and maintain prosecution flexibility. Keep continuations pending so claims can be updated as the law evolves. Invest in specifications that capture the technical improvement with granularity. And recognize that the strongest portfolios combine utility patents with trade secrets, copyrights, and contracts to create a defense-in-depth that does not hinge on any single doctrine. For more on the contested IP questions raised by generative systems, see our article on copyright infringement claims against generative AI.

Practical Checklist: Navigating Section 101

Specification drafting:

  • Identify the specific technical problem the invention solves.
  • Describe the limitations of prior-art approaches in concrete terms.
  • Explain how the claimed invention overcomes those limitations through a specific technical mechanism.
  • Acknowledge alternative approaches that remain open to others, to defeat preemption arguments.
  • Include performance benchmarks, comparative data, or measurable improvements.
  • Never characterize your own technical elements as "conventional," "generic," or "well-known."

Claim drafting:

  • Frame claims around a technical improvement to computer functionality or another technology.
  • Include the "how," not just the "what"—specific algorithmic steps, data structures, or arrangements.
  • Ensure every technical improvement described in the specification appears as a claim limitation.
  • Draft claims at multiple specificity levels (broad, medium, narrow) for layered protection.
  • For AI claims, focus on improvements to the model itself, not the application of generic models to new domains.

Prosecution:

  • Hold the examiner to the prima facie burden; traverse boilerplate or grouping-less rejections.
  • Scrutinize the examiner's abstract-idea characterization—challenge overgeneralizations and the grouping fit.
  • Cite analogous eligible USPTO examples; distinguish ineligible ones.
  • Invoke Enfish, BASCOM, and Desjardins for software/AI claims directed to technical improvements.
  • Demand Berkheimer-compliant factual support for any "conventional" finding at Step 2B.
  • Request examiner interviews to understand concerns and explore amendments.
  • Consider 37 C.F.R. § 1.132 declarations for technical evidence of improvement.
  • Keep continuations pending to preserve claim flexibility as the law evolves.

Portfolio management:

  • Conduct eligibility audits of existing patents to identify high-risk claims.
  • Review prosecution histories for damaging admissions of conventionality.
  • Consider reissue or continuation strategies for key patents with eligibility vulnerabilities.
  • Layer patent protection with trade secrets, copyrights, design patents, and contracts.
  • Monitor PERA and other legislative developments; maintain filing flexibility.

Frequently Asked Questions

Can software still be patented in the United States after Alice? Yes. Alice did not categorically bar software patents; it barred patents on abstract ideas merely implemented on a generic computer. Software that improves the functioning of a computer or solves a technological problem in an unconventional way remains patentable, as Enfish, DDR Holdings, McRO, BASCOM, and Finjan all confirm. The key is to claim a specific technical improvement, described in concrete terms, rather than a result achieved on a generic machine.

What is the difference between Step One and Step Two of the Alice test? Step One asks whether the claim is directed to a judicial exception such as an abstract idea. If it is not, the claim is eligible and the analysis ends. If it is, Step Two asks whether the claim nonetheless contains an "inventive concept"—elements that, individually or as an ordered combination, add "significantly more" than the abstract idea itself. Enfish is the model for winning at Step One; BASCOM is the model for winning at Step Two through an unconventional arrangement of known components.

What is preemption, and why does it keep coming up? Preemption is the policy worry that drives the whole doctrine: a patent that monopolizes a basic building block—an abstract idea, a law of nature—would block more innovation than it promotes. The narrower and more specific your claim, the weaker any preemption objection, which is why McRO's particular rule set survived. Arguing that your claim leaves the field open to alternatives is one of the most persuasive moves available.

Why does Berkheimer matter so much in prosecution? Berkheimer v. HP Inc. held that whether a claim element is "well-understood, routine, and conventional" is a question of fact, not something the USPTO or a court may simply assert. As a result, an examiner who labels your claim elements "generic" must back it up with specified evidence (a specification admission, a court decision, a publication, or official notice). That requirement gives applicants a powerful, evidence-based way to traverse Step Two rejections, and its companion case Aatrix shows it can defeat early invalidity motions in litigation, too.

Are AI and machine-learning inventions patentable? They can be, but the bar is exacting after Recentive Analytics v. Fox Corp. Simply applying off-the-shelf machine learning to a new data set is not eligible. The claims must describe a genuine technical improvement to the AI itself—a novel architecture, a new training method, a measurable efficiency gain. The precedential Ex parte Desjardins decision shows the kind of model-level improvement (solving "catastrophic forgetting" during sequential training) that the USPTO will treat as eligible.

What is the USPTO 2019 PEG, and is it binding on courts? The 2019 Patent Eligibility Guidance is the USPTO's internal procedure for applying Alice, notable for confining "abstract ideas" to three groupings (mathematical concepts, methods of organizing human activity, and mental processes) and for its "practical application" inquiry. It binds examiners and is reflected in the MPEP, but it does not bind courts—the Federal Circuit said so in In re Rudy. A patent allowed under the 2019 PEG can still be invalidated under Section 101 in litigation, which is why claim drafting must satisfy both the agency and Federal Circuit precedent.

Should I rely on a patent or a trade secret to protect my algorithm? It depends on whether the advantage can be detected in a shipped product and how long you expect it to last. If a competitor could reverse-engineer or independently discover the technique, a patent's 20-year exclusivity may be worth the disclosure. If the advantage lives in proprietary training data, model weights, or a server-side process that never leaves your control, a trade secret can last indefinitely and avoids the Alice gauntlet entirely. Many companies do both—patenting the detectable, claimable improvements and keeping the rest secret. See our guides to trade secrets in remote and cloud environments and the legal protection of software.

Will PERA fix the Alice problem? PERA would abolish the judge-made eligibility exceptions and replace them with a narrower statutory list, which would make far more software and diagnostic inventions eligible. But it has not been enacted, its final text could change, and the Supreme Court could read any new statute narrowly. Treat PERA as a reason to preserve prosecution flexibility, not as present law.

Conclusion: Key Takeaways

More than a decade after Alice, the landscape of software and business-method patent eligibility remains challenging—but it is not hopeless. The Federal Circuit's restrictive approach and high invalidity rates demand meticulous attention to claim drafting, prosecution strategy, and portfolio management. Yet the Enfish–DDR–McRO–BASCOM line of cases demonstrates that software improvements are patent-eligible. The Berkheimer decision turned the conventionality inquiry into a factual question that applicants can win. The precedential Desjardins decision signals a meaningful shift in USPTO examination philosophy back toward Sections 102, 103, and 112. And the possibility of legislative reform through PERA, while uncertain, represents a potential rewriting of the rules entirely.

The key to success, today as in every era of patent law, is precision. Eligibility analysis turns on how an invention achieves its results, not merely what results it achieves. Claims directed to specific technical improvements—described with detailed specificity, supported by rich specifications, prosecuted with strategic sophistication, and narrow enough to escape preemption—can navigate the Alice framework. Claims that describe results without methods, that invoke computers as tools rather than as the subject of improvement, and that characterize their own elements as conventional will not. For AI and machine-learning innovations the challenge is sharper still: Recentive demands that applicants distinguish genuine advances in AI technology from the application of known techniques to new domains, and Desjardins shows how to do it. The companies and inventors who internalize these distinctions—and translate them into the language of patent claims—will be the ones who secure the protection their innovations deserve.

Related Articles

This article is for general informational purposes only and does not constitute legal advice. Patent eligibility analysis requires case-specific evaluation of the claims, specification, and prosecution history, and the law in this area is unsettled and fast-moving. Consult qualified patent counsel regarding your specific circumstances.