Drafting AI patents that survive step 2A of the USPTO's subject matter eligibility analysis has changed materially since mid-2025. Under Director Squires, the USPTO issued memoranda reminding examiners that they must provide a sufficient factual basis before rejecting claims under 35 U.S.C. § 101, and the agency signaled it intends to raise the bar for § 101 rejections in AI and machine learning cases. That shift does not make AI claims automatically eligible — it makes well-drafted claims more likely to survive, and poorly drafted ones still die at step 2A or fall back into Alice/Mayo territory at step 2B. This guide gives you the drafting playbook: what examiners now must show, how to structure claims that recite a patent-eligible improvement, which claim formats work, what to avoid, and when to act given the current examination climate.
The Direct Answer: What Step 2A Requires for AI Claims in 2026
Also worth reading: What is the USPTO SMED declaration guidance for AI patents and how does it affect § 101 eligibility? · What is the definitive guide to AI patent eligibility case law and USPTO policy in 2026? · How do you navigate AI patent eligibility strategies across major jurisdictions in 2026?
Step 2A of the USPTO's eligibility analysis (per MPEP 2106) has two prongs. Prong One asks whether the claim recites a judicial exception — an abstract idea, law of nature, or natural phenomenon. Most AI claims are grouped under the "mental processes" or "mathematical concepts" groupings of abstract ideas, sometimes under "certain methods of organizing human activity." Prong Two asks whether the claim recites "additional elements" that integrate the exception into a practical application. If Prong Two is satisfied, the claim is eligible and the examiner never reaches step 2B; if not, the examiner evaluates whether the additional elements amount to significantly more.
The 2025-2026 guidance changes the burden allocation, not the legal test. Examiners are now reminded that a rejection under § 101 requires evidentiary support — a rationale grounded in the specification, claim language, or knowledge generally available to skilled artisans — rather than conclusory assertions that an algorithm is "merely math." In practice, this means claims that clearly tie the AI model's output to a concrete technical result (a control action, a measured physical change, a specific hardware configuration) face fewer boilerplate rejections. But the guidance also cuts both ways: examiners who do reject are expected to articulate a more rigorous basis, so vague claims give them less room to sustain the rejection on appeal while giving applicants less to rebut.
The core drafting principle remains unchanged from the case law underlying the guidance: recite what the invention does to technology, not what a human could do with a pen and paper. An AI model that classifies images is vulnerable; an AI model whose classification output directly adjusts a manufacturing parameter within stated tolerances is far stronger. Your job at drafting time is to make that integration explicit, measurable, and technically specific in the claim itself.
Why the 2025-2026 Guidance Shifted the Examiner Burden
Throughout 2024, practitioners reported high rates of § 101 rejections in TC 2100 (computer architecture, software, and AI art units), often with one-line rationales citing abstract idea groupings without mapping claim limitations to any grouping key. Commentaries published through Holland & Knight, JD Supra, ArentFox Schiff, Dykema, and The National Law Review during 2025 documented a consistent pattern: rejections that failed to identify which specific limitation was a judicial exception, or that treated every neural network as inherently a mathematical concept regardless of claimed context.
Director Squires' memoranda responded to this by reminding examiners of existing requirements in MPEP 2106 and the 2019 Revised Patent Subject Matter Eligibility Guidance: identify the exception, map it to specific claim limitations, and explain why the remaining elements fail to integrate it into a practical application. The memoranda did not create new law — they enforced procedural discipline. For applicants, the practical consequence is that a well-built specification now has real leverage. If the examiner's rejection lacks the required factual basis, a supervisor-level interview or a pre-appeal brief can succeed where the same argument would have been ignored two years earlier.
Do not over-read the shift, though. The Federal Circuit's § 101 jurisprudence has not become friendlier to software, and district court outcomes remain unpredictable. The USPTO guidance governs prosecution, not litigation. A claim drafted only to satisfy a friendly examiner may still be invalidated years later if it reads on nothing but an abstract workflow. Draft for both audiences: the examiner reviewing today and the Article III judge reviewing in 2031.
Practical Drafting Step 1: Anchor Claims in Technical Improvement Language
The single highest-value drafting move is stating the technical problem and technical solution in the claim body itself. Compare these two structures. Weak: "A method comprising receiving data, applying a machine learning model to the data, and outputting a prediction." Strong: "A method for reducing inference latency in a convolutional neural network deployed on resource-constrained edge hardware, comprising pruning filter channels based on activation sparsity thresholds computed per layer, wherein the pruned network maintains classification accuracy within 0.5% of the unpruned baseline on the validation set."
The strong version does several things at once. It names the technical field (edge inference), identifies the measurable engineering objective (latency reduction with bounded accuracy loss), specifies the mechanism (sparsity-based channel pruning), and includes quantitative constraints (0.5% threshold) that make the claim concrete rather than mental-process-shaped. Each of those elements gives the examiner something to map away from the abstract idea groupings in Prong One and toward a practical application in Prong Two.
Your specification should support this structure with figures showing signal flow through actual hardware components, performance tables comparing the invention against prior-art baselines, and at least one end-to-end example with real numbers. When the examiner writes a rejection, your response can quote the spec's own technical metrics back as evidence of integration — exactly the kind of factual record the new guidance says rejections must engage with. Applications drafted without these anchors force attorneys to argue eligibility from thin claim language, which rarely works even under a friendlier examination climate.
Practical Drafting Step 2: Choose Claim Structures That Survive Prong Two
Not all AI claim architectures perform equally at step 2A. The table below compares the four dominant formats practitioners used across 2025 filings and their observed outcomes:
| Feature | System/hardware claim | Method tied to physical action | Training-method claim | Data-structure/CRM claim |
|---|---|---|---|---|
| Typical Prong One outcome | Often no abstract idea found if hardware-specific | Exception found (mental process/math) | Exception found (mathematical concept) | Frequently rejected as abstract |
| Prong Two strength | High when components are specific | Moderate to high with concrete actuator/measurement steps | Low unless training improves the model's own technical operation | Low unless structure is functionally tied to hardware |
| Litigation durability | Strongest | Moderate | Weakest | Weakest |
| Best use case | Edge devices, medical imaging pipelines, robotics | Industrial process control, diagnostics | Novel architectures with provable efficiency gains | Avoid as lead claim; keep as fallback |
File multiple independent claims across categories in the original application. Adding new claim categories after a final rejection invites § 112 or new-matter problems and burns budget. A typical robust filing includes two to three independent claims spanning system, method, and non-transitory computer-readable medium formats, each independently satisfying Prong Two rather than relying on dependent-claim fallbacks alone.
Specification Strategy: Building the Factual Record Before the First Office Action
Under the clarified expectations, the specification is your primary defensive weapon. Examiners must ground rejections in evidence, and the best counter-evidence is your own disclosure. Include a dedicated "Technical Improvements" section early in the description that enumerates, with numbers, what the invention achieves: percentage reductions in compute cost, latency improvements in milliseconds, accuracy deltas against named baselines, memory savings in megabytes. Vague statements like "improves efficiency" carry little weight; "reduces GPU memory consumption by 38% relative to dense baselines at iso-accuracy" gives your attorney a quotable fact.
Describe the invention's operation at the level of data flow between named components. Figures should show tensors moving between modules mapped to hardware blocks, not just flowchart boxes labeled "process data." Where the invention involves a novel loss function, regularization scheme, or architecture modification, explain why the mathematics produces the measured improvement — this supports the argument that the claim is directed to a specific technological implementation rather than the math itself.
Also draft against the future. Assume a litigator will need to prove infringement and a judge will assess eligibility from this document. Define every technical term on first use, avoid trademarked product names, and include alternative embodiments covering plausible design-arounds. The modest extra drafting time — typically adding five to ten hours of attorney work — pays for itself many times over in reduced office action cycles and stronger litigation posture.
Common Mistakes That Still Trigger Step 2A Rejections
The most frequent error remains drafting results-oriented claims that could be performed mentally or with generic software. A claim to "determining a risk score using a trained model and displaying it" describes a mental process plus generic presentation, and no amount of examiner guidance will save it. Every functional limitation needs an implied structural or algorithmic home in the claim.
Second, applicants often bury the technical improvement in the specification but omit it from the claims. Examiners evaluate claims, not intentions. If your pruning technique reduces latency by 40%, that fact belongs in at least one independent claim, not only in Figure 7's description. Third, over-reliance on "wherein" clauses tacked onto otherwise abstract claims fails because courts and examiners treat bolt-on limitations as insignificant extra-solution activity. Integration must be organic: the AI computation should be the thing producing the technical effect, not running alongside it.
Fourth, some applicants overcorrected after the 2025 guidance and filed claims loaded with hardware jargon ("configured processor," "memory operatively coupled") around substantively abstract methods. Examiners and courts see through this; the Federal Circuit has repeatedly held that generic computer implementation is not significantly more. Specificity means describing particular algorithms, data structures, signal timings, and component interactions — not sprinkling the word "processor" into a business-method claim. Finally, do not skip the pre-filing prior art search in adjacent art units; a surprising number of AI § 101 disputes collapse once a § 102 rejection reveals the claimed improvement was already known, converting an eligibility fight into a novelty fight you cannot win.
Prosecution Tactics: Interviews, Responses, and When to Push Back
When a § 101 rejection arrives despite good drafting, respond in layers. First, demand the required rigor: point out where the rejection fails to map limitations to a specific grouping key or lacks factual support, citing the examiner's own training obligations under the recent memoranda. Second, argue Prong Two on the merits using the specification's technical metrics as evidence of integration. Third, propose an examiner interview with a concrete amended claim set ready — interviews resolve a large share of TC 2100 eligibility disputes faster than written rounds, and supervisors now review rejections lacking adequate basis more actively than before.
Budget realistically. A first-action § 101 rejection with a strong response package typically costs $3,000–$8,000 in attorney fees depending on complexity, versus $10,000–$25,000+ if the dispute escalates to an RCE cycle, appeal brief, or PTAB proceeding. Given the clarified guidance, investing early in a thorough response is usually cheaper than appealing. Track pendency too: TC 2100 first-action pendency has run roughly 20–30 months in recent years, so filing quality matters more than speed — a clean first application that avoids two RCE cycles saves 12–18 months of total pendency.
Timing consideration: file sooner rather than later if your invention is ready. The current examination climate favors applicants, but guidance can shift with leadership changes, and prior art accumulates daily. An application filed in Q3 2026 with a technically anchored claim set positions you ahead of competitors who waited for further clarification that may never come in the form they hope.
Cost-Benefit Reality Check: Is the Current Climate Actually Better?
Honesty requires nuance. The 2025-2026 memoranda improve prosecution odds for well-drafted applications, but they do not change statutory law, and they can be revised by a successor director. Practitioners should treat the current environment as a window, not a guarantee. Allowance rates in AI-heavy art units have historically lagged other TC 2100 areas, and even favorable examiner treatment at the USPTO provides zero protection in district court, where post-grant § 101 challenges remain common and expensive — inter partes review and district court invalidation defenses routinely add six-figure costs.
For startups, the rational strategy is portfolio triage: spend full drafting effort on the two or three inventions closest to revenue, use provisional applications (filing fee $65–$130 for micro/small entities as of 2026, plus $2,000–$6,000 in typical attorney fees) to preserve dates on speculative ideas, and convert provisionals within twelve months only where market signals justify it. For established companies with licensing programs, breadth matters more than allowance speed, so invest in the multi-category claim sets described above and accept longer prosecution timelines. Either way, the marginal dollar is best spent on specification quality and claim architecture — the two factors most within your control — rather than on expedited examination fees ($2,290–$4,580 for Track One depending on entity status) that accelerate a weak application into faster rejection.
Key Takeaways for Your Next AI Filing
Step 2A survival in 2026 comes down to three controllable variables: claims that name a specific technical improvement with quantitative support, specifications that build the factual record examiners must now engage with, and claim sets spanning system, method, and CRM formats so no single rejection is fatal. The USPTO's recent guidance raises the quality bar for rejections, which rewards applicants who gave examiners something rigorous to reject — and exposes applicants who did not. Draft as if the friendly examiner might leave tomorrow, because procedurally, they can.