A neural network claim most commonly lands in Prong One through the mathematical-concepts grouping. Training and inference are typically recited as mathematical relationships: inputs are transformed through layers of weighted calculations, parameters are adjusted by optimization routines, and outputs are generated by applying learned functions. When the claim language describes the network primarily in these terms, the sorting procedure places those limitations in the mathematical-concepts bucket, yielding a “yes” at Prong One.
To apply this sorting procedure before filing or responding, map every claim limitation—not just the preamble or the stated goal—to one of the three groupings. Treat each step of data preparation, model training, inference, and output use as a separate unit of analysis and ask which grouping, if any, it recites. This limitation-by-limitation mapping reveals whether the neural network is claimed as an abstract idea or as part of a broader arrangement that must be evaluated further.
If the mapping shows that the network appears only as a mathematical optimization or calculation, the claim is sorted into Prong One as reciting an abstract idea, and the eligibility analysis must proceed to Prong Two. If the limitations cannot be fairly placed in any of the three groupings, Prong One is answered “no,” and the claim is eligible without reaching Prong Two.
What the July 2024 Update Actually Changed
The USPTO's 2024 guidance update on AI-related inventions took effect on July 17, 2024. Per the Mintz analysis of the update, it was prompted by Executive Order 14110 on safe, secure, and trustworthy AI. Treat that date as the operative trigger for any filing or response strategy that relies on the update's examples.
The update does not reflect any new USPTO practice or procedure and is stated to be consistent with existing guidance. That no-new-procedure finding matters because the 2019 PEG's Step 2A framework still governs eligibility analysis; the update did not replace that framework, add a new procedural step, or change the order in which eligibility is analyzed. For practitioners, that means prior Step 2A arguments remain usable; only the illustrative AI fact patterns are new.
What changed is the addition of AI-specific worked examples showing how the framework applies to AI-related claims. Those examples are the update's operative contribution: they illustrate application of the existing analysis rather than announcing a new eligibility rule, a new category of patentable subject matter, or a new drafting shortcut. Use them as illustrations, not as substitutes for claim-specific analysis.
The worked examples are framed as applications of the same Step 2A analysis, not as a separate or stricter test. Check the claim against the existing framework first, and use the examples to anticipate how an examiner may map AI limitations to the abstract-idea categories and to concrete applications.
Because the framework did not change, the filing or response question remains whether each limitation is tied to a concrete application rather than left as an abstract idea. The update's examples may help anticipate that mapping, but they do not create a new safe harbor or excuse a limitation-by-limitation review before filing or responding.
A practical check is to compare each AI limitation in the draft to the fact patterns in the update's examples, then ask whether that limitation is doing work in a concrete application. If it only restates a model, training routine, or optimization objective, the example may signal risk even though the underlying rule is unchanged.
Drafting Paths Compared: Which Survives
When evaluating a neural network claim under Step 2A Prong One, the drafting strategy determines whether the claim reads as a concrete application or as an abstract idea. Comparing four common claim-drafting paths against the 2024 USPTO AI subject-matter eligibility guidance shows that tying the network to a specific technical application provides the clearest route to survival, with Path D emerging as the strongest overall approach.
| Path | Drafting Strategy | Prong One Risk | Rationale |
|---|---|---|---|
| A | Claim the network abstractly (e.g., “a processor configured to run a neural network that computes an output”). | Highest | The network reads as a mathematical concept with no integrating application. |
| B | Tie the network to a specific technical field and input type (e.g., network traffic anomaly detection or adaptive signal processing). | Lower | The application context supplies the concrete-use anchor reflected in the 2024 worked examples. |
| C | Recite specific training steps, data transformations, or architectural features without a defined technical use. | Moderate to high | Added detail may still describe mathematical or mental processes if it is not directed to a practical application. |
| D | Combine a field-specific application (Path B) with particularized implementation details (Path C). | Lowest | The claim anchors the network in a concrete use and explains how the recited features operate within that context. |
Path A carries the highest Prong One risk because it presents the neural network as a generic computational tool. Before filing, check whether each limitation does more than invoke a model that computes an output; if the claim language could describe any neural network in any context, it is vulnerable to being characterized as an abstract idea.
Path B reflects the approach the 2024 examples reward. By specifying the technical field and the nature of the input, the claim frames the network as part of a concrete application rather than as a standalone mathematical construct. Path C can strengthen a disclosure, but reciting training or architectural detail alone does not guarantee eligibility if the claim lacks a defined practical use.
For the strongest position, drafters should pursue Path D: pair the concrete-use anchor of Path B with the implementation specificity of Path C. In a response or pre-filing review, confirm that every neural network limitation maps to the claimed application and that the specification supports how those limitations function together in that technical context.
Costs and Numbers to Check Before Filing
Treating a Step 2A Prong One rejection as a budgeting problem means counting the cost of getting it wrong in prosecution cycles, not in guessed dollar amounts. The arithmetic is straightforward: each rejection you did not pre-empt adds at least one Office Action response, and if the eligibility issue remains unresolved, it can add a request for continued examination (RCE) or an appeal. Because official fees and professional rates change, the only reliable total is one you compute from figures you verify at the time you file or respond.
Start by logging the current government fees from the U.S. Patent and Trademark Office (USPTO) Fee Schedule at uspto.gov. Before budgeting a §101 response cycle, confirm the current RCE fee, any excess-claim fees that apply to your application, and, if you anticipate escalation, the fees associated with an appeal. Record the effective date shown on the schedule and the entity status under which you qualify (micro, small, or large), since the same fee item can carry different amounts depending on status.
| Figure to verify | Where to check | What to confirm |
|---|---|---|
| RCE fee | USPTO Fee Schedule (uspto.gov) | Current amount, entity status, and effective date |
| Excess-claim fees | USPTO Fee Schedule (uspto.gov) | Current thresholds, per-claim amounts, and entity status |
| Appeal-related fees | USPTO Fee Schedule (uspto.gov) | Applicable stages, current amounts, and entity status |
To estimate your exposure, build a simple worksheet: multiply the number of anticipated response cycles by your expected professional cost per response, then add the verified USPTO fees for each cycle. Recompute the total whenever a figure changes, and keep the source and verification date next to each entry so the budget reflects the schedule in force rather than an outdated assumption. If a figure is unavailable or uncertain, leave it as a line item to confirm rather than substituting an estimate.
Use this arithmetic as a pre-filing check. Compare the cost of an additional drafting or review pass against the cumulative cost of one or more response cycles. Framing the decision this way makes the financial consequence of an avoidable Prong One rejection explicit and gives you a documented basis for allocating time and resources before the first Office Action arrives.
Where Prong One Breaks: Edge Cases
The application-anchor rule is a limitation-by-limitation test, not a label test. A neural network claim survives Prong One only when at least one limitation performs a concrete application; the surrounding field, purpose, or industry name cannot do that work. Map each limitation separately, because an examiner can group one limitation as the abstract idea while another limitation supplies the concrete application.
A field tag alone breaks the rule. A recital such as "in the field of network security" supplies context but no operation. If the remaining limitations describe training, weighting, inference, or optimization, the claim still maps to the mathematical-concepts grouping. The field name does not become a concrete-application limitation merely because it sounds technical, and it cannot rescue a claim whose operative steps are all mathematical.
A mixed mental-process and hardware claim can still win, but only if the hardware limitation is the one doing the concrete application. Do not average the limitations. Identify which limitation the examiner grouped as the abstract idea, then argue from the limitation that applies the network's output to a machine, process, or article. If that hardware limitation is removed and only math remains, the claim breaks; if it stays and performs the application, the claim can survive even though a mental-process step appears elsewhere.
Post-training use is the closest call. A claim that merely says the trained network is "used to predict" or "used to classify" often remains in the mathematical-concepts grouping, because prediction is still a mathematical operation. It wins only when the use limitation ties the output to a specific technical act—such as controlling a valve, selecting a dosage, or routing a packet—and that act is the limitation doing the concrete application. The closer the use is to an ordinary business or diagnostic conclusion, the more likely it is to be treated as abstract.
| Edge case | Application anchor | Prong One result |
|---|---|---|
| Field tag only ("in the field of network security") | None beyond the field recital | Breaks; still maps to mathematical concepts |
| Mental-process step plus hardware step | Hardware limitation performing the concrete application | Wins if that limitation is the one grouped and argued |
| Post-training use | Output tied to a specific technical act | Wins only if the use is concrete, not mere prediction |
Before filing or responding, list every limitation, mark the one that performs the concrete application, and test whether deleting it leaves only mathematics. If it does, the anchor is not real and Prong One breaks. If the marked limitation is the one the examiner grouped differently, reframe the argument around that limitation rather than around the field tag or the model's purpose.
Anomaly Detection Claim
Consider the input claim: “A processor configured to detect network anomalies by feeding traffic features into a trained neural network and classifying each connection as anomalous or normal.” Before filing or responding to an Office action, map each limitation against the abstract-idea groupings under the USPTO’s 2024 AI subject-matter eligibility update and MPEP § 2106. The worksheet below captures that Prong One mapping and preserves the reasoning you will need if the analysis advances to Prong Two.
| Claim limitation | Checkpoint | Prong One mapping | Filing note |
|---|---|---|---|
| Feeding traffic features into a trained neural network and classifying each connection as anomalous or normal | Checkpoint 1 | Maps to mathematical concepts because classification is performed as computation | Record the mathematical character explicitly; do not rely on this limitation alone for eligibility |
| Traffic features and detecting network anomalies | Checkpoint 2 | Supplies a concrete application in network security outside the abstract-idea groupings | Tie the recited features to the specific technical environment and the anomaly-detection function |
| Claim as a whole | Checkpoint 3 | Prong One support is present, but integration remains a separate question | Treat any Prong Two integration argument as a fallback, not the only argument for eligibility |
Checkpoint 1: Isolate the neural network operation. “Feeding traffic features into a trained neural network and classifying each connection as anomalous or normal” describes a computational process that maps inputs to labels. In a Prong One mapping, that limitation falls within mathematical concepts. If the claim were limited to this operation, the record would show the network appearing only as mathematical optimization or classification, leaving the claim exposed at Step 2A.
Checkpoint 2: Identify the limitations that anchor the computation to a specific use. “Traffic features” and “detect network anomalies” tie the classification to network-security monitoring rather than to data processing in the abstract. These limitations supply a concrete application outside the abstract-idea groupings, so the worksheet should connect the claimed inputs and output to the practical task of identifying anomalous connections in network traffic. This connection is the limitation-by-limitation basis for surviving Prong One.
Checkpoint 3: Preserve a fallback position. Even when the Prong One mapping supports a concrete application, document how the additional elements integrate the recited computation into that application. Under the USPTO’s Step 2A framework, that integration analysis belongs at Prong Two; it strengthens the record but should not be treated as the only argument for eligibility. A complete worksheet makes the primary Prong One theory and the secondary Prong Two theory independently reviewable.
Five If/Then Rules for Prong One
Each of the five if/then rules below starts from a limitation-by-limitation map of your neural network claim against the 2024 guidance's three abstract-idea groupings and tells you the next move — file, re-draft, or respond. Run all five before every filing and every response.
If every limitation maps to one of the three abstract-idea groupings, restructure around a concrete application before filing. When the network, its inputs, and its outputs all land in the groupings, expect a Prong One rejection — and a Prong Two integration argument will have nothing to integrate, because the claim recites no application to integrate. Re-draft the claim, not the response: bring the concrete application into the claim itself before the application is filed.
If the network appears only as a mathematical-optimization step, rewrite it before anything else. Training, weighting, minimizing a loss function — optimization language with no field attached reads as a mathematical concept, whatever the preamble promises. Tie the network to a specific input type and a technical use, and the mapping changes; leave it abstract, and the claim stays in the groupings.
If the claim recites the network plus a specific input type and technical use, map the application limitation first and lead your response with it. That limitation is your Prong One anchor: identify it explicitly in the mapping chart, then build the argument from it. Keep the Prong Two integration argument in reserve — spending it before Prong One is resolved uses your best material on a question the examiner has not reached.
If your only anchor is a field-of-use recital, treat the claim as unanchored and add an application limitation. "For use in medical imaging" or "applied to financial transactions" tells the examiner where the invention might live, not what it does. Recite the concrete practice instead: the specific data source the network ingests, the machine it operates, the technical outcome it produces.
If generic hardware is doing the work, the mapping does not carry the claim out of the groupings. "A processor, a memory, and a neural network" recites components, not an application. The network must carry the technical use itself — the input it receives, the transformation it performs, the machine it drives — or the claim stays where the mapping placed it, however the hardware is arranged.
What to do next
| Step | Action | Why it matters |
|---|---|---|
| 1 | Map every limitation of your neural network claim to one of the three abstract-idea groupings in the 2024 guidance: mathematical concepts, certain methods of organizing human activity, or mental processes. | Creates the limitation-by-limitation record required to perform a proper Step 2A Prong One analysis. |
| 2 | Identify any limitations that describe the neural network solely as a mathematical-optimization engine without tying it to a concrete application. | Pinpoints the specific claim language most likely to fall within an abstract-idea grouping under the 2024 guidance. |
| 3 | Restructure those limitations around a concrete application so the claim reflects how the neural network is applied, not just how it optimizes mathematically. | Directly addresses the canonical decision rule for surviving Step 2A Prong One. |
| 4 | Compare your revised claim language against the USPTO’s worked examples revised in the 2024 AI eligibility update. | Benchmarks your claim against the examples the USPTO uses to illustrate eligible and ineligible neural network applications. |
| 5 | Review your analysis against MPEP 2106 to ensure each limitation is evaluated consistently with the 2024 guidance. | Aligns your eligibility reasoning with the USPTO’s governing framework before submission. |
| 6 | Document the complete limitation mapping and concrete-application rationale before filing a new application or responding to the USPTO. | Provides a clear, defensible basis for eligibility during examination under Step 2A Prong One. |
Frequently Asked Questions
Does describing a neural network mainly as weighted calculations and optimization routines place it in Prong One?
When the claim language describes the network primarily in these terms, the sorting procedure places those limitations in the mathematical-concepts bucket, yielding a yes at Prong One.
Should only the preamble be mapped when applying the sorting procedure?
To apply this sorting procedure before filing or responding, map every claim limitation, not just the preamble or the stated goal, to one of the three groupings.
How should data preparation, model training, inference, and output use be analyzed?
Treat each step of data preparation, model training, inference, and output use as a separate unit of analysis and ask which grouping, if any, it recites.
What happens if the network appears only as a mathematical optimization or calculation?
If the mapping shows that the network appears only as a mathematical optimization or calculation, the claim is sorted into Prong One as reciting an abstract idea, and the eligibility analysis must proceed to Prong Two.
When is Prong One answered no without reaching Prong Two?
If the limitations cannot be fairly placed in any of the three groupings, Prong One is answered no, and the claim is eligible without reaching Prong Two.
What does limitation-by-limitation mapping reveal?
This limitation-by-limitation mapping reveals whether the neural network is claimed as an abstract idea or as part of a broader arrangement that must be evaluated further.
Quick answers
| Through which grouping does a neural network claim most commonly land in Prong One? | A neural network claim most commonly lands in Prong One through the mathematical-concepts grouping. |
| How are training and inference typically recited in neural network claims? | Training and inference are typically recited as mathematical relationships: inputs are transformed through layers of weighted calculations, parameters are adjusted by optimization routines, and outputs are generated by applying learned functions. |
| What approach should be taken to apply the sorting procedure before filing or responding? | Map every claim limitation—not just the preamble or the stated goal—to one of the three groupings, treating each step of data preparation, model training, inference, and output use as a separate unit of analysis. |
| When must the eligibility analysis proceed to Prong Two? | If the mapping shows that the network appears only as a mathematical optimization or calculation, the claim is sorted into Prong One as reciting an abstract idea, and the eligibility analysis must proceed to Prong Two. |
| When is a neural network claim eligible without reaching Prong Two? | If the limitations cannot be fairly placed in any of the three groupings, Prong One is answered 'no,' and the claim is eligible without reaching Prong Two. |
Also worth reading: 2024/25 §101 Guidance: 58% AV-Cluster Rejection vs Overlooked 63%: 2024/25 §101 Guidance: 58% AV-Cluster · 2026 USPTO AI Guidance: AV Claim Architecture Under 101: 2026 USPTO AI Guidance: AV · USPTO 2025 AI Rule: Binary Choice, 40% Rejection Mean: USPTO 2025 AI Rule: Binary