# How Should Companies Build an AI Patent Eligibility Strategy in 2026?

patentreviewpro.com · September 24, 2026

> A Practical Answer for AI Patent Eligibility Strategy An effective AI patent eligibility strategy begins with the business problem, not with a claim to...

## A Practical Answer for AI Patent Eligibility Strategy

An effective AI patent eligibility strategy begins with the business problem, not with a claim to “an AI system” or a particular model architecture. As of 24 September 2026, patent offices and courts still ask whether an application discloses a patentable technical contribution rather than an abstract idea, mathematical rule, or computer program that operates only by conventional computer implementation. The same application can therefore pass one jurisdiction and fail another, even when the description is identical. A defensible strategy combines claim design, inventorship records, data governance, filing decisions, and a realistic view of prosecution cost. It does not treat AI as a special category that automatically receives or defeats patent protection.

**Also worth reading:** [Are AI Inventions Facing Higher U.S. Patent Eligibility Risk in 2026?](https://patentreviewpro.com/knowledge/are_ai_inventions_facing_higher_us_patent_eligibility_risk_in_2026.php) · [How Should Deepfake Technology Be Drafted for U.S. Patent Eligibility?](https://patentreviewpro.com/knowledge/how_should_deepfake_technology_be_drafted_for_us_patent_eligibility.php) · [What Does the 2026 USPTO AI Patent Eligibility Guidance Actually Change for Applicants?](https://patentreviewpro.com/knowledge/what_does_the_2026_uspto_ai_patent_eligibility_guidance_actually_change_for_applicants.php)

The central U.S. test remains the two-step framework associated with Alice and Mayo: a court first asks whether the claim recites a judicial exception such as an abstract idea, and then asks whether additional elements supply an inventive concept. A machine-learning model trained to make a prediction does not, by itself, clear that second step. Eligibility becomes more credible where the claim is tied to a particular technical problem, uses a specified structure or method that produces a technical effect, and includes limitations that prevent the claim from covering any generic implementation of the underlying idea. The EPO and UKIPO apply related exclusions with different wording, but each also considers whether the claimed subject matter makes a technical contribution.

A strong strategy also recognizes that eligibility is only one part of patentability. Novelty, inventive step, industrial applicability, sufficiency, clarity, entitlement, and unity must still be satisfied. AI inventions often present difficult prior art because public papers, code repositories, and product releases disclose model structures or equivalent methods before an examiner searches. The practical goal is not to promise broad protection for every AI feature. It is to identify the narrow set of technical contributions that can be claimed clearly, prosecuted under the applicable law, and enforced against commercially meaningful alternatives.

## How the United States, EPO, and UKIPO Compare

The three principal patent systems use related commercial-law exclusions, but they do not apply identical wording or case law. A comparison made on 24 September 2026 should therefore be treated as a working prosecution map, not a guarantee of identical outcomes. International counsel should check the relevant statute, examination guidelines, and current case law before relying on any of these positions in a filing or portfolio decision.

| Feature | United States | European Patent Office | United Kingdom |
| --- | --- | --- | --- |
| Primary eligibility test | Alice/Mayo two-step judicial framework | EPC exclusions, including mathematical methods and computer programs “as such,” interpreted in light of technical contribution | Patents Act 1977 exclusions for programs for computers “as such,” mental acts, and mathematical methods, subject to the statutory and case-law tests |
| Typical AI risk | Claim appears directed to an abstract idea implemented with generic computers | Claim appears directed to mathematics or data processing without a technical effect | Claim appears directed to a rule, mental process, or program of instructions without a technical contribution |
| Useful claim emphasis | Concrete technical improvement, unconventional arrangement, or specific implementation | Interaction between the method or device and physical or technical processes | Technical contribution, especially where the result goes beyond ordinary computer operation |
| Data considerations | Description must enable the invention; public disclosure can create prior-art or statutory-bar problems | Sufficiency and support remain important; mathematical notation and model operation must be understandable | Sufficiency and clarity remain important, particularly where the description relies on black-box assertions |
| Human contribution | A natural person must be an inventor under current USPTO guidance and law | A natural person must be the inventor, although a European patent can concern a process implemented with AI assistance | A natural person must be the inventor; company entitlement depends on employment and assignment rules |
| Term, assuming validity and grant | 20 years from the earliest effective nonprovisional filing date | 20 years from the international filing date | 20 years from the filing date, subject to renewal and other statutory rules |

The U.S. framework is often the most demanding for claims that use the language of prediction, classification, optimization, or data analysis. The EPO frequently permits a claim directed to a technical process when the application explains how the mathematics is implemented in a technical system. The UKIPO takes a similarly technical-contribution approach but exercises its own judgment under the Patents Act 1977. These are tendencies, not safe harbor rules: an application that survives a U.S. rejection may still encounter an EPO or UKIPO objection if its technical contribution is not visible in the claims.
International applications also divide ownership and enforcement. A single PCT application can defer national or regional decisions, but it does not create one worldwide patent, and it does not make inconsistent national eligibility law disappear. The EPO can grant a European patent, while enforcement may depend on national rights or the Unitary Patent in participating states. A company should select jurisdictions based on where competitors operate, where users and servers are located, where infringement can be proved, and where the cost of translation and opposition is acceptable.

## What Makes an AI Claim More Likely to Be Eligible

The strongest claims connect an AI method to a specific technical problem and a defined technical result. Examples include a controller that changes an industrial actuator setting, a network system that reduces a measurable communications failure, or a medical device that performs a physically meaningful operation using a trained model. The claim should identify the input, the processing steps, the technical resource being changed, and the output in terms that a competitor can compare. That structure is often stronger than adding words such as “artificial,” “neural,” or “automated” to an otherwise abstract instruction.

AI inventions are particularly vulnerable when the claim is limited to collecting data, applying a mathematical formula, and generating a recommendation. A generic processor, a conventional memory, or a rule that the model is trained on labelled data may not add enough to the underlying idea in the United States. In Europe and the United Kingdom, the examiner may instead ask whether the claim is directed to a computer program or mathematical method “as such.” The applicant’s response should identify the technical effect, explain why the effect follows from the claimed arrangement, and avoid asserting a technical result without a corresponding limitation in the claim.

Specification quality is central to that response. The description should disclose the relevant features, training approach, data relationships, control flow, and technical effect with enough detail to carry the claim. It need not disclose every production example, but it should not depend on an unexplained assertion that a black box produces a novel result. The EPO’s Enlarged Board decisions in G 1/19 and G 1/21 illustrate why AI-specific examination must focus on the actual claim and its technical context, not merely the commercial label “machine learning.”

The claims should also be written at more than one abstraction levels where appropriate. A system claim can capture hardware and control relationships, while a method claim can define a technical process and a narrower computer-readable medium claim can cover software distributed with a particular technical arrangement. These alternatives should not simply repeat one another. Each should have a different statutory role, and each should be supported by the description. A broad claim may be useful commercially, but an examiner’s rejection can impose costs or narrow the family to a level that no longer matches the product roadmap.

## Building the Application and Prosecution Workflow

The first practical step is a technical and legal audit before drafting begins. The team should identify the product version, the human problem being solved, the data sources, the people who contributed to the inventive concepts, and the competitor architectures it must distinguish from. It should also record which features are already public in papers, code, demonstrations, customer disclosures, or sales material. This audit often reveals that the strongest patentable concept is not the model’s ordinary prediction function but a particular control loop, memory architecture, sensor arrangement, or data-generation technique.

Drafting should then proceed from a claim map to the specification, rather than attaching a generic AI narrative to an initially broad claim. Counsel should define plausible claim variants and test each against the U.S. eligibility framework, EPC exclusions, and the UKIPO position. The application should explain the technical problem, the interaction between model components and other system elements, and the reason the claimed arrangement produces the stated effect. Particular attention should be paid to the source and structure of training data, the operation of a model when the data is incomplete or adversarial, and the way an output is applied to a physical or network process.

Before filing, the inventors, employers, universities, contractors, and data providers should be checked for ownership and entitlement issues. A named inventor is not necessarily the owner, and an AI vendor’s contribution does not automatically become a company patent. A complete record should show who conceived the claimed features, what contributions were made by human individuals, and what rights were transferred through assignment or employment. A laboratory notebook, design history, experiment log, and dated technical discussion can help establish that record. These materials are not substitutes for legal review, but they make later inventorship disputes considerably harder.

The filing decision must also account for publication timing. In the United States, many disclosures are evaluated against a 12-month inventor grace period, but the grace period is limited and does not protect every disclosure. Patent applications are ordinarily published 18 months after the earliest claimed priority date, although an applicant may request treatment under specified conditions. A company should therefore obtain advice before a paper, conference talk, open-source release, customer demonstration, or repository publication. If secrecy is important, a provisional application may preserve an early date, but a provisional is not examined and does not mature into a patent unless the applicant files an appropriate nonprovisional application within the applicable 12-month period.

## Inventorship, AI Assistance, and Data Rights

The law does not treat every use of AI as equivalent. A human can use an AI tool to explore alternatives, draft language, analyze results, or suggest a model configuration without becoming an inventor merely because the tool generated an output. Inventorship generally turns on contribution to the conception of the claimed invention, not on the amount of typing, coding, or documentation performed. USPTO guidance and U.S. case law require human inventorship, and a “solely AI-generated” invention cannot be resolved simply by listing a company as the applicant. The application must identify natural persons who made the relevant inventive contributions.

That distinction makes contribution records important. If one engineer designed the technical problem and another selected the specific arrangement that solved it, both may need to be named even if a model or tool proposed alternatives. If a contractor suggested the claimed feature but did not participate in conception, the contractor may still create a contractual or ownership issue even if the feature is not attributed to the contractor as an inventor. Counsel should not insert a name to satisfy a filing formality. The named inventors should be those who actually contributed to conception, subject to applicable law and the proper correction procedure.

Training data raises separate questions that patent law does not answer. A dataset may contain personal information, copyrighted material, database rights, contractual restrictions, or trade secrets. Granting a patent on a method does not grant a right to train on a dataset or use all of the data described in the application. Copyright status can differ from patent status, and a public dataset can be free of some contractual restrictions while remaining subject to copyright or database rights. The U.S. Copyright Office has rejected the proposition that AI output receives copyright protection without human authorship, and the 2025 D.C. Circuit decision in Thaler v Perlmutter reinforced the human-authorship requirement in that dispute. The patent specification should therefore avoid presenting data access as automatic permission.

Confidential data needs a separate release decision. Publishing a sample or a metric may improve sufficiency and help establish technical effect, but it can also destroy secrecy and make later enforcement harder. A staged approach is often sensible: preserve confidential details, disclose enough to support the claims, and consider whether selected examples can be released after filing. The choice depends on the value of public disclosure to validity, the risk of a requirement for greater disclosure, and the likelihood that competitors can infer the protected method from the filing. No single rule works for every company.

## Comparing Patents with Other Protection Options

Not every commercially valuable AI capability should become a patent application. Patents can exclude independent implementation, but they require public disclosure, cost money, and expose the invention to validity challenges. Trade secrecy may be preferable for model weights, orchestration rules, tuning processes, and operational information that can remain hidden. The right question is whether secrecy is commercially realistic and whether independent discovery or reverse engineering is likely. A feature published in a technical paper, API documentation, or customer contract is generally a poor trade-secret candidate even if the company labels it confidential.

| Decision factor | Patent route | Trade-secret route | Defensive publication route |
| --- | --- | --- | --- |
| Main benefit | Potential exclusion of independent commercial use in covered territory | Potentially indefinite protection without disclosure | Prevents others from obtaining a later patent on the published subject matter |
| Main cost | Drafting, examination, translations, renewals, and possible opposition | Controls, access management, employment measures, and incident response | Loss of secrecy and limited ability to exclude others |
| Best subject | Defined technical contribution with a commercial market and practical detection | Weight values, tuning data, internal workflows, or implementation know-how | Low-value technical information that is likely to appear in the public domain anyway |
| Main risk | Eligibility, prior art, validity, cost, and disclosure of the core method | Leakage, lawful reverse engineering, and inability to prove secrecy | Public teaching may be followed by commercial use, and the disclosure may not defeat every patent route |
| Practical timing | Before material disclosure, with a controlled filing timetable | Before any unauthorized disclosure, often during development | When secrecy has little value and preventing a monopoly is useful |

Copyright can protect source code, documentation, interfaces, and some human-authored expression, but it does not generally protect the functional idea behind an AI method. It can also be difficult to apply to generated material where human authorship is uncertain. A copyright claim is therefore often complementary to a patent, not a replacement. Open-source licensing and contractual terms may provide another layer of protection, although an open-source license can restrict proprietary use of code and may conflict with a planned proprietary claim if the relevant material is essential.
A portfolio should consequently mix protection types according to the asset’s lifetime and detectability. Fast-changing model parameters may be better managed as trade secrets or controlled know-how, while a novel sensor-control architecture or specialized hardware interaction may justify patent review. The patent decision should also consider acquisition and licensing value. A publicly disclosed but difficult-to-detect method may have weaker practical value than a secret improvement that a customer can observe, even if the latter would be harder to exclude through litigation.

## Common Mistakes That Weaken AI Patent Strategy

One frequent error is assuming that use of neural networks makes an invention eligible. A trained network is an implementation technology, not a claim limitation that automatically supplies technical character in every jurisdiction. Another error is selecting a claim level based on the company’s preferred product implementation rather than on the commercially important concept. If the product changes before filing, a narrowly drafted claim may become irrelevant, while a broad claim may be rejected as an abstract or purely mathematical idea. Claim design should test both a defensible core and at least one commercially realistic alternative.

A second common mistake is treating the specification as a marketing document. Statements that a system is “faster,” “more accurate,” “secure,” or “intelligent” do not establish a patentable technical effect unless the application explains how the claimed structure produces that effect. Unbounded accuracy claims can also conflict with ordinary variation in real data. Counsel should use measurable mechanisms and examples that can be tied to the claimed features. This is especially important for generative AI, where the definition of a meaningful output can depend on the task, the evaluation method, and the surrounding technical process.

Inventorship and disclosure errors are more damaging because they can affect the whole application. Naming an AI system as an inventor, relying on a vendor’s blanket assignment, or failing to identify a university contributor can create entitlement and validity problems. Filing after a public demo, using an expired priority chain, or treating a provisional filing as the end of the process can also defeat intended protection. Companies frequently discover these problems only during a due-diligence review, when the cost of correction is much higher.

Finally, a strategy focused on one jurisdiction is a commercial mistake, not merely a legal one. U.S. eligibility risk, EPO technical-effect objections, UKIPO computer-program exclusions, and post-grant opposition procedures operate on different schedules. A company that assumes an identical outcome everywhere may spend years pursuing claims that would have had little value if drafted for the intended market. A modest initial review in each relevant jurisdiction is usually more useful than a large portfolio of untested applications.

## Timing, Cost, and When to Act

A company should begin the review before the first external disclosure, often while the system is still in engineering development. The early review need not produce a filing on the same day. Its purpose is to identify the critical inventive concepts, preserve evidence, decide which data can be disclosed, and establish a deadline for a provisional or nonprovisional filing. Waiting until a launch announcement or conference submission removes options and can force a rushed description that fails to explain the technical contribution.

A reasonable internal timetable is measured in stages rather than fixed promises. During the first zero-to-three-month period, a cross-functional team can document the problem, architecture, human contributions, and disclosure history. Between roughly three and nine months, counsel can prepare claims, test prior art, evaluate jurisdiction-specific eligibility, and decide whether a provisional is justified. The following nine-to-eighteen-month period can include a formal application, office-action responses, national-phase planning, and any required translations. The actual schedule depends on the maturity of the invention and the examination queue.

Official fees are only one part of the budget. USPTO fee reductions for qualifying small and micro entities are commonly calculated at 60% and 80%, respectively, on many applicable official fees, but eligibility for those rates and the current fee schedule must be checked at filing. A utility patent commonly incurs several thousand dollars in official USPTO charges over its life, while drafting and prosecution by private counsel can add approximately $8,000 to $30,000 or more for a straightforward, noncontested matter. EPO and UKIPO official fees are often lower than U.S. fees, but translation, attorney work, validation, renewal, and opposition costs can narrow the difference. International portfolio spending should be compared with expected licensing or litigation value rather than with another company’s filing budget.

The best time to act depends on the asset. File when a defined technical contribution is sufficiently mature to disclose, a competitor or market makes exclusion valuable, and the company can meet the applicable deadlines. Preserve secrecy when the value depends on remaining unknown and detection is unlikely. Defer the decision when the engineering is experimental, but document why the decision is deferred and calendar the next review. A useful decision record should state the eligible subject matter, rejected alternatives, public-disclosure status, expected cost, likely jurisdictions, and the person responsible for revisiting the decision.

No AI patent strategy can guarantee a grant, enforceability, or commercial advantage. It can improve the probability that a worthwhile technical contribution is identified, described, claimed, and maintained through prosecution. The strongest approach as of 24 September 2026 is disciplined: treat AI as a set of specific technical solutions, separate invention from software and data rights, and test each jurisdiction’s actual law rather than relying on a global assumption.

## Quick answers

### Is an AI invention automatically patent-eligible in 2026?

No. In the United States, an AI claim can be rejected if it recites an abstract idea implemented with generic computers under the Alice/Mayo framework. The EPO and UKIPO also examine whether the claim is a mathematical method or computer program “as such,” and a technical effect usually needs to be visible in the claim and supported by the description.

### Can a company patent an invention conceived by AI?

Current U.S. guidance requires a natural person to be an inventor, and patent law generally does not recognize an AI system as the inventor. Humans can still use AI for research, coding, and drafting, but they must make and document the inventive contributions claimed in the application.

### When should a company file before disclosing an AI product?

The review should begin before a paper, conference talk, demo, repository release, or customer disclosure that could trigger a filing deadline or prior-art problem. A provisional application may preserve an early date, but it must be followed by an appropriate nonprovisional filing within the 12-month priority period.

### Is a provisional application enough for an AI patent?

No. A provisional application is not examined and does not itself become a patent. It can establish an early filing date, but the applicant must later file a nonprovisional application and satisfy all requirements for patentability, including sufficiency, eligibility, inventorship, and ownership.

### Are AI patents worth the cost compared with trade secrets?

The choice depends on whether the technology can remain hidden and whether competitors can detect its use. Patents offer potential exclusion rights in covered jurisdictions but require public disclosure, while trade secrets may protect internal methods longer but offer no reliable remedy against independent development or lawful reverse engineering.

Canonical: https://patentreviewpro.com/knowledge/how_should_companies_build_an_ai_patent_eligibility_strategy_in_2026.php
Markdown: https://patentreviewpro.com/knowledge/how_should_companies_build_an_ai_patent_eligibility_strategy_in_2026.php/index.md