The Direct Answer
An effective AI patent strategy combines legal protection, commercial secrecy, and disciplined filing decisions rather than treating every model improvement as a patent candidate. The central question is not simply whether an invention uses artificial intelligence, but whether a human has made a patentable contribution, whether the claimed subject matter is eligible, whether the disclosure satisfies the written-description and enablement requirements, and whether the commercial value justifies the cost of enforcement. As of 24 September 2026, companies should assume that an AI system may be a tool used to invent, but it cannot be named as the inventor of a U.S. patent application. A practical portfolio therefore combines selected patent filings with trade-secret controls, copyright protection for software and documentation, contractual restrictions on data and model access, and rapid publication decisions for research that should remain open.
Also worth reading: What are the current AI patent inventorship requirements for global intellectual property filings? · What are the definitive legal standards for AI patent inventorship in 2026? · AI Patent Eligibility in 2026: What Could the Supreme Court Change?
The strategy should begin with an invention-disclosure process that records human contributions, technical problems, experimental evidence, dates, and ownership. It should then separate three issues that are often confused: inventorship determines who must be listed, ownership determines who can license or enforce the right, and eligibility determines whether the claim can be patented at all. A company may have valid inventorship and ownership while still presenting a weak Section 101 case, or it may have a technically eligible application that fails because the specification does not explain how to reproduce the invention. The best route depends on the nature of the asset, the markets served, the likelihood of reverse engineering, and the time required to commercialize it.
For a B2B intellectual-property rights team, the objective is a repeatable portfolio process, not a one-time filing campaign. Counsel and product leaders should agree on decision thresholds, deadlines, and escalation rules before a product launch. A rights and registry SaaS platform can help organize disclosures, docket deadlines, assignments, family relationships, and status information, but it does not replace legal judgment about eligibility or inventorship. The durable strategy is the combination of documented human contributions, technically grounded claims, controlled disclosure, and a filing calendar that is reviewed at least quarterly.
Why AI Changes the Patent Analysis
AI changes both the speed at which technical improvements are created and the difficulty of proving who created them. Machine-learning systems can generate candidate architectures, optimize parameters, identify unusual material combinations, or propose control policies in hours rather than months. That speed increases the value of preserving evidence before employees move, vendors publish details, or competitors file similar applications. It also increases the risk that a company will disclose more than necessary in a patent application or internal invention memo. The legal question is consequently tied to information governance as much as it is to patent prosecution.
The U.S. human-inventor requirement is the first threshold. The Supreme Court’s decision in Thaler v. Perlmer and subsequent USPTO guidance establish that an inventor must be a natural person who contributes to conception, not merely a person who owns the resulting system or supervises an automated process. The February 2026 USPTO policy materials referenced in the supplied research context reinforce the practical need to identify and document human contributions to the claimed features. A named employee must understand the invention and make a material contribution to at least one claim; a person who only supplies funding, direction, or access to a computing platform is not automatically an inventor. AI-generated suggestions can inform the application, but they should not be presented as independent inventive acts.
Eligibility is a separate inquiry under 35 U.S.C. § 101. An applicant should expect the USPTO to examine whether the claim is directed to a judicial exception, such as a mathematical concept, and, if it is, whether the claim recites significantly more than the exception by imposing a particular inventive concept. A generic instruction to use a neural network to predict an outcome will not automatically become eligible because it is implemented on a computer. By contrast, a claim tied to a specific technical improvement in memory usage, network reliability, industrial control, signal processing, or energy consumption may present a stronger case. The relevant facts must appear in the claim and be supported by evidence in the specification.
Ownership also requires attention. Employment agreements, contractor agreements, university licenses, collaboration agreements, and customer-funded development terms can create overlapping rights or gaps in the chain of title. The same distinction applies to training data and model weights, which may have contractual restrictions even when the underlying patent application is owned by the company. Copyright does not solve every gap: the U.S. Copyright Office’s human-authorship rule means that protection for purely AI-generated expression may be limited, while human selection, arrangement, modification, or compilation can be protected. Patent, copyright, trade-secret, and contract rights should therefore be planned as a package rather than as competing alternatives.
A Practical 90-Day Process
The first step is to create a cross-functional review group involving patent counsel, product engineering, research, security, finance, and commercial teams. A useful operating cadence is a 30-day intake and evidence-preservation phase, a 60-day technical and prior-art assessment phase, and a 90-day filing-or-protection decision. During the first 30 days, teams should freeze relevant source-code versions, model checkpoints, laboratory records, design documents, and correspondence. The purpose is not to create litigation evidence immediately, but to make later verification possible. Inventors should record which human proposed a feature, which experiments were run, and why a particular technical limitation was selected.
At the 60-day point, counsel should conduct a preliminary search covering patents, non-patent literature, product documentation, standards, and public code. The search should ask whether the proposed feature was already known, whether the improvement is merely a routine application of a known model, and whether the commercial advantage comes from implementation details rather than the alleged invention. A company can use a simple scoring rule: require a reproducible technical effect, a plausible route to exclusivity, and a decision on whether secrecy is more valuable than publication. For example, a 70 percent threshold for technical reproducibility can be used as an internal screening rule, although it has no legal meaning by itself.
By day 90, the team should choose a route and document the reasons. A provisional application can be appropriate when the invention is still being developed, provided the written description supports the later claims and the business can preserve the 12-month priority deadline. A nonprovisional or PCT filing may be preferable when the technical picture is mature and international protection matters. A trade-secret route may be better for a model-training pipeline whose value depends on confidential data or tuning methods. The decision record should identify the human inventors, assign ownership, estimate the first three years of cost, and state what evidence will be needed during prosecution.
The process should also establish a publication gate. Conference papers, demos, open-source releases, grant applications, and customer documentation can become prior art or trigger public disclosure deadlines in some foreign jurisdictions. A single chart in a paper may disclose the core invention even if the paper does not use patent terminology. Engineers should submit publication plans for review at least 60 days before external release when a patent application is under consideration. This is one of the highest-return controls in an AI portfolio because a missed disclosure window cannot normally be repaired by paying more to counsel later.
Drafting Claims That Survive Examination
AI claims should describe a technical solution, not a business aspiration expressed in technical language. Weak drafting says that a system uses a model to classify data, optimize a process, or predict a user preference. Stronger drafting identifies the input, the processing structure, the technical constraint, the operation performed, and the measurable technical result. A useful claim architecture includes a broad independent claim directed to the technical improvement, narrower claims covering particular model or hardware arrangements, and fallback claims protecting a control method, system, or apparatus when the product is implemented differently. Alternatives should be supported by the original disclosure, not added at the last minute as unsupported wishful thinking.
The specification should explain why the proposed solution works and how a skilled engineer would reproduce it. If the application relies on improved accuracy, latency, memory use, energy consumption, safety, or network performance, the application should state the metric, baseline, test conditions, and acceptable range. An example might describe an accuracy improvement from 92 percent to 96 percent under a defined data distribution, but the numbers must come from actual experiments. Results that are not measured should be described as expected or theoretical, not presented as demonstrated performance. Experimental detail also helps with Sections 102 and 103, because a broad claim may be novel yet obvious over several known references.
Prosecution should address eligibility before examiner objections become expensive to correct. Counsel can develop a technical-effects record showing how the claimed arrangement changes computer operation or solves a technical problem, while avoiding the mistake of treating the mere use of a computer as the inventive concept. The application should distinguish the claimed improvement from a field-specific mathematical rule or generic automation instruction. If the invention is genuinely difficult to explain because it relies on a model, the specification should still identify the relevant architecture, data flow, constraints, and human-designed control points. An examiner is more likely to engage with a concrete technical mechanism than with a statement that the system uses deep learning.
A portfolio team should also review whether patent protection is being asked to do work better suited to trade secrecy. Patent applications disclose the mechanism needed for enablement and public enablement after publication, while trade-secret protection can last indefinitely if secrecy is maintained. In practice, many AI systems are better divided into a public interface and proprietary components, with the latter kept out of the application. This is not a choice between filing everything and filing nothing; it is a choice about which layers of the stack should become public, which should remain controlled, and which combinations should be protected in more than one form.
Comparing U.S., European, and UK Treatment
The United States, European Patent Office, and United Kingdom do not apply identical eligibility standards, although all require a real technical contribution and proper attribution of human work. The comparison below is a planning summary, not a substitute for jurisdiction-specific advice. International filing decisions should be made before the Paris Convention priority period expires and should account for local translations, representation, validation costs, and national implementation requirements.
| Feature | U.S. USPTO | European Patent Office | United Kingdom IPO |
|---|---|---|---|
| Core eligibility approach | Alice/Mayo two-step analysis under Section 101 | Inventive-step analysis focused on technical contribution and the closest prior art | Technical contribution and inventive step, including computer-implemented invention guidance |
| AI treatment | A human must be identified as the inventor; AI assistance does not itself establish inventorship | Human inventive contribution is required, and the claimed feature must make a technical contribution | Human contribution is expected, and the claimed feature should provide a technical effect |
| Typical evidence | Architecture, control logic, benchmark data, and a concrete computer or industrial improvement | Technical effect, comparison with the closest prior art, and implementation details | Technical effect, technical problem, and how the computer-implemented feature solves it |
| Common deadline planning | Paris priority at 12 months; PCT national phase generally at 30 or 31 months | Paris priority at 12 months; European and national-phase costs differ by route | Paris priority at 12 months; PCT and national-phase fees and translations require separate budgeting |
| Strategic consequence | Prepare for examination focused on judicial exceptions and prior art | Avoid claims that reduce to mathematics or business rules without a technical contribution | Build a clear technical-effect narrative and expect examination of computer-implemented subject matter |
A useful international plan normally preserves a 12-month Paris Convention deadline, then uses a PCT filing to buy time for market and cost decisions. National-phase entry commonly occurs around 30 or 31 months from the priority date, depending on the jurisdiction, but the exact date must be calculated from the filing record. The PCT does not create one worldwide patent; it can provide a substantial search and examination report before national decisions. If commercial protection is only needed in two countries, direct filings may be cheaper. If a product will be sold in five or more jurisdictions, the administrative cost of a coordinated family may be easier to justify, although translations and local representation can still dominate the budget.
Alternatives, Costs, and Portfolio Trade-Offs
Trade-secret protection is often the first alternative for training recipes, curated datasets, model weights, annotation processes, and internal evaluation methods. It can provide protection without a public disclosure date, but it requires access controls, confidentiality agreements, logging, offboarding procedures, and incident response. A trade-secret decision should be reviewed whenever a partner receives access, a code repository is made public, or a patent application is drafted. A patent may still protect a public-facing technical mechanism while the company keeps the model-training process secret, provided the application does not disclose the secret itself.
Copyright and contract are also relevant. Copyright can cover source code, documentation, original diagrams, human-authored training materials, and certain compilations, but it does not give exclusive rights to an abstract algorithm or automatically protect purely AI-generated expression. Open-source licenses may permit use of code while imposing attribution, notice, reciprocity, or patent-license conditions. A company should review the license before incorporating a third-party model or dataset, and should avoid using customer data in training without an appropriate contractual basis. These rights are not substitutes for a patent when the goal is to exclude independent implementations of a technical mechanism.
For planning purposes, a straightforward U.S. utility filing often consumes approximately $3,000 to $8,000 in official fees and outside search or drafting expenses, while a more technically involved family can exceed $15,000 during the first several years. A PCT and subsequent national-phase filings can bring the initial international cost into the $15,000 to $40,000 range, depending on the number of countries, translations, entity status, amendments, and prosecution complexity. These are budgeting ranges rather than official quotations, and USPTO fees change periodically. Large, small, and micro entity status can materially change official fees, but status must be valid at the time of filing; a company should not select a category merely to reduce cost without meeting the applicable size requirements.
Registry and portfolio software may be priced by user, portfolio volume, matter type, or a negotiated annual subscription, so no responsible answer should claim one universal SaaS price. The relevant evaluation is whether the system records assignment events, monitors family deadlines, supports role-based access, preserves audit history, and exports information that counsel can verify. A low subscription fee can be offset by missed deadlines or inconsistent ownership records. Buyers should compare implementation effort, data migration, security controls, customer support, and integration with engineering systems, not just the monthly license.
Common Mistakes and Timing Triggers
The most damaging mistake is naming an AI system, a model developer, or a customer as the inventor. Another common error is treating a person who supervised the project as the inventor without identifying the human conception they performed. Counsel should also reject applications that merely say the model is novel because it was trained on a particular dataset. A dataset, parameter value, or prompt is not automatically an invention, and a new use of a known model may still face an obviousness or technical-effect objection. Inventorship records should be corrected promptly if the application names the wrong person, using the applicable USPTO correction procedure.
Timing failures are equally expensive. A provisional filing must be prepared within 12 months of the earliest contemplated public disclosure or foreign filing, and a later nonprovisional must claim the appropriate priority. A PCT application generally needs to be filed within 12 months to preserve priority, while national-phase entry is commonly due at 30 or 31 months. A product launch, investor presentation, standards submission, or open-source release may constitute public disclosure even when no patent is mentioned. Companies should create alerts at 90, 60, and 30 days before each major deadline, with a responsible attorney and business owner assigned to every alert.
Another mistake is filing too early, before the company knows whether the feature works outside a benchmark. Claims supported by speculative performance can fail enablement or written-description requirements, while a rushed application can disclose commercially important details. The opposite error is waiting for a product to become successful before preserving priority, because competitors may have already filed and public disclosure may have started. A balanced approach is to file a focused provisional when the core mechanism and experiments are sufficiently defined, then use the following year to validate the product and decide which claims deserve national or international spending.
The portfolio should also be reviewed when the company changes direction. A model that becomes an internal service may need secrecy controls rather than a new filing; a customer-requested feature may trigger a joint-development agreement; and an acquisition may require review of chain of title. A five-year horizon is useful for budgeting, but a patent’s nominal term is generally measured from the earliest relevant nonprovisional filing and can be shortened by prosecution delay, fees, disclaimers, or non-payment. Renewal and maintenance decisions should be made using expected revenue, remaining term, enforceability, and the availability of alternative protection.
Building a Repeatable Governance System
The strongest AI patent programs operate like a small investment committee rather than an informal email chain. Counsel should chair a monthly intake review, product leaders should explain business value, engineers should verify technical detail, and finance should approve expected spend. A standing committee can use a written rubric covering human inventive contribution, technical effect, novelty distance, freedom to operate, secrecy risk, market coverage, and total cost. The rubric should permit a decision not to patent. A feature that will be obsolete in eight months may be better documented as a trade secret or released as open source than subjected to a four-year prosecution campaign.
Metrics should measure both speed and quality. Management may review the number of disclosures received, percentage completed within 60 days, filings made before public disclosure, percentage of applications with verified human inventorship, family cost variance, and outcomes at first office action or allowance. A program that produces many filings but has repeated eligibility or enablement objections has a volume problem, not necessarily a success. Conversely, a small number of carefully selected families can outperform a large low-value portfolio. The committee should also track incidents involving source-code leaks, unauthorized model access, missing assignments, and missed deadlines.
Ownership data should be treated as an operational asset. Assignors, inventors, employers, universities, vendors, and customers should be recorded in a way that can be audited, and every material agreement should be linked to the relevant disclosure or application. Registry SaaS can provide a shared view for legal and product teams, but the organization still needs a policy for who may approve a filing, amend an inventor record, execute an assignment, or send an application to a foreign office. Access permissions should follow least privilege, and exports should preserve timestamps and version history.
The practical answer is to act now on the next three AI-related technical improvement, not on every idea in the backlog. Identify the human contributors, preserve records, run a focused prior-art and eligibility review, and make a deliberate choice between patent, trade secret, copyright, contract, or publication. Repeat that process at least quarterly, and escalate any launch or disclosure that falls within 12 months of a possible filing date. This approach does not guarantee a patent or eliminate legal risk, but it gives a company a defensible process for preserving value while avoiding both automated inventorship errors and unnecessary filing costs.
Sources and Current Guidance
The supplied research points to reporting from JD Supra, Massachusetts Lawyers Weekly, IPWatchdog.com, Skadden, and Holland & Knight on AI patent eligibility, prosecution, and the relationship between AI and intellectual property. The sources below are publisher and institutional entry points for current guidance; counsel should verify the latest USPTO, EPO, and UKIPO materials before filing because eligibility guidance and fees can change. The U.S. Copyright Office’s current human-authorship position should also be checked when software, training material, or generated documentation is part of the commercial asset.