One Disclosure, Three Standards: AI Patent Drafting for the USPTO, EPO and Indian Patent Office (IPO)

You filed the same machine-learning invention in three offices and received three unrelated objections. The USPTO Examiner said the claim recites an abstract idea without integrating it into a practical application. The EPO Examiner said the computational model makes no technical contribution and, separately, that the training data is insufficiently disclosed. The IPO Examiner said the claim falls under Section 3(k). None of these is a prosecution problem you can fully argue your way out of. AI patent drafting for the USPTO, EPO and IPO is a drafting problem, because the material each office wants must be present in the specification on the day you file. This article sets out what each office currently requires, explains why the divergence cannot be retrofitted, and describes how a drafting workflow can build all three positions into a single priority document.

What the USPTO requires

Eligibility runs through Alice/Mayo, and for AI claims the decisive point is almost always Step 2A, Prong Two: whether the claim integrates the judicial exception into a practical application. The 2024 Guidance Update on Patent Subject Matter Eligibility, Including on Artificial Intelligence (effective 17 July 2024, with Examples 47 to 49) is the practical text. Read alongside the examples, three things carry the weight. The claim should be limited to a particular field of use rather than reciting the model in the abstract. The specification should contain a technical explanation of how the invention improves the technology. And the claims should recite limitations that actually implement that improvement. The Federal Circuit's 2025 decision in Recentive Analytics tightened the point. Applying a known machine-learning model to a new data environment is not enough, and asserting that the model produces better accuracy or speed does not convert an abstract idea into eligible subject matter. What survives is a claim tied to a specific technical mechanism, described in the specification with enough detail that the improvement is visible rather than asserted. The USPTO issued a further reminder memorandum to examiners on evaluating eligibility under §101 in August 2025, and has continued to update its position since. Note one thing the USPTO has been explicit about: how an invention was created, including whether AI tools were used, is not part of the eligibility analysis. Inventorship is a separate question, and there the rule is firm: an AI system cannot be named as an inventor, and a natural person must have made a significant contribution.

What the EPO requires

The EPO treats artificial intelligence and machine learning as mathematical methods, addressed at G-II, 3.3.1 of the Guidelines for Examination. The current edition came into force on 1 April 2026. The structure is a two-hurdle test. A claim directed to a method involving technical means, or to a device, has technical character as a whole and is not excluded under Article 52(2) or (3). Clearing that hurdle is easy and worth almost nothing. The second hurdle is the inventive step under the COMVIK approach: features that do not contribute to technical character cannot support the inventive step. A computational model earns its place only if it contributes to a technical solution to a technical problem: either by being applied in a field of technology, or by being adapted to a specific technical implementation. Two consequences follow for drafting. First, the specification must identify the technical problem and explain how the algorithmic feature solves it. A description that explains what the model does, but not what technical effect it produces, leaves the examiner nothing to work with. Second, sufficiency under Article 83 has a specific edge for machine learning: where the technical effect depends on particular characteristics of the training dataset, those characteristics needed to reproduce the effect must be disclosed, unless the skilled person could determine them without undue burden. The Guidelines also warn that terms such as "neural network" or "support vector machine" are read carefully, because they usually denote abstract models with no technical character of their own. Then there is Article 123(2). Whatever technical explanation you did not include at filing, you cannot add later.

What India requires

Section 3(k) excludes a mathematical or business method, a computer programme per se, or algorithms. The word doing the work is per se, and the route past it is a demonstrated technical effect or technical contribution. The Indian Patent Office issued revised Guidelines for Examination of Computer Related Inventions on 29 July 2025, replacing the 2017 guidelines. They matter to drafters for three reasons. They set out a step-wise methodology for assessing Section 3(k), illustrated by roughly twenty examples in the body and a further forty in an annexure distinguishing allowable from non-allowable claims. They add a dedicated chapter on emerging technologies including AI, machine learning, blockchain and quantum computing. And they add a chapter on sufficiency of disclosure, requiring claims to be supported by technical description. They also incorporate the jurisprudence (Ferid Allani, Microsoft Technology Licensing, OpenTV) confirming that an invention otherwise caught by Section 3(k) remains patentable where it shows a technical effect, a technical contribution, or a technical solution to a technical problem. Two practical points recur: where a claim uses means-plus-function language, the structural features must be disclosed in the specification; and system and method claims can each be granted where the specification supports them. The old shorthand that Indian practice simply demands novel hardware is no longer an accurate frame. What it demands is implementation detail sufficient to show technical effect.

The reason this is a drafting problem

Look at what the three offices want and notice that the overlap is smaller than the union. The USPTO wants a technical explanation of the improvement, tied to a field of use, implemented in the claim limitations. The EPO wants the technical problem stated, the algorithmic contribution explained, and, where the effect depends on it, the relevant training-data characteristics disclosed. India wants implementation detail sufficient to establish technical contribution, with structural support for any means language. A specification drafted to satisfy one of these will be thin for the other two. And each of the three has a hard door that closes at filing: Article 123(2) at the EPO, written description under §112 in the US, and the sufficiency requirements applied in India. Prosecution can argue about what the specification means. It cannot add what the specification lacks. The divergence has to be resolved in the first draft, before you know which office will object to what. That is why this is a drafting-depth problem rather than a workflow problem, and why the stage that determines the outcome is the one worth investing in. USPTO EPO IPO Central test Practical application under Step 2A, Prong Two Technical contribution to a technical solution (COMVIK) Technical effect or contribution, past per se under s.3(k) Governing text 2024 AI SME Update, Examples 47–49 Guidelines G-II, 3.3.1 (in force 1 April 2026) Revised CRI Guidelines, 29 July 2025 What the specification must carry Technical explanation of the improvement Technical problem stated; training-data characteristics where the effect depends on them Implementation detail; structure for means language Claim-level requirement Limitations that implement the improvement Features contributing to technical character Support in the technical description Door that closes at filing §112 written description Article 123(2) Sufficiency of disclosure

How Esgenix approaches AI patent drafting for the USPTO, EPO and IPO

Esgenix is an agentic patent workflow platform. An agent here means a piece of software given a defined job, the tools to do it, and a standard to meet: it completes the job and reports back rather than waiting for the next instruction. Several run in sequence on different underlying models: one searches the prior art, one drafts the claims, one writes the description, one produces the figures, one grades the result. US, European and Indian practice are native, which matters here specifically: the three sets of requirements above are applied to the same disclosure in the same session, rather than to three separately drafted texts. Patent AI is moving from assistance toward workflow execution, and products differ mainly in where they draw the line between machine and professional. It sits beside the attorney while the attorney writes and proposes the next passage. Esgenix runs the drafting work itself and stops at defined gates for the attorney's decision. Where a tool is shaped around assistance, the professional stays present through the drafting itself. With Esgenix, the attorney makes five decisions and the platform does the production between them. The gate that matters most for this subject is the first. At intake, the platform reads the disclosure and flags what is missing (an unexplained step, an undefined term, a result asserted with no mechanism) and proposes how to close each gap. That is precisely the failure mode that produces all three objections above. An inventor writes that a model improves prediction accuracy and does not say by what mechanism. The USPTO calls it an abstract idea, the EPO calls it no technical contribution, and India calls it Section 3(k). Catching it at intake is the difference between one email to the inventor in hour one and three unanswerable objections in year two. At the strategy gate you set claim breadth on the scope slider: an on-screen control that widens or narrows the claim set as you drag it, with support re-checked against the disclosure at each position. At the claims gate you approve or amend the set; nothing proceeds without it. At the specification and figures gate you review the draft with provenance visible on every claim element: provenance meaning each element carries an identifier tracing it to the passage of the disclosure it came from, so you verify support by following a pointer rather than reading two documents in parallel. That check is the practical answer to the Article 123(2) and §112 problem: you can see, before filing, which limitations rest on solid disclosure and which do not. The fifth gate is a signature, by a natural person, and it cannot be delegated. The EPO's Guidelines state plainly that parties and their representatives remain responsible for the content of applications and submissions regardless of whether a document was prepared with AI assistance. Esgenix is built on that assumption rather than around it. Two further mechanisms matter for AI inventions. The prior art search runs across 220 million documents from 108 jurisdictions and matches on meaning rather than keywords, and its output feeds the drafting agents in the same session, so the claim set is positioned against the closest art from the first draft, and the specification carries fallback positions where the art is dense. An independent model grades the finished draft against the standards a senior patent attorney would apply, returning it to be rewritten when it fails. In a field where the failure mode is a plausible-sounding paragraph that never states a mechanism, that quality gate is doing exactly the work the three offices are asking for. Figures are generated with no baseline drawing required: the system reads the specification, decides what the figure set should be, draws it, and reconciles every reference numeral against the text: useful in a field where the "drawing" is a data-flow architecture nobody sketched. Esgenix has completed ISO 27001 Stage 1, covering the platform and the handling of invention and client data.

A specific next step

Take an AI application you filed in all three offices in the last two years and read only the priority document. Check three things: whether it states a technical problem and the mechanism by which the model solves it; whether it discloses the training-data characteristics the effect depends on; and whether every means-type limitation has structural support in the description. Where the answer is no, the objection you eventually received was written on the day you filed. Bring that same original disclosure to a working session with us and run it through the five gates with all three jurisdictions selected. Then compare what comes out against what you filed.