Brazil: Implications of the draft AI guidelines for agrochemical inventions

AI is driving innovation in every single industry. In the agrochemical sector, AI is being increasingly used across the entire value chain to shorten development cycles, reduce costs and chemical use, improve regulatory compliance, and promote sustainable crop protection. In R&D, machine learning and generative models accelerate target identification, mode-of-action discovery, and denovo molecule design while optimizing synthesis planning. Safety and regulatory work benefit from AI-based toxicology and ecotoxicity predictions, as well as drafting support for dossiers. In product development, AI optimizes formulations for efficacy, stability, and sustainability. Precision agriculture applications include computer-vision sprayers for targeted herbicide use, decision support for application timing, and resistance management modeling. Beyond the field, AI also improves demand forecasting, supply chain planning, and market intelligence.

The use of AI in the development of inventions raise, however, profound questions in patent law. Across the globe, patent offices are grappling with issues of inventorship, patentability, enablement, and non-obviousness. Questions arise about how much human contribution is required to allow patent protection for an AI-assisted invention; what must be disclosed about datasets; and how to assess obviousness in a world where powerful tools are arguably available to the skilled person. Data ownership, trade-secret interfaces, and plausibility standards further complicate protection strategies.

Against this backdrop, the Brazilian Patent and Trademark Office (Instituto Nacional da Propriedade Industrial – INPI) released in August 2025 a draft guidance for examining AI-related patent applications (″Draft AI Guidelines″), which is open for public consultation until October 19. This article offers a practical reading of those draft guidelines from the perspective of agrochemical inventions.

Inventorship: AI-generated is a no; AI-assisted requires significant human contribution

The Draft AI Guidelines state unequivocally that outputs generated autonomously by AI, where a human being did little more than trigger a system, are not eligible for patent protection. Inventorship, under Brazilian law, must always be attributed to a natural person. By contrast, inventions assisted by AI can be patentable, but only where a human made an ″intellectual contribution.″

The draft fit squarely within the emerging international consensus. It mirrors the EPO’s DABUS decisions, which refused applications naming an AI as inventor, and the UK Supreme Court’s ruling that an inventor must be a person. In the United States, the USPTO’s 2024 inventorship guidance likewise permits protection for AI-assisted work only when at least one natural person made a ″significant contribution″ to each claim. Mere operation or review of AI output is insufficient.

The INPI’s proposal further explains that human intervention may happen in the identification of a technical problem in the state of the art, in the configuration of the system to reach a specific goal, in the validation of proposed results, or even in translating those results into a concrete solution with industrial application. If the AI model proposes alternatives, the inventive conception and the definition of what is claimed must result from human intellectual work.

This means if an AI autonomously suggests a promising compound, that suggestion alone cannot be the basis for a patent in Brazil. Patentability requires identifiable human contribution: in crop-protection or formulation workflows, it is the scientist who must frame the screening regime, impose agronomic and regulatory constraints, interpret the system’s recommendations in light of field realities, and converge on a particular composition or application method. These are the moments of ″intellectual contribution″ by a human being that the INPI’s draft guidelines require.

Under the Draft AI Guidelines, patent applicants are further required to provide enough description to show how the AI-assisted solution materializes in the real world. They also must demonstrate that the technical effects are truly achieved by said solution, and not merely hypothesized by the model. This is a nod to the risk of ″algorithmic hallucination.″

This requirement may, however, be dismissed when there is reliability in the result—for instance, when the efficacy was already known by a skilled person at the filing date.

Enablement: applicants must provide sufficient information

The Draft AI Guidelines acknowledge AI’s ″black box″ character yet emphasize that applicants must describe the invention in a clear and sufficient manner to enable a skilled person to reproduce it without incurring in undue experimentation.

That is, enablement does not require reproducing the exact performance metric, but the specification must provide enough technical details that allow a skilled person to reproduce the technical effect—such as the training/benchmark data (i.e., detail provenance, structure, and how relevant variables are obtained); the input–output correlation (i.e., explain why inputs plausibly yield the claimed outputs); and, depending on the case, essential model choices, parameters, training and validation procedures, and interactions with other technical components.

In the agrochemical field, an AI model might suggest formulations or spray regimes optimized against simulation or greenhouse surrogates. But if the application claims field-level efficacy or resistance management, and provides no credible route to practice those conditions, enablement is vulnerable.

Another issue relates to parameter-defined inventions (e.g., ″formulation with stability index ≥ X″ or ″sprayer system reducing drift by Y%″). If the application does not disclose a standardized method to measure those parameters, the scope is unclear and not reproducible.

Claims directed to ″training data,″ ″datasets,″ or ″databases″ are treated as presentations of information and excluded from patent protection. Claims must be framed around the technical application rather than around the dataset or model itself—for instance, ″method of identifying kinase inhibitors using […]″ or ″process for formulating microencapsulated pesticide with […].″