How to File a Patent Application for an AI Invention in 2026

This guide is about the filing work. Whether an AI claim survives 35 U.S.C. 101 is covered in the machine learning eligibility guide, and choosing counsel in the guide for AI startups. Most weak AI applications are weak because the disclosure meeting stopped at the architecture diagram.

• Data pipeline. Sources, labelling, cleaning, augmentation, sampling. If a preprocessing step is why the thing works, that step is part of the invention.
• Architecture and training. Model type and size, layer structure, loss function, objective, schedule, and the problems you had to solve to get convergence. The workarounds are often the patentable part.
• Inference and deployment. What the model sees at runtime, what it outputs, latency and memory constraints, and what consumes the output.

• The enablement bar: a skilled engineer could rebuild the system without your team. This is the one AI applications most often miss.
• Architecture in reproducible detail, including the parts you consider obvious, the training data characteristics that matter, and at least one worked example with representative inputs and outputs.
• Ranges rather than single points: layer counts, learning rates, thresholds, and window sizes, so the claims have room.
• Alternatives you considered. Every one written down is claim scope available later; every one left out is not.

Draft a system claim, a method claim, and a non-transitory computer-readable medium claim, then decide separately whether each covers training, inference, or both. Training and inference are different acts, frequently performed by different parties on different machines. A single claim that requires one actor to train the model and then run it against a customer's data can be very hard to enforce. Split them, and keep each claim performable by one party. Figures matter too: a system block diagram, a training flow, and an inference flow give the examiner something concrete to map claims onto.

Only natural persons can be inventors. The USPTO's revised inventorship guidance for AI-assisted inventions, published at 90 FR 54636 and effective November 28, 2025, states that AI systems are tools used by human inventors and do not qualify for inventor status. It rescinded the February 13, 2024 guidance in its entirety, so there is no special AI inventorship test to pass. The ordinary requirement of human conception governs as it always has. Name the people who conceived it, and keep records of when they did.

Get the technical content down before you worry about claim language. A provisional with the real architecture, the real training details, and one worked example is worth far more than a polished summary. Use the twelve months to see whether the approach holds. If a decision date matters, Track One is the speed lever, at $903 for a micro entity per the USPTO fee schedule. Attorney fees are in the cost guide.

Move the capture upstream into engineering. A disclosure template that asks for the data pipeline, the training procedure, and the deployment path produces filings that need one drafting cycle instead of three. Decide as policy whether you file on training, inference, or both, based on where your competitors sit in the stack. And keep inventorship records contemporaneous: across a large team, who conceived what is easier to answer from notebooks than from memory two years later.