IP Strategy Memo: Patenting AI-Assisted Code

KNOWLEDGE (WHAT)

The Core Legal Reality (USPTO 2024/2025)

The USPTO has issued critical guidance regarding AI-assisted inventions over the last two years, clarifying the boundaries of inventorship and subject matter eligibility.

AI Cannot Be an Inventor: The bedrock principle is that only natural persons (human beings) can be named as inventors on U.S. patent applications. [USPTO Feb 2024 Guidance]
The Tool Doctrine: AI systems, including generative models, are treated legally as tools—analogous to software compilers or laboratory equipment—that assist human inventors. [USPTO Nov 2025 Revised Guidance]
Conception is Key: An inventor using AI tools must demonstrate that they conceived the complete and operative invention. Identifying a general problem is not enough; the human must shape the conception, such as through specific prompting, training, or post-processing.
Subject Matter Eligibility (The Alice Trap): You cannot patent "an AI that does accounting." Claims must integrate the AI into a practical application that provides a specific technological improvement, avoiding rejections for claiming abstract ideas. [USPTO 2024 Subject Matter Eligibility Guidance]

The Operators vs. The Bureaucracy

The current legal framework creates a distinct dynamic between founders/operators and the patent system.

The USPTO's Balancing Act: The agency is attempting to encourage AI innovation while preventing the "over-attribution" of inventorship to individuals who merely press a button on an AI interface.
The "Named Inventor" Requirement: As the founder (Stephen), you must be listed as the inventor. The burden is on you to document your specific contributions (architecting the RAG schema, designing the graph layout) to prove the invention wasn't entirely autonomous.
Global Discrepancies: While the US strictly requires human inventors, filing international patents requires careful navigation, as priority claims cannot link back to foreign patents that name AI as an inventor.

GENERATIVE (WHAT IF)

Execution Strategy for ETA & LEDGER

Given that raw code is covered by copyright and abstract ideas are unpatentable, what is the optimal IP strategy for your tech-enabled accounting roll-up?

1. Patent the Architecture, Not the Code

Do not attempt to patent the Python scripts. Instead, patent specific, novel methods. For example:

The Compliance Bypass: A specific data-routing architecture that allows an offline, localized LLM to query a restricted Graph Database (client financials) without data leaving the environment, thereby satisfying SOC/GAAP privacy constraints.
State-Management Validation: A multi-agent loop that translates fluid natural language into rigid state-tracking metrics required for verified CPE credit in a gamified environment.

2. The Operator's Move: Trade Secrets > Patents

Patents take years and require public disclosure of your methods. For software operators, Trade Secrets are often the superior moat.

Lock down your custom system prompts, RAG database schemas, and fine-tuned weights.
Host the models locally to maintain absolute control over the proprietary workflows.
By the time a patent would be approved, AI capabilities will have shifted; keeping the architecture a closely guarded Trade Secret allows you to iterate faster than competitors can copy you.