Writing Patents with AI: What Works (and What Doesn’t)

AI has become surprisingly good at generating text, summarizing documents, and assisting with research. In the LegalTech space, this has naturally led to a lot of discussion around whether AI can “draft patents.”

After working at the intersection of patent law, startups, and software teams for nearly two decades, I’ve found that the real answer is more nuanced. Patent drafting turns out to be a poor fit for one-shot automation — but a much better fit for workflow-level assistance.

This post breaks down where AI helps, where it struggles, and what that means for builders working on LegalTech or other document-heavy products.

Why patent drafting is hard to automate

Patent applications aren’t just legal documents. They are technical disclosures that must satisfy several constraints simultaneously:

• Describe an invention in enough detail to enable others to reproduce it


• Align the disclosure with best practices in the relevant technical field (software, mechanical, pharma, etc.)


• Demonstrate why the invention is novel over prior art (existing technologies that may be similar)


• Provide sufficient implementation detail while preserving flexibility for future claim amendments

Unlike many forms of writing, later sections of a patent application depend heavily on earlier decisions — especially in the claims. That dependency makes “generate the entire document” approaches brittle and difficult to control.

Where AI actually helps in patent drafting

In practice, AI performs best in patent drafting when it is used to assist discrete steps, rather than replace human judgment:

• Enhancing invention details provided by users by filling gaps and clarifying vague features


• Helping users think through alternatives, variations, and edge cases


• Translating engineering or technical concepts into clearer, patent-style descriptions


• Reducing blank-page friction during early drafting


• Drafting individual sections or sub-sections in a step-by-step workflow

This mirrors how developers often use AI today — not to ship production code unreviewed, but to accelerate iteration and exploration.

Where AI still struggles

AI tends to struggle in patent drafting when asked to:

• Decide what is legally or strategically important


• Identify which features are likely to make an invention novel


• Interpret large bodies of prior art to assess novelty or differentiation


• Make claim-scope tradeoffs without sufficient context

These limitations aren’t unique to patent drafting. They appear in many domains where correctness, strategy, and downstream consequences matter.

A workflow-first approach

Rather than asking “Can AI draft a patent?”, a more productive question is: Where in the drafting workflow does AI reduce friction without introducing unacceptable risk, and maximize the robustness of the disclosure?

That framing has guided the early-stage development of Idea2PatentAI, which integrates a step-by-step, AI-assisted patent drafting workflow designed to help users structure and document their inventions more clearly. The workflow emphasizes user control, allowing inventors to guide the process of patent drafting with AI assistance and focus the attention of the AI model on aspects of an invention that matter most — such as features that are novel or that differentiate the invention from existing solutions.

The emphasis is on guidance, iteration, and transparency — not automation for its own sake.

Why this matters beyond patents

This pattern shows up across LegalTech and developer tooling more broadly:

• AI works best when it augments decision-making, rather than replaces it


• Granular, reviewable outputs consistently outperform monolithic, one-shot generation


• Domain constraints matter more than raw language ability

As builders, aligning AI tools with real-world workflows tends to produce better outcomes than chasing full automation.

Community Feedback

I’d be interested to hear how others are thinking about this problem. If you’ve experimented with AI-assisted patent drafting or similar document-heavy workflows, I’d be curious to hear what has worked well — and what hasn’t.