AI tools now draft useful disclosures, embodiments, and first-pass specifications. They still break at claim strategy, enablement, and inventorship.
Yes for the first half of the job, no for the half that decides whether the patent is worth anything. Current AI tools are good at the inventor’s side of a patent application: prior-art triage, structuring the disclosure, expanding embodiments, and a first-draft specification. They break at claim strategy, at enablement and written description, at inventorship, at confidentiality, and at citations, which they will invent with a straight face.
Technical founders ask us a version of this most weeks now, usually as “we have a good model and a $20 subscription, why pay a patent attorney five figures?” Our read is that AI has made the inventor’s contribution to a patent bigger and cheaper, and left the drafter’s contribution about where it was.
In our experience, four things.
Vendors put the time savings at 30 percent per draft, and 40 to 60 percent for experienced users on individual tasks (Solve Intelligence). The same write-up concedes that picking the right claim scope for a specific competitive landscape stays with the attorney. The savings are real. They land on the parts of the job that were never the hard part.
The claims are the patent. Everything else is support. Claim strategy is the piece no current tool does, because it is not a writing problem.
Good claims answer business questions. Which competitor products need to read on claim 1? What does a design-around look like, and does a dependent claim close it? Would a buyer or licensee three years from now pay for this? The model does not know your competitors, your roadmap, or your exit.
Our house view is that claim writing is closer to an art than a form, and the drafter’s skill shows up directly in what the patent is later worth. Breadth has to be managed in both directions: too broad and the examiner rejects it, too narrow and a competitor steps around it. Two patterns we see in AI-drafted claim sets: steps split across two parties (divided infringement, which in our experience is much harder to enforce), and coverage of the product as built rather than the markets it could reach.
Thompson Patent Law’s July 2026 take is that AI can produce text that looks like a patent application while functioning as nothing of the kind (Thompson Patent Law). We agree with the distinction, and part ways with their flat “no” on the disclosure half. That is real work, AI does it well, and a founder who does it well walks into the attorney’s office with a cheaper and better engagement.
The statute is short. 35 U.S.C. 112(a) requires the specification to describe the invention and how to make and use it “in such full, clear, concise, and exact terms as to enable any person skilled in the art” to do the same (35 U.S.C. 112). The USPTO’s examination manual adds the line that bites AI drafts hardest: “The more one claims, the more one must enable” (MPEP 2164).
Two failure modes. First, the model expands embodiments past what anyone on the team knows how to build, the claims follow, and breadth outruns the teaching. Second, and sneakier, the model fills gaps in the disclosure with plausible technical detail the inventors never conceived and cannot back up. Baker Botts flagged this in April 2026: AI-generated material muddies the question of which parts of the claimed invention the named humans actually conceived (Baker Botts).
The provisional version of this trap is the one founders hit most. The government fee for a provisional is $65 for a micro entity, $130 for a small entity, and $325 undiscounted (USPTO fee schedule), so the temptation is to file the AI draft tonight and fix it later. But the early date only holds for what the provisional actually teaches; a thin one protects little, and what you disclosed in it can end up as prior art against your later filings (what a provisional does). And the USPTO’s guidance on AI tools notes that patching the specification or drawings after filing may be new matter (USPTO, April 2024).
This one has settled down. On November 28, 2025 the USPTO published revised inventorship guidance for AI-assisted inventions (90 FR 54636) and rescinded its February 2024 guidance in its entirety (USPTO announcement). There is no separate or modified standard for AI-assisted inventions. AI systems are tools used by human inventors, analogous to laboratory equipment, software, or research databases, and only natural persons can be inventors. When one person develops an invention with AI help, the only question is whether that person conceived it under the ordinary standard; the joint-inventorship factors from Pannu do not apply, because the AI is not a person. With several people, ordinary joint-inventorship rules apply among the humans.
For a founder: name the humans who conceived each claimed invention, keep notes on who decided what, and do not let the model’s embodiments become claims nobody on the team conceived. The USPTO’s separate April 2024 guidance on AI tools in filings, still in force with no new rules, says there is no general duty to tell the office you used AI, but the duty of candor applies if the AI’s role becomes material to patentability, for example where it introduced embodiments the inventors did not conceive (USPTO, April 2024). That is a legal call. Ask your patent attorney; we are a strategy firm, not a law firm.
Before you paste your invention into anything, ask the process questions first. Hugh AI, our conversational agent built on ipCG’s methodology and the Invent Anything archive, is a low-commitment way to work through what a disclosure needs, how to expand embodiments, and when to bring in a practitioner: try Hugh AI.
The consumer tier of a general-purpose assistant is the wrong place to describe an unfiled invention. The USPTO’s 2024 guidance says it plainly: AI systems may retain what users enter, the owner can use that data to train models or hand it to third parties, and servers outside the U.S. can pull in export-control and foreign-filing-license rules (USPTO, April 2024). Baker Botts recommends enterprise platforms with contractual confidentiality commitments and an internal rule against consumer tools for disclosures (Baker Botts). We run the same way: our AI policy commits that client data is never used to train third-party models and that every AI-assisted deliverable gets professional review before it goes out.
Whether typing an invention into a chatbot is a public disclosure that starts a novelty clock is unresolved. Smart & Biggar’s October 2025 analysis says courts have not addressed it, and the stakes differ by jurisdiction: the U.S. gives inventors a 12-month grace period for their own disclosures, while the European Patent Office allows almost none (Smart & Biggar). Our read: do not run that experiment on your own company. Use a tier with a written no-training commitment and keep a log of what you entered and when. File before you widen the circle.
The models make things up, and the patent system has started punishing it. In August 2026 the USPTO’s Office of Enrollment and Discipline posted its first AI-related discipline order: a patent attorney was publicly reprimanded after a claim construction chart he drafted with generative AI cited specification passages, figures, and prosecution history that did not exist or were not where the chart said (IPWatchdog). In February 2026 a Kansas federal judge in Lexos Media IP v. Overstock.com fined the plaintiff’s lawyers a combined $12,000 for a brief full of nonexistent quotations and citations (Missouri Lawyers Media).
For a founder the relevant version is prior art. When a tool hands you the closest references, open every one. Confirm the number exists and the passage says what the summary claims. The USPTO’s rule for anyone who signs a filing is that simply relying on the accuracy of an AI tool is not a reasonable inquiry (USPTO, April 2024). Hold yourself to the same standard before a search decides what you file.
Where does that leave the five-figure question? Our disclosure guide treats application drafting as the largest single cost of most patents and puts the prosecution time a strong disclosure saves at 30 to 40 percent. That is the savings to go after. A good disclosure shrinks the attorney’s hours. It does not remove them.
If you have an AI draft on your desk right now, do not file it tonight. Have someone who drafts claims for a living read the claims against the specification and against your two nearest competitors’ products. A few hours of that review is cheap next to a provisional that holds a date for the wrong invention.
If you want a second set of eyes on the strategy side (the disclosure, the embodiments, what to claim for which market, what to keep as a trade secret), that is the work we do. Talk with our team and bring the draft.
Work with ipCapital Group
From invention to monetization, our team has guided 2,000+ engagements across the full IP lifecycle. Start with a free 30-minute discovery call.
Written by
Seth Cronin