AI Review
Written by Ben Esplin
What Inventors Give Me Now II (The Review)
In my earlier post, “What Inventors Give Me Now,” I wrote about how language model technology has changed the disclosure materials inventors prepare for me, and my ability to absorb these materials. They help me get up to speed on unfamiliar technical ground and synthesize richer, more detailed disclosure materials. That gives inventors less reason to simplify their inventions for my benefit and more opportunity to explain what matters. But that post covered only one stage of patent application preparation: the inventor helping me understand the invention. Another stage that has been transformed by AI agents is the review inventors give my work before we finalize and file an application.
The feedback I receive increasingly goes beyond technical corrections and engages with the strategic choices reflected in a draft. With help from agents powered by language model technology, inventors can develop more specific questions about claim scope, the treatment of prior art, or how an argument characterizes the invention. We often discuss these suggestions and discard them for one reason or another. But some contain genuinely useful insights, and that mixed experience is what interests me. The feedback does not have to be consistently right to make the review more valuable.
Inventors have always contributed to patent strategy. They understand the technology, the alternatives, and often the commercial stakes in ways the attorney needs to learn. But recognizing that a draft does not quite capture something important is different from being able to explain where the problem lies. An agent can help turn “I’m not sure this covers what matters” into a question about a particular limitation, a proposed alternative, or an explanation of what the inventor thinks the draft leaves out. That does not make the resulting analysis authoritative, but it can give us something more concrete to examine together.
Suppose an inventor questions whether a claim is unnecessarily tied to one implementation and returns a proposed revision. The revision might not work: it could overlook a reason for the existing language or introduce a different problem. Yet the discussion might reveal an alternative or more specific description that is inserted in the specification. Rejecting the proposed language would not make that exchange a failure, because the useful contribution was the concern it brought into focus. Other times, the proposed criticism may simply be right. I may have framed something poorly, overlooked an implication, or made a choice that deserves reconsideration, which is part of the reason to ask for review in the first place.
There is a cost to this process. A developed critique takes time to evaluate, and a longer review is not necessarily a better one. Often suggestions rest on mistaken premises, miss context, or question deliberate tradeoffs that still make sense after another look. The attorney still has to distinguish a useful insight from a plausible-sounding objection, and earning trust from the inventor is a key to making these interactions efficient. I would therefore not measure the value of AI-assisted review by how much feedback it generates or how many suggested edits survive. I would ask whether it helps us identify a worthwhile change, uncover a concern, or better understand a decision.
That brings me back to the relationship I described in the earlier post. Language model technology can help me engage more deeply with an inventor’s technical knowledge, and it can help an inventor engage more deeply with the work I produce from it. Neither direction requires our expertise to become interchangeable. I do not expect an inventor’s agent to resolve the strategic questions in a patent application, but I welcome help bringing worthwhile questions into the conversation. Sometimes the answer is to leave the draft alone, and sometimes the draft gets better. The opportunity is a review process in which more of the inventor’s understanding can inform that decision.