The AI-Enabled Patent Office: What the USPTO's AI Strategy Means for High-Volume Prosecution

• Robert Stoll, citing his contacts at the USPTO: a soft rollout to SPEs began in the fourth quarter of 2025, with an internal system used for claim construction, summaries, priority data checks, mapping under Section 102, combinations under Section 103, step-by-step Section 101 analysis, and Section 112 clarity and support. He sees no problem with Section 102 mapping or Section 112 clarity, has significant concerns about Section 103 combinations and Section 101, and argues for pilots applicants can opt into.
• Bernard Tomsa: once a tool's output reaches examiners it behaves like guidance, but unlike guidance it cannot be read, challenged and corrected. And if AI makes repeated rejections costless for the office, nothing slows a cycle of non-final actions.
• Tomsa's proposal: test the office's tools on thousands of past cases with known outcomes before rollout, then give examiners standardized prompts, so practitioners know every examiner queries the model the same way.
• On models trained on past decisions: a PTAB decision is not automatically a good decision, so the training data has to be reviewed by practitioners before it is treated as ground truth.
• On strategy, practitioners usually keep arguments narrow to limit prosecution history estoppel. Facing an AI-assisted examiner that may simply concede the one error you point out and issue another non-final action, Tomsa expects responses may need to lay out the full position, accepting some estoppel risk.
• Stoll does not expect pendency to drop to a few months soon: roughly 1.2 million applications, by his count, still need an examiner's analysis.

For as long as most practitioners have been answering office actions, the safest response has often been the shortest one that works. Point out what the cited art does not teach, put the burden back on the examiner, keep the file wrapper thin. Every argument made on the record can later narrow how the claims are read, and a short response gives an accused infringer less to work with.

That habit rests on an assumption about who sits on the other side: a human examiner who, once shown the flaw in a rejection, usually allows the claim or issues a better action. At our July panel with IPWatchdog, Bernard Tomsa of Brooks Kushman asked what happens to that habit when the examiner works with an AI system that can concede the error you point out and produce a new rejection just as quickly. His answer, and the rest of the panel's, is the subject of this piece.

Why the thin response became the default

Prosecution history is read against the patentee. Narrowing amendments can surrender equivalents (Festo Corp. v. Shoketsu Kinzoku Kogyo Kabushiki Co., 535 U.S. 722 (2002)), and clear statements made to distinguish prior art can disclaim claim scope (Omega Engineering, Inc. v. Raytek Corp., 334 F.3d 1314 (Fed. Cir. 2003)). The rational response is to say as little as the rejection requires. If the examiner relied on a passage that does not teach the limitation, saying so is usually enough.

What changes when the examiner's tool is AI

According to Robert Stoll, former Commissioner for Patents, citing his contacts at the office, a soft rollout to supervisory examiners began in late 2025, with an internal system covering claim construction, prior art mapping under §102, combinations under §103, step-by-step §101 analysis and §112 support. Stoll sees little risk in the mapping work and real risk in obviousness combinations and eligibility, where the output depends on judgment.

The panel's concern was less about errors than about what happens after one is corrected. Tomsa described a recent case in which the examiner's citation to page 3 of a secondary reference was wrong, but the teaching was there on pages 4 through 7. A human examiner told that page 3 is wrong rarely comes back with the same reference and new page numbers. A system that treats your response as one more prompt can do exactly that: agree it was wrong, find the passage elsewhere, and issue another action. When the corrected rejection is a new ground not necessitated by an amendment, the next action cannot be made final (MPEP 706.07(a)). The cycle repeats as another non-final, and nothing in the process makes the repetition costly for the office.

The trade-off: a fuller record against another round

Tomsa's conclusion was that practitioners may have to lay out their full position, not only the flaw in the rejection, and accept some estoppel risk to end the conversation about a reference. That is a trade-off, not a new best practice. A fuller response can shorten prosecution but puts more on the record. A thin response preserves scope and may cost another round of fees and months.

Which way to go is a case-by-case call. It depends on facts only the attorney holds: how important the claim is commercially, how likely litigation is, what the client is willing to pay for, and how this examiner has behaved so far.

Six questions before filing a response

1. Is the error you are pointing out the only thing between this reference and the claim? Read the whole reference, not only the cited passage. If the teaching appears elsewhere, a citation-error argument buys a new action, not an allowance.
2. For §103, are you answering the combination and its rationale, or only the individual teachings?
3. Which statements in the draft characterize the claims, and is each of them needed to overcome the rejection?
4. Would an interview resolve the point faster than another written round (MPEP 713)?
5. What does the client prefer: speed and lower prosecution cost, or the narrowest possible record? Ask rather than assume.
6. Is the reasoning consistent with earlier actions in the file? A shift between actions is worth documenting, whoever or whatever produced it.

What the panel asked the office to do

• An opt-in pilot rather than a rollout at scale (Stoll, Tomsa): applicants who want speed can accept the trade-off, others keep the current process.
• Testing before rollout: run the tools on thousands of past cases with known outcomes and compare the results (Tomsa).
• Standardized prompts, so every examiner queries the model the same way and practitioners know what produced a rejection (Tomsa).
• Reviewed training data: a PTAB decision is not automatically a good decision, so practitioners should vet what a model is trained on (Tomsa).
• Practitioner input: the office does listen, and the panel urged practitioners to comment while the approach is still being set.

What stays the attorney's call

Stoll does not expect the backlog to disappear quickly: by his count, roughly 1.2 million applications still need an examiner's analysis, with or without new tools. Whatever the pace, how much to put on the record in a given response remains a judgment about a client's priorities, made by the practitioner who signs it.

This article covers prosecution practice and tooling. It is not legal advice.

Robert Stoll, President, Stoll Patent Consulting; former Commissioner for Patents, USPTO

Bernard Tomsa, Partner, Brooks Kushman

FX Leduc, CEO, DeepIP

Gene Quinn, Founder, IPWatchdog (moderator)