What happens when the patent examiner gets AI first?

IP Insights | Oct 5, 2026 | Written by: Tyler Teske

Patent AI has spent years helping people find prior art. The USPTO may now be testing something much more consequential: AI that helps interpret claims, map prior art to them, and develop the reasoning behind a rejection. For applicants, the question is who sees that reasoning first.

In Brief

  • The USPTO has used AI to help find prior art for years. A former Commissioner for Patents reports it may now be testing AI for claim construction, §102 mapping and §103 combinations: the reasoning behind an office action.
  • That reasoning is far harder than search. It means reading long technical documents, understanding the technology, and applying a legal standard, all at once, where one early mistake carries through to the conclusion.
  • The bigger opportunity is timing. Most applicants file first and learn later. Applying examiner-style analysis before filing reverses that: learn first, then decide what to file, while the claims and specification can still change.

The Next Wave of Patent AI

The first wave of patent AI was search.

Find the right documents faster. Rank the closest prior art. Route applications to the right technology group. Useful improvements, but fundamentally retrieval problems.

The next wave may be very different.

According to a former USPTO Commissioner for Patents who spoke with contacts inside the Office, the USPTO has been quietly testing an internal AI system with supervisory examiners for tasks including claim construction, §102 mapping, possible §103 combinations, §101 analysis and §112 review.

The USPTO has not publicly confirmed that deployment. But if the account is accurate, this is not simply a better patent search engine.

It is AI moving into the reasoning that produces the office action.

Search asks: What prior art might matter?

Examination asks something else entirely:

  • What does this claim actually mean?
  • Does this reference disclose every required limitation?
  • Can these references be combined to reject the claim?
  • What scope survives?

That distinction is enormous. A search engine retrieves evidence. An examiner uses that evidence to decide what happens to the claim.

If AI begins helping the USPTO with the second job, applicants have a new problem: the examiner may arrive with machine-assisted claim analysis before the applicant has ever seriously tested its own claims.

Patent AI is moving up the examination stack USPTO tasks stacked from retrieval to reasoning. Established: keyword search, AI similarity search, AI-assisted classification, automated search reports to applicants (pilot). Reported, not confirmed: summary of the invention, claim construction, §102 mapping, §103 combinations, §101 eligibility and §112 clarity analysis. Patent AI is moving up the examination stack USPTO tasks: publicly adopted vs. reportedly being tested Reported, not confirmed Established §101 eligibility and §112 clarity analysis §103 combinations §102 mapping Claim construction Summary of the invention Automated search reports to applicants (pilot) AI-assisted classification AI similarity search Keyword search Reasoning Retrieval
USPTO AI tasks · established vs. reported The lower tasks primarily retrieve and organize information. The upper tasks, reported but not confirmed, move closer to the substantive reasoning reflected in an office action.

In Plain English: What §102 and §103 Rejections Are

If you work in patents, skip ahead. If you don’t, these two terms are the heart of the story.

A patent’s claims are the numbered sentences at the end of the application. They define the legal boundary of what the patent protects. Each claim is made of required elements, often called limitations.

§102: anticipation. A claim is anticipated, and therefore not new, if a single earlier document discloses every limitation of the claim, arranged as the claim requires. The key words are every and arranged. A document that shows most of the elements, or shows all of them but in separate devices, does not anticipate.

§103: obviousness. A claim can be rejected even if no single document shows it, if the invention would have been obvious to a skilled person. Typically, the examiner argues that one document shows most of the claim, another shows the rest, and a skilled engineer would have had a reason to combine them, with a reasonable expectation that the combination would work. Finding the pieces is not enough; the reason to combine them has to be articulated.

What happens after a rejection. A rejection arrives in an office action. The applicant can argue the examiner is wrong, amend the claims (usually by narrowing them), request an interview, go another round, appeal, or abandon. What cannot change is the application itself: after filing, no new technical matter can be added, and every amendment and argument becomes part of the permanent file history.

Search Can Be Wrong. Examination Can Change the Patent.

Search is probabilistic retrieval. Examination is chained reasoning.

A search system can return twenty mediocre references and still be useful if one of them is important. The bad results get discarded, at the cost of a few minutes.

Examination is less forgiving. First the claim has to be interpreted correctly. Then each limitation has to be mapped to actual disclosure in the prior art, not merely to language that sounds similar. For §103, there also has to be a sound reason to combine references. Then all of it has to support a legal conclusion.

Get the claim construction wrong and the mapping can be wrong. Get the mapping wrong and the rejection can be wrong. Respond to the rejection by narrowing the claim, and the consequence may become permanent.

Put another way: retrieval tells you the pieces exist. Examination decides whether the pieces legally and technically fit together.

It is easier to see than to describe.

A Worked Example

The claim, references, page numbers and paragraph citations below are invented for illustration. They are simplified, but the reasoning is the same reasoning an examiner applies.

The Claim

Claim 1. A wearable device comprising:

  • [A] an optical sensor configured to collect physiological measurements from a user;
  • [B] a processor configured to derive a physiological metric from the measurements;
  • [C] a motion sensor; and
  • [D] wherein the processor modifies the physiological metric based on motion detected by the motion sensor.

§102: One Reference, Every Limitation

The examiner’s search surfaces Reference A, a 45-page published application for a fitness band. Here is what the mapping looks like:

The claim says Reference A says Why it’s a judgment call
[A] An optical sensor “a photoplethysmography (PPG) module with a green LED and photodiode” (p. 11) Different words. You have to know that a PPG module is an optical sensor.
[B] Derive a metric from the measurements “computes heart rate from the PPG waveform” (p. 19) A clear match.
[C] A motion sensor Shown mainly in Figure 6, labeled “3-axis accelerometer” (p. 27) It’s in a drawing, not the text. Easy to miss.
[D] Adjust the metric based on motion “accelerometer data may be used to compensate the calculated heart-rate signal during periods of movement” (p. 30) None of the claim’s words appear. You have to decide whether it means the same thing.

All four limitations appear in the same device, arranged as claimed. Claim 1 is likely anticipated. But notice what it took: reading 45 pages, recognizing a technical synonym, finding a limitation that lives mostly in a figure, and judging that a passage using entirely different words discloses the claimed relationship.

The Opposite Trap: The Right Words in the Wrong Place

Now suppose a different reference, Reference A′, contains the phrase “modify the heart-rate value based on detected motion,” nearly the claim’s exact words. But that passage describes a chest-strap ECG monitor in a separate embodiment, a device with no optical sensor at all.

Every word of limitation [D] appears somewhere in the document. It still does not anticipate, because the limitations are not disclosed together in one device, arranged as the claim requires. A system that matches keywords would report anticipation. It would be wrong.

Different words can still disclose a limitation. Identical words can still fail to. That is why mapping is a judgment about meaning, not a search for matching text.

§103: When the Pieces Are in Different Documents

Now suppose Reference A discloses [A], [B] and [C], but says nothing about adjusting the metric for motion. A search also surfaces Reference B, a paper on chest-strap ECG monitors that teaches correcting a heart-rate signal for motion artifacts.

[A] [B] [C] [D]
Reference A ✓ ✓ ✓ ✗
Reference B ✓

A search engine has done its job: every piece exists somewhere. That is not enough for §103. Examination has to answer a further question:

Would a person skilled in wearable sensors have had a reason to apply Reference B’s motion-correction technique to Reference A’s optical device, and would they reasonably have expected it to work?

There is a plausible argument yes: motion artifacts are a well-known problem for optical heart-rate sensors during exercise, and Reference B addresses that problem. There is also a plausible argument no: Reference B’s technique was built for an electrical ECG signal, and an optical signal is corrupted by motion in different ways, so a skilled engineer might not have expected it to transfer.

Which argument wins depends on the technology, the evidence and the legal standard. It cannot be read off a search result. And it is exactly the kind of judgment the reported USPTO system is said to assist with.

Why Knowing This Before Filing Changes the Patent

The value of examiner-style analysis is not simply that it may predict a rejection.

It is that before filing, the applicant can still act on what the analysis reveals.

Go back to the example. If that analysis arrives in an office action, the applicant is negotiating with a fixed application. If it arrives before filing, the same information becomes drafting guidance.

Suppose the analysis shows that the broadest version of the claim is already covered by prior art, but one particular technical feature appears to distinguish the invention: say, a motion-correction method designed specifically for optical signals. Before filing, counsel can restructure the independent claim around that feature, strengthen the specification describing it, add alternative embodiments, and build dependent claims around the most credible fallback positions.

Or suppose the opposite happens. The analysis finds no strong prior art against a feature the drafter had treated narrowly. Counsel now has evidence that the application may be leaving valuable scope unclaimed, and can test a broader claim before filing.

The analysis may also change a more basic decision: whether the invention is worth patenting at all, whether foreign filing is justified, whether the application deserves more investment, or whether the inventors need to do more work before anything is filed.

Once the application is filed, many of those choices shrink. The specification is fixed. New matter cannot be added. The inventors may have moved on to the next project. The initial claim architecture is already on the record.

That is the difference between learning from the art and reacting to it.

Two Failure Modes, Not One

The goal is not to eliminate rejection. The goal is to understand the boundary before the claim is committed. A claim can fail in two directions:

  • Too broad: it collides with the art, draws a rejection, and gets narrowed under pressure, late, on the examiner’s terms. Known before filing, counsel can narrow deliberately, around the feature that matters most.
  • Too narrow: it may sail through to allowance while leaving claim scope narrower than the prior art actually required. Nobody notices, because an easy grant looks like a success. Known before filing, counsel can claim more while there is still time.

Examiner-style analysis before filing helps counsel locate that boundary while there is still time to draft toward it.

Even the USPTO’s own public AI search pilot, ASAP!, was built on the value of earlier information: participating applicants received potential prior art before substantive examination, in time to amend before an examiner picked up the case. The approach we call Examination Upfront™ takes that idea further, from a list of references to the reasoning about them, and earlier still, before the application is filed.

Today, most applicants file first and learn later. Examination Upfront™ reverses that sequence: learn first, then decide what to file.

Why Practitioners Are Paying Attention

The account of the USPTO system comes from Robert Stoll, who spent 34 years at the USPTO and rose from examiner to Commissioner for Patents, speaking at a July 28, 2026 IPWatchdog webinar. He described “a somewhat soft rollout” to supervisory examiners in late 2025, and he was clear about the direction: “we’re going to move more and more and more into AI involved in the examination.”

The panel was not anti-AI, and Stoll did not treat every task the same way. He was comfortable with AI-assisted §102 mapping and §112 clarity checks (“I see no problems”), while his “significant concerns” centered on §103 combinations and §101 eligibility. But the practitioners on the call were already asking what this world looks like.

Stoll warned that once the tools tell an examiner a claim is obvious, “they get it into their mind that it is in fact a 103.” Bernard Tomsa, a partner at Brooks Kushman, echoed the point more bluntly: in the examiner’s hands, the output “becomes gospel.”

Tomsa also asked the question that should worry every applicant:

“If AI makes it a costless enterprise to continue rejections, what puts the brakes on that?”

And he named why it matters: “A patent is not a renewable right in the sense that if I lose it now to a bad decision, I can’t go get it again.”

The applicant-side conclusion came from FX Leduc, CEO of the AI patent assistant DeepIP, and it was unusually direct. Law firms and companies “can’t face the PTO without leveraging heavily these tools as well. You need to be armed with the same kind of weapons.”

That is the emerging asymmetry. The examiner may increasingly begin with a structured, machine-assisted view of the claims. Most applicants still learn how their claims read against the art when the first office action arrives, after every choice described above has narrowed.

It is worth being clear about where the panel’s argument ends and ours begins. The panel’s concerns were mostly about what happens after filing: examiners anchoring on machine output, cheaper repeated rejections, and lasting consequences in the prosecution record. The conclusion we draw is earlier in the process. If this kind of reasoning is coming to examination, applicants shouldn’t wait for the office action to see it.

We’ve Been Working on the Other Side of That Asymmetry

Much of commercial patent AI has concentrated on drafting and search.

At PatentVest, we have been working on the broader examination workflow that comes before and after search: understanding the invention, construing the claims, decomposing them into limitations and technical concepts, building search strategies, screening candidate prior art for relevance, mapping that art back to the claims, and evaluating likely §102 and §103 issues before filing. In other words, the work the example above walks through, done on a company’s own draft claims.

We call the approach Examination Upfront™.

The premise is simple. If an examiner is eventually going to subject the claims to this analysis anyway, the company should see as much of that evidence and reasoning as possible before it files, while it can still act on it.

The resulting work product is a Predictive Examination, reviewed by our IP analysts and attorneys for the company’s in-house counsel and patent team. PatentVest Inventa™ is the internal system we built to make that workflow practical at scale. The machine does the repetitive decomposition and mapping across long documents; people own the judgment calls the example shows are unavoidable, which is the model the webinar panel argued any serious use of these tools should follow.

Search is one stage. The decision comes from the full chain. Four stages in sequence. 1, Understand: invention analysis, claim construction, limitations and concepts. 2, Find, where much of patent AI focuses: search strategy, prior-art retrieval, relevance screening. 3, Reason: limitation mapping, §102 analysis, §103 analysis. 4, Decide: analyst review, attorney review, Predictive Examination. Search is one stage. The decision comes from the full chain. 1 · Understand 2 · Find 3 · Reason 4 · Decide Invention analysis Claim construction Limitations + concepts Search strategy Prior-art retrieval Relevance screening Limitation mapping §102 analysis §103 analysis Analyst review Attorney review Predictive Examination Where much of patent AI focuses
Examination Upfront™ · the full pre-filing chain Search is one stage of four. The value comes from understanding the claims before searching and reasoning over the art after.

Can the Reasoning Be Tested?

The panel’s own standard for these tools was the right one. Tomsa described it simply: run the system on cases already decided and ask, “Did it get it right in the past?”

So we did. PatentVest ran a retrospective test at scale. Within the tested candidate reference sets, PatentVest Inventa™ identified 88% of the examiner-cited references associated with §102 anticipation rejections. We will publish a study describing the methodology and results shortly.

After the example above, it is worth being concrete about what that means. For each examiner-cited reference, the system had to do what the §102 table shows: read a long technical document against the claim, find the passages that matter, recognize disclosures written in different words, and decide whether every limitation was present and arranged as claimed.

The examiner references were included in the candidate sets, so the experiment isolated the claim-analysis layer rather than independent search recall. Finding the art and reasoning correctly over the art are different technical problems. We wanted to isolate the second one and test whether it could be automated at scale.

The result is an early validation signal, not a universal accuracy claim. It does not measure search recall, false positives across every reference, or full prosecution outcomes. And the cases where the system and the examiner disagree matter as much as the matches: they are exactly the cases analysts and attorneys exist to examine, which is why every client result is reviewed by people before anyone relies on it.

Who Sees the Reasoning First?

The patent industry spent years using AI to find prior art. The USPTO may now be testing AI to help determine what that art means for your claims. Those are not the same problem.

Whether or not the reported internal system is ever officially confirmed, the direction is clear: patent examination is moving toward machine-assisted claim analysis.

That leaves applicants with a choice. They can meet that analysis for the first time in an office action, after the specification is fixed and the budget is spent, and react to it. Or they can apply disciplined, human-reviewed, examiner-style analysis to their own claims first, and learn from it while they can still change what they file.

The examiner may get AI first. The applicant doesn’t have to see the reasoning last.

About PatentVest

PatentVest is building its prosecution model around Examination Upfront™: applying claim-level scrutiny before filing, while the company still has options. For what this means for how a patent program is run, read The Most Important Patent Analysis Happens Too Late.

PatentVest offers portfolio screening as a service for in-house IP teams. Not every invention disclosure merits a patent filing, and an independent screening step lets in-house counsel decide, before any budget is committed, which inventions justify the investment.

To discuss portfolio screening for your invention pipeline, get in touch.

Sources

  • The AI-Enabled Patent Office: What the USPTO’s AI Strategy Means for High-Volume Prosecution, IPWatchdog webinar, July 28, 2026. Robert Stoll at about 04:00 to 07:15; Bernard Tomsa at about 08:50 to 11:00, 18:00 to 18:30 and 23:30 to 24:45; FX Leduc at about 16:00 to 16:20.
  • Automated Search Pilot Program (ASAP!), USPTO.
  • AI-assisted examination tool, USPTO.
  • 35 U.S.C. §§102 and 103; USPTO Manual of Patent Examining Procedure §2131 (anticipation) and §§2141–2143 (obviousness); Net MoneyIN, Inc. v. VeriSign, Inc., 545 F.3d 1359 (Fed. Cir. 2008) (elements must be arranged as in the claim); KSR Int’l Co. v. Teleflex Inc., 550 U.S. 398 (2007) (articulated reasoning required for obviousness).
  • Worked example: illustrative only; the claim and references are invented.
  • Validation figures: PatentVest internal retrospective study, September 2026.