By Michael S. Borella –

The U.S. Patent and Trademark Office (USPTO) is trapped in a perpetual battle on two fronts.  First, there is their application backlog, which can extend patent pendency by months or years.  Second is the systemic challenge of patent quality.  These two problems are not independent of one another. Low-quality examination is characterized by incomplete searches and rushed analysis, which can lead to rounds of Requests for Continued Examination (RCEs), appeals, and even costly post-grant litigation.  Conversely, under pressure to reduce pendency, examiners may conduct less-thorough examinations, thereby harming quality.

For decades, the solutions have been incremental, bureaucratic, or both.  Examiner interviews, accelerated examination programs, and various pilot programs have nibbled at the edges of the problem.  The most recent, the “Streamlined Claim Set Pilot Program,” continues this tradition by offering a simple, but flawed bargain: trade your claims for a spot at the front of the examination queue.

These solutions misdiagnose the illness.  The problem is not the number of claims; instead, it is the profound inefficiency of the first substantive interaction between the applicant and the examiner.  A first non-final office action is often the output of many examiner hours being spent on cold-start prior art searches. In many cases, this office action contains prior art rejections that could have been avoided.  If the applicant had known of the closest prior art, they likely would have narrowed their claims to focus examination on more protectable aspects of the invention.

With the rise of generative artificial intelligence (AI), there are new options on the table.  One that we’d like to propose for discussion is what we call AI-First Triage (AIFT).  This change would fundamentally restructure the examination front-end to accelerate prosecution, improve patent quality, and create a more collaborative, efficient process for applicants and examiners alike.  Particularly, AIFT is designed to front-load information sharing between the USPTO and the applicant and incentivize early, substantive amendments where appropriate.  The process would be simple, universal, and mandatory, updating the current examination procedure.  Further, it would be provided at no charge for all new applications.

The main principles of AIFT would be as follows.

• AI Search (The “Zeroeth” Office Action): Upon filing and completion of formalities, every new patent application would be processed by an in-house, private generative AI search engine. This tool would analyze the specification, drawings, and, most critically, the as-filed claims.
• The AI Advisory Action: Within a few weeks of filing, the applicant would receive an initial AI advisory action.  This office action would not contain statutory rejections.  Instead, it would be an AI-generated search report.  The report would contain the results of the AI search, mapping the most relevant prior art references found by the AI search to specific claim elements.  For example, the report might include a claim chart mapping each claim element to a specific location in a prior art document where the AI search engine contends that it found disclosure of the element.  Notably, the AI search would only identify prior art and would not provide any legal reasoning.  In other words, the AI search would not provide a § 101 or § 112 analysis, nor would it provide a motivation to combine should it use multiple references against a claim.
• The Applicant’s Choice: The applicant would then have a two-month, non-extendable window to respond to this AI advisory action.  There would be two simple options.  The first would be to file a preliminary amendment that makes a substantive narrowing amendment to at least one independent claim in view of the AI search.  The applicant would have the option of including remarks explaining why the narrowing amendment results in claims that the cited art does on read on.  The second would be to take no action and not file a response. In the latter case, the applicant can still file a preliminary amendment at any time before the first action on the merits.
• The Examination Queue: If the applicant chooses the first option, the application is placed into a prioritized queue for examination by a human patent examiner.  If the applicant chooses the second option, the application is placed in the standard (slower) examination queue.
• The Non-Final Action: Regardless of which option the applicant chooses, the next office action is issued by a human examiner and is a non-final office action.  The examiner is free to use or discard the AI search results, conduct their own search, and formulate their own rejections.  This preserves the applicant’s established right to respond to a complete, human-led examination on the merits.

AIFT addresses a bottleneck in the examination process by speeding up the prior art search.  This creates a virtuous incentive for applicants to engage with prior art and improve patent quality from the start.  This up-front sharing of information and incentive for early amendment is likely to more quickly to disposal of an application (i.e., allowance or abandonment) by identifying prior art before the examiner lifts a finger.  Additionally, we expect that such a procedure would reduce the rate of RCEs, which are one of the greatest contributors to pendency, making the entire examination process more efficient.

AIFT can also improve patent quality.  First, it would augment examiners by providing an AI search across extensive databases.  Examiners could then spend less time looking for needles in haystacks and more time conducting an analysis of the prior art.  Second, by presenting applicants with relevant prior art early, applicants are encouraged to amend broad or vague claims early in the process.  Third, the initial AI advisory action could act as a filter, prompting applicants to abandon clearly anticipated or clearly obvious inventions early, freeing up examiner resources for other applications.

Despite AIFT having advantages, there are several concerns that may hinder its deployment.  Applicants would face a significant dilemma if presented with low-quality or irrelevant AI-cited prior art, forcing them to either make an unnecessary amendment to gain docket priority or be penalized with a longer wait time for challenging the AI’s findings.  Thus, the quality of the AI search would be key and some degree of human oversight may be needed.  Further, filing a preliminary amendment in response to an AI advisory action would create prosecution history estoppel before the applicant ever interacts with a human examiner.  This may incentivize some applicants to ignore the AI search results.  Moreover, some applicants might strategically file overly broad initial claims specifically so they can concede a narrowing amendment in response to the AI search.  This gaming of the system would result in a faster examination without having to file a focused claim set up front, effectively turning the collaborative step into a mere procedural hurdle to be cleared for a docketing advantage.

While potentially problematic, each of these issues could be mitigated.  AI is a learning technology and the USPTO’s AI search model could be updated with examples of when a human examiner successfully used the AI-cited prior art to motivate an applicant to amend their claims, as well as when the human examiner overrode the AI-cited prior art and used different art to reject the claims.  Applicants could mitigate the estopped issue by filing a response with minimal or no substantive remarks, relying on the claim amendments to speak for themselves.  Also, Applicants could make it clear on the record that the amendments were made without prejudice and merely to expedite prosecution.  Finally, if necessary, applicants can be disincentivized from gaming the system by modifying the procedure as follows.  The first human-produced office action can be final when the applicant does not amend the claims in response to the AI advisory action and the human examiner’s prior art rejections are based entirely and solely on the prior art references that were previously cited in the AI advisory action (if the human examiner adds any new prior art to the rejections, their action must be non-final).  It is recommended that the human examiner would need the head of their technical center to approve such a first-action final.

The USPTO cannot solve its dual pendency and quality crisis by merely repackaging old ideas or coming up with another pay-to-play docket manipulation approach that fails to improve search quality, claim clarity, or examiner decision making.  The proposed AIFT, in contrast, is a re-imagining of the examination process that places a powerful information technology tool at the very beginning, creating a collaborative triage step to help applicants produce better claims for human examination, and to help examiners conduct better searches. USPTO should stop tinkering with the queue and instead innovate at the core of the examination process itself.