.jpg)
Every conversation about AI in patent work eventually arrives at the same question - the elephant in the room. At a conference this spring, an audience member said it out loud, while a panelist was mid-sentence:
“If AI is doing the searching, what are we actually here for?”
Without skipping a beat, the panelist answered: the steps you can write out yourself are the ones AI should be doing. The hard part was never the steps - it's knowing which ones matter. That's still the expert's job.
It's a clean answer, and a true one. But it exposes a deeper rift between the patent software users (like the individual asking the question) and the patent software purchasers (like the panelist). Ask the people doing the searching and they'll tell you they're worried about what's left for them. Ask leadership and you'll often find the opposite problem - they've already decided AI can do more than it actually can.
Neither group is wrong to worry, exactly. They're just answering a different question than the one the panelist did. "Human in the loop" isn't a line for a proposal - it's the difference between a workflow that functions and one that doesn't. The expert isn't disappearing from the work; the work is changing shape around them, and figuring out which steps still need a human is itself the job now. In an industry where a missed reference has real consequences, most firms still treat that as a talking point instead of the job to be done in the first place.
The more useful question isn't whether AI replaces the expert. It's what experts can now do that wasn't viable before - faster, at a higher standard, with more of the work they used to have to skip.
That's the part most of the industry hasn't figured out yet.
Patent search has historically been an incredibly manual process dependent on individual judgment at almost every step. A request would come in from R&D or an engineering team, a searcher would build a string of keywords, synonyms, and connectors, run it in a search engine, and get back thousands of results. They might filter by classification code, date range, and field, and then their expert judgement would come in - reading claims, interpreting scope, screening forwards and backwards citations one at a time until they have a manageable set, maybe 50 documents, to hand to the attorneys building the opinion. Then they move to the next request and do it all again.
There’s genuine value in that work. But the overwhelming majority of a searcher’s time is spent on execution: building, running, filtering, screening, not on the interpretation that the whole exercise exists to produce. The profession has been spending its most expensive resource, expert judgment, on the parts of the job that need it least.
When people hear “AI is changing patent search,” most assume that means the role is being replaced. The reality is almost the opposite.
AI is taking over the manual-heavy middle section of the workflow: generating input, running searches across millions of records in minutes, applying classification and semantic filters. That’s the part that used to eat up the majority of the time. But handling that portion of the workflow is not the same as doing the job.
The parts that require human judgment still belong to the professional - understanding what's being asked, and what's at stake behind the request. Interpreting what the results mean, and the business implications that follow, is the "why," not the "how." That's still a distinctly human skill.
AI-powered search was never meant to automate the bottom of the value chain and stop there. It's meant to move the practitioner up it. When AI absorbs the execution, the practitioner's time goes back to the work only they can do. More than that, it opens up work that wasn't possible at scale before AI existed.
An AI search surfaces a result with a 97% relevance score. On paper it’s a strong match: the same core approach, overlapping terminology, the right timeframe. The system flags it with high confidence, recommended for citation and flagged for attorney review.
A quick surface check confirms the basics. Right technology space, right timeframe, plausible assignee. The AI did its job, and it did it in seconds. Is this novelty destroying or a red herring?
Enter: human judgement.
Read the reference closely and the match falls apart. It discloses the same general approach, but only for a narrower configuration than the one the subject application actually claims, and the specific limitation those claims turn on isn’t taught anywhere in it. Close on the surface, but it doesn’t read on the invention itself. It doesn’t anticipate. And the more valuable read is the one sitting just past that correction: if the closest prior art keeps clustering around that other configuration, the claimed space looks relatively open. That’s not a search result anymore. That’s a patentability signal worth putting in front of the attorney before the prosecution strategy is locked.
AI surfaced the document correctly and quickly - it did its job. But it required a discerning practitioner to turn it into something the client can act on.
It isn’t a checklist to run through. It’s a mental structure that works across three layers:
Surface check. Does this even make sense: right space, right timeframe, right assignee? For an experienced searcher this is fast and close to automatic. If something’s off here, the AI missed on the basics.
Claim-level scrutiny. What does the reference actually disclose? Species distinctions, sequence specificity, method versus composition, priority date nuance. The AI reads the words. You understand what they mean.
Strategic read. What does the result set say about the landscape: assignee clusters, timing gaps, prosecution patterns, and what should be done about it? This is the layer that turns a search into a recommendation.
When measuring what AI is really doing for this profession, NLPatent keeps coming back to three layers: Quantity, Quality, and Quantum.
Quantity and Quality are table stakes. More searches, faster, more consistently: that’s real, but it’s surface level , and before long every serious team will have it. Quantum is the part that matters: the things you can do now that simply weren’t possible before. A few that stand out:
Reading a whole landscape, not a result set. Surfacing the strategic signal across millions of records: assignee clusters, white space, timing gaps, instead of handing back a stack of documents and hoping someone connects the dots.
Turning every search into a recommendation. Not “here’s what matched,” but “here’s what this means for the matter and here’s what to do about it.” That used to be a luxury reserved for the most important projects. Now it can be the default.
Spending judgment where it counts. When execution can be condensed to minutes rather than hours, the bulk of time goes to interpretation: the part of the job that actually moves a prosecution strategy.
That’s not the same job done faster. That’s a different and better deliverable.
The mistake that comes up most often isn’t a technical one - it’s a mindset one. Some practitioners disengage because they’ve decided AI is on a clear path to making them redundant, so they are skeptical and reluctant to adopt. Others over-trust it and assume the machine handled all three layers. Both represent the same error wearing different clothes - treating AI as a replacement for judgment instead of an input to it.
There’s a question underneath all of this that the industry isn’t spending nearly enough time on. The whole argument this piece rests on is expert judgment: the analytical scrutiny, the strategic read, the surface check that’s “fast and close to automatic” for someone experienced. But that instinct isn’t innate. It got built over years of doing the manual, sequential, tedious work that AI is now absorbing.
So now that the manual process is all but replaced with AI, how is that discerning judgement developed in trainees?
Significant energy has been spent on what AI can do, and almost none on how to transfer that knowledge to the next generation. With people entering a field where AI is already embedded in the workflow, a new searcher who has never built the strings by hand, never screened records one at a time, never sat with a bad result long enough to understand why it’s bad: how do they learn what good looks like? And if they can’t recognize good, what exactly are they supposed to be verifying when the output lands on their desk?
That surface check that feels automatic to a veteran is automatic precisely because it was earned the slow way. Take away the slow way and you don’t get a faster expert. You get someone who can run the tool but can’t catch it when it’s wrong, which, in a profession where errors carry real weight, is the one thing the human in the loop is there to do.
Most organizations are treating this as a training problem. It’s actually a succession planning problem. Mentorship is part of the answer, not all of it. The teams who build judgment into the next generation now, deliberately, while everyone else assumes the tool will handle it, are the ones who will still have experts worth pairing with AI in ten years. That’s not a defensive move. That’s an edge.
Getting there is complicated, and this is the part that’s easy to skip past. AI doesn’t move anyone up the value chain on its own. It requires strategically adapting it to the workflow: deciding which steps to hand off, where to insert review, what “good” looks like at each layer. That’s its own kind of expertise. The teams pulling ahead aren’t the ones who bought the best tool. They’re the ones who redesigned how they work around it.
Interpretation and judgment can’t be automated, and human eyes still have to verify the output, especially in a legal context where errors carry real weight. But “human in the loop” was never meant to be the ceiling on AI’s value. Used well, It helps us go beyond it. The professionals who make this shift, from executor to strategist, from search vendor to IP advisor, aren’t being replaced. They’re delivering a level of work the role couldn’t reach before.
The role isn’t disappearing. It’s being redefined, and the only real question is whether you redefine it on your terms, or let someone else do it for you.
Sign up to receive alerts when we publish more blog posts about how NLPatent has brought success to the field of IP and AI.
.svg)
.png)
.svg){kind=small}
