We were a development team of three, and between design, testing, and external work, all three of us were stretched thin.
In the middle of all that, a request came down from the patent team to review several hundred competitor patents. The job was a first pass, separating those with potential for infringement from those without — a stage the standard process had set aside two months for. It wasn’t even our job, and none of the three of us were patent experts. It might seem you could just ask AI, “Does this infringe or not?” But that wasn’t an option. If you hand over the judgment and it’s wrong, it’s the company that ends up with the lawsuit.
The starting point was a single legal principle: the all-elements rule. If a competitor’s claim consists of A, B, C, and D, and our product has no D, it isn’t infringement. Miss even one, and infringement doesn’t hold. That became a clear ‘erasing rule’ we could give AI. At first we tried filtering by keyword, but not everything that mentioned “battery” was about our battery. So we asked separately whether the method was the same and whether the purpose was the same, and kept only what passed both. That made an irrelevant 40% disappear.
- First-pass filter
- Separate checks on method and purpose · about 40% irrelevant, excluded
- Risk sorting
- Element mapping · about 5% high-risk, isolated
- The human share
- Three people · confirmed the high-risk group, split it off for second-stage detailed review · one day
For the patents that remained, we held their elements up against our product one by one. We sorted them from non-infringing through low, medium, and high risk, and isolated the high-risk 5%. But this was strictly a first-pass screen. The detailed review, comparing claims line by line, comes after that, and it’s still human work. What I was most careful about was this: if AI wrongly erased something because “this element is clearly different,” a truly infringing patent would get pushed into the 40% pile screened out in the first pass and drop off the priority list for follow-up review. So keeping the last 5% in human hands wasn’t laziness. It was a safety mechanism.
AI erased, and people kept the judgment. We didn’t hand it over — we divided the work.
The lever this time wasn’t “AI judged the patents for us.” Because AI cleared away 40%, the three of us could focus squarely on the truly risky 5%. At least in this case, having AI erase what was “clearly not a match” was safer than asking it for the answer. Erasing can be done by rule; judgment is hard to hand over. And even that only worked because we had a clear rule: “Miss even one and it’s out.” This division of labor, not asking AI for the answer but having it erase what’s clearly not a match first, cut a two-month review to a single day, and it’s now my default. I’m still looking for rules like that in other kinds of work.
A researcher on a team that takes technology into mass production. Working where a handful of people weather a flood of requirements, I’m trying to set AI up not as a search tool but as hands and feet. I write more about what didn’t work than what did, and more about process than conclusions.