When handling thousands of cases per yr with a small inhouse team, my primary strategy for ensuring patent quality at scale was to pick trusted outside counsel. I had to trust they focused on issues that I cared about, and I'd spot-check the work. (I don’t think this changes with AI. Spot-checking just becomes checking every draft.) In any event, I always wondered: what problems were still making it into patents when nobody was looking? [numbers below...]

To test OC, I’d ask for patent drafts of a technically simply case and the most technically difficult case they'd handled.

For the simple case, I’d look for extra embodiments. That showed technical creativity and whether they were thinking beyond the disclosures.

For the difficult case, I’d look for how well they explained the concepts: the level of detail, clarity of thought, and ease of reading. I also wanted to see what they considered technically difficult, since this would indicate the range of technology they had handled before.

For both cases, I’d look for clear and concise claiming, lack of divided infringement issues, and detailed support for claim terms.

If an OC did well in all of those areas, I’d invite them to a disclosure meeting to see how they interacted with inventors. Were the questions thoughtful and improve the patent draft? Did they make the process easy for inventors?

Once trust was established, I’d spot-check the work. But even after I trusted OC, I wondered how often problems were still making it into patents. After all, nobody’s perfect and there was no AI to help at the time.

Well, we ran a test! We looked at 240k patents granted in the last 12 mo and benchmarked them against 10k litigated patents. We excluded chemistry, pharma, and biotech because some of the flags we tested for are standard practice there.

We found 1 in 8 granted patents had at least one flag. The biggest issue by far was narrowness. This may not be anyone’s fault in particular, since 101 and prior art may affect certain tech areas more than others. But the other issues I worried about are more likely within OC’s direct control like divided infringement and lack of explicit support. These happened in <3% of patents.

Company differences were huge: among large portfolios, flag rates varied by 28x.

The outside counsel you pick matters a lot.

My one takeaway for inhouse counsel is: if you only review one thing, review claim length, specifically how many words are added during prosecution.

Congrats to NVIDIA for showing what disciplined patent drafting looks like at scale: high volume, short claims, near-zero defect flags, and one of the cleanest profiles in the dataset.

Report here