Human-in-the-loop by design

In a law firm, "the AI got it mostly right" is not good enough. The bar is traceability and accountability. The trick to automating legal work isn't removing the human — it's placing them exactly where they add the most value.

Patent attorneys spend enormous effort on two things AI is genuinely good at helping with: drafting first versions of applications, and searching prior art. But you can't hand either over blindly — a missed reference or a sloppy claim has real legal consequences. The answer isn't full automation or none. It's a system that does the heavy lifting and routes the judgment calls to a human.

The dial that matters: confidence threshold

Every AI output comes with a confidence. The core design decision is where you set the bar: above it, the system proceeds automatically; below it, a human reviews. Set it too high and you save no time. Too low and you let risky work slip through. Drag the threshold and watch how a batch of 20 items routes — and where the danger is.

Where the human belongs

Human-in-the-loop done well isn't "a person checks everything" — that just moves the bottleneck. It's a deliberate division of labor:

• The system drafts and searches — producing a first application and a ranked set of prior-art references, with citations.
• The attorney reviews and decides — spending their expertise on judgment, not on typing first drafts from scratch.
• Everything is traceable — every claim links to its source and every decision is logged, so the work stands up to scrutiny.

What it buys you

The result isn't "attorneys replaced." It's attorneys doing far less low-value work, with a complete audit trail. Here's roughly where a patent workflow's time goes, before and after.

More insights