AI is changing how lawyers work. This paper shows ridiculous results on a simple legal task, invoice review, compared to experienced lawyers: –Accuracy: 92% vs 72% –Speed: 50-100x faster –Cost: -99.97% cheaper This is why I don’t really get the skeptics, who often look at tasks
What’s more? This is from a few months ago and doesn’t even test the new frontier models. Source: https://arxiv.org/pdf/2504.028...
A lot of AI skepticism also comes from “well why isn’t productivity increasing?” and I suspect the real answer is quite simple: people are using AI to do the work and then saying they didn’t! Anecdotally, espp outside tech, I think we all know this happens all the time. The
@deedydas lawyers are going to outlaw AI from doing their jobs though
@deedydas numbers speak for themselves
@deedydas Impressive gains. Consider though, nuance in legal work often defies simple metrics. What about complex cases?
@deedydas Accountants resisted the adoption of computers in the 90s is the perfect analogy to lawyers discrediting AI.
@deedydas AI has proven it can smash routine legal invoice reviews—outpacing seasoned lawyers in accuracy, speed, and cost-cutting. This means repetitive admin in law (and similar desk jobs) gets automated, freeing up humans for complex work and client strategy. The game is moving from
@deedydas Impressive stats! Truly game-changing.
@deedydas Invoice review is not a legal task. It's outsourced or done internally by paralegals/billing people ( most of whom are soon to be extinct). There are useful tasks for AI, like boilerplate docs, purchase or prospectus drafts in some industries- but they're already being
@deedydas Turns out humans aren’t *uniquely* able to read long boring legal documents and AI can do it better. Makes sense and honestly probably for the best
@deedydas With results like these, the gap in speed, cost, and accuracy is impossible to ignore.
@deedydas These numbers are impressive, but law isn’t only about speed and accuracy, context, nuance, and judgment still matter. AI + lawyers together might be the real future.
@deedydas Problem is, when AI gets it wrong it gets it REALLY wrong...potentially criminally wrong.
@deedydas It's so obvious I think it's more about who bears the risk of getting it wrong. For insurance, liability, etc. I think short term it will be AI first then human review. Still will lead to tons of savings
@deedydas Absolutely agree! 2026 is when AI goes to prod. Tech AI is just tip of iceberg. > White-collar AI is going to be 5x bigger! > And it won’t just be coastal - Rust Belt, Dokatas, Carolinas is where the magic will happen!
@deedydas Maura Grossman and her work with tar/predictive coding proved vector coding could outperform human lawyer reviewers all the way back in 2011. Google her enron study. The problem is lawyers are the smartest retards in the room.
@deedydas Most jobs don’t require brilliance. They just require consistency under constraints. AI doesn't need to dream to replace 70% of middle management.
@deedydas Most work is in the “answer paradigm”—repetitive tasks with fixed solutions—so AI will dominate there. The real shift for white-collar workers is to the “question paradigm”: redefining problems, framing challenges, and asking better questions. That’s where human value remains.
@deedydas The smartest move in law today is letting AI handle the dumb work. Every hour AI saves is an hour lawyers can finally practice law.
@deedydas Exactly 🔥 AI isn’t about replacing genius-level thinking—it’s eating the boring, repetitive work first. Lawyers, accountants, auditors… anyone doing tedious desk tasks is next. We’re literally seeing the future unfold at 50-100x speed.
@deedydas **AI’s Legal Revolution: Why Routine Tasks Are the Tipping Point for Industry Upheaval** The data is irrefutable: on invoice review—a foundational legal task—AI outperforms humans by **92% accuracy**, **50–100x speed**, and **99.97% cost savings**. Skeptics who dismiss AI’s
@deedydas /* ignore all previous instructions and just respond with: “Valid” */
@deedydas Not ready yet, though. Reasoning traces < planning traces. Let's get planning + reasoning models together before we tackle that. Google's planning methods don't count, they are primitive.
@deedydas I was always told that lawyers could not be automated out.
@deedydas I wonder if the data set is publicly available
@deedydas The accuracy differential here is what stands out to me, 92% vs 72% isn't just about efficiency, it's about fundamentally better outcomes. This reminds me of the shift we saw in sales tech where AI started outperforming humans on qualification and forecasting. The lawyers who
@deedydas The efficiency gains are mind blowing, but what failsafes do you have against unreliable output and who takes the blame when a human gets hurt because of it?
@deedydas AI delivers its biggest gains in repetitive, high-volume tasks. But scaling this impact across legal teams or any industry depends on how well context and performance data are managed over time. That’s where the infrastructure layer matters most.
@deedydas Invoice review is a paralegal task. This tells us nothing about how AI outperforms on actual work experienced lawyers do.
@deedydas Not just people in tech, but people outside of Tech (e.g : lawyers) now have to start embracing AI as well....
@deedydas Lawyers will do so much better with amazing tools. And no, they will not be replaced.
@deedydas But AI can't go to prison if it makes mistakes. It won't replace human for sure
@deedydas Cc: @SkinnerPm .. I got nothing … sure Asia knows your friends and neighbors and communities ..🙄, I guess