drew

Lawyer, hacker; building Cursor for lawyers

[drew @urandomd]

Winners in legal will be practice groups with well-structured data and purpose-built #AI-native startups with profitable flow.

Good taxonomies and labels over generational legal datasets at legacy firms fuels differentiated applications like the Kirkland & Ellis + Palantir Technologies agentic/ontological system. Agents working in free text in off-the-shelf legal products like Hargora tend towards distraction and hallucination, while structured ontologies from traditional Knowledge Management work provide guardrails and focus that maximize AI agents' utility. Practice groups will consolidate towards the firms with the most and best structured data as partners flock to the winners over rational fears of being left behind. Consider, for example, the result of an "Investor" ontology that allows an AI agent to look through SPVs to determine which sovereign wealth fund sits behind a particular transaction. Now compare that with an off-the-shelf agentic harness that stops at the surface. The first is powerful while the second is useless.

On the other hand, AI-native startups like Kyra Law, Norm Ai and Lawhive will attack from below with a clean slate to structure their pipelines using 2026 technologies. Their pipelines and user interfaces will be fine-tuned in real time on immediate feedback from lawyers without the drag of partnership dynamics and economics of traditional law firms.

Stagnating in the middle will be at least two groups. First, the legacy firms that can't utilize their data. Hargora is too blunt and generic of an instrument to provide differentiation. Perhaps necessary but not sufficient. Second, the startups that never get enough profitable deal flow to become relevant. Unit economics matter here. If a startup can leverage technologies to make, for example, NDA negotiation a profit center rather than a loss-leader, then that startup is a disruptor. Otherwise, its technology stack is just another sustaining innovation that legacy firms will deploy more effectively.

[drew @urandomd]

@ryanmckeen TLDR? Did it work out?

[drew @urandomd]

@tlbtlbtlb Strong argument against frequent scanning at the individual level, but there is real utility in measuring things at the population level. The additional information could help quantify the "harmlessness" of the assessment and lessen that knowledge burden. A virtuous circle.

[drew @urandomd]

Folks in legal love the initial interaction with CLI-equivalents like Cowork and Codex, but they always end up back in Word. It surprises me every day that A/ and OAI aren't going after the true competitor there.

[Justin Cromer @just_cromer]

@rfleury @rms80 @jonathan_wilke It’s dramtically easier to ship a CLI than a GUI for the same reason you report above. It is definitely an inferior interface, but has a much higher ROI with an extremely volatile product vs trying to ship a “decent” GUI.

[drew @urandomd]

@ZachAbramowitz The U is for unconditional surrender

[drew @urandomd]

@Andercot "hop on a call" is uniformly suggested by the person who values others' time the least

[drew @urandomd]

@Eli_Albrecht On the contrary, a soft "threat" can be a solid professional courtesy for regulated adversaries. I've had clients on the receiving end that avoided regulatory filings because the matters were settled before officially "threatened".

[drew @urandomd]

@gabepereyra @MollySOShea @nikogrupen I appreciate more and more the business model of selling a result instead of raw tokens. That gives you an incentive to innovate a layer above the models. The providers, on the other hand, are incentivized to black box the "reasoning" and send a big bill.

[drew @urandomd]

@paulg Seems like looking for the I/o bottlenecks and portability has shown historical innovation paths.

[drew @urandomd]

@OliverBCushing Only temporarily. There’s an information asymmetry that will work itself out in a few years as the basic tools become more widespread. Everyone know who the bad/slow paralegals in the firm are, for example, and they get routed around.

[drew @urandomd]

Daniel is dead on. The adoption path is using AI to automate the next layer up of boring / routine tasks while keeping the knowledge work human-centric.

It wasn't worth it in the past to make that document into a template and write a brittle script to update it once a year. Now, however, agents can take the first pass at a broad class of these types of verifiable tasks that brought no joy to any human doing them before.

[Daniel Lemire @lemire]

It is becoming clear that AI is powerful. However, most organizations will be unable to benefit from it for a long time.
AI can handle many boring or routine tasks:
• Check that the website has been updated with the latest procedure.
• Open this Word document, update the team leader’s name, regenerate the PDF, and post it on the internal site.
For every such task, organizations face serious roadblocks. Much of the work cannot yet be done by AI because the necessary tools do not exist in the organization.
Processes are often incredibly fuzzy and poorly documented. Organizations know more than they have written down. AI cannot easily figure out systems that live mainly in people’s heads.
Security and regulations are major hurdles too. A single constraint can quickly erase AI’s benefits. AI will never be perfect, so trust will remain an issue. Extensive human verification will be required—which is fair but can eliminate the gains in some cases.
According to a June 2025 Pew survey, half of U.S. adults are more concerned than excited about the increased use of AI in daily life. This suggests roadblocks will appear, often for political reasons.
We will get through it, but it will take a long time.

[drew @urandomd]

@zerohedge A lot of competitors coming into the space. Seems like every technology product company is becoming a consulting-led sales org now with the popularization of "forward deployments"

[drew @urandomd]

@BitGrateful Nothing sounds more awesome than running a low-margin law firm...

[drew @urandomd]

@ZachAbramowitz "You're right to push back! We should allow students to use AI"

[WillC @willchen500]

There is going to be an AI arms race among Biglaw firms, at least until the bubble pops. Each will be building their own ICBM or at least advertise that they are doing so. Just look at the amount of legal engineers and software engineers some of these places are hiring.

Don't be deluded - much of the motivation is simply client marketing rather than building any real tech.

Harvey and Legora were mortar launchers that served that purpose while being relatively cheap due to steep discounts funded by VC money. But sadly the industry perception is now that they are not real solutions.

[drew @urandomd]

"Our eng org is starting to spend as much on tokens as we do on headcount".

This is a problem for modern SAAS, especially LLM wrappers. They need to raise at ZIRP valuations to keep the party going, but they have high COGS (inference) and have also lost control of their technology stacks to the LLMs, so capex has become opex. They seem like they should be priced more like energy companies than technology today.

[Gabe Pereyra @gabepereyra]

We don’t plan to become a law firm or acquire one - we think the best business for Harvey is to help every law firm become AI-native not become one.

If Kirkland and every other top firm spend $500M on AI that means this market is going to massive and there will be plenty of room for OAI / ANT / Harvey / Legora / Mike to all be successful because there will be many more firms that can’t afford to spend $500M and will need similar platforms.

The part most people are missing is how expensive AI is about to get. Our eng org is starting to spend as much on tokens as we do on headcount. If the same happens for law firms there is going to be a massive opportunity to help firms manage and optimize that spend. We think there is value in building the agent infrastructure, all of the firm level management infrastructure (how do you manage your token spend / agents / data / lawyers by client, matter, practice area) and custom models to reduce this spend (you don’t have to train a model better than mythos - you just need open source models that can perform diligence better than mythos and our recent post training results show this is probably possible).

There is a world where Kirkland builds an awesome platform, we help them with some of it and we help a lot of other firms with it as well and so do all the other legal tech players and everyone wins without anyone needing to build a law firm

[drew @urandomd]

@mmjukic Objectivity? I love your work, but I think asking outside experts for opinions before forming your own does make some sense.

[Ryan Fleury @rfleury]

He’s making a differently pointed statement, but it applies; part of the human condition is craving what is not abundant. Unless you subscribe to the science fiction notion that automation will simply enclose the entire set of possible works—an absurdity—then it is clear that humans will flock to what is difficult, admirable, and beautiful for humans to do.