
28 | legal engineer @mercadolibre
Santa Fe, Argentina
Joined December 2012
- Tweets16.5K
- Following758
- Followers2.9K
- Likes24.4K
Hernán retweeted
Building AI tools isn't practicing law. My students know that. But they now have a skill most lawyers don't.
Tomorrow is the last day of my Fall law school class. Six teams built six working prototypes with partners at Ice Miller:
Here's what they learned, in order.
Thinking tools came first. Before touching a model, they decoded the work.
They lived a real example first; organizing themselves to tackle a mock client NDA request. This made them focus not just on the legal elements of the NDA itself but how a request comes in, how to organize to attack it, how to orchestrate the work, what done looks like. This is fundamental as most people are focusing on just AI and the legal work.
I taught them SIPOC (look it up) to map who supplies each step, what comes out of it, and who receives it. It's a really helpful lens to decode work. But we did not rely on one view of the work. Rather there are always three versions of every process: how it’s written down, how people describe it, and how it actually runs. In order to successfully integrate AI, you need to find out how it actually gets done.
Then I gave them Jobs to be done so they learned to look beyond just the functional impact of AI (typically people are thinking speed) but value means more than speed: what the lawyer needs done, how they want to look to their client, and how they want to feel do this work. These matter in terms of adoption and lawyer-self worth.
Then planning tools.
Before anyone built anything, the workflow had to pass a gate: a clear trigger, a clear deliverable, it repeats, and it includes at least one judgment call and one handoff. Bigger than a single task, smaller than a whole matter. Most AI projects fail right here, because they aim at something too vague to build or too small to matter.
Then they mapped the future state, which I called the ingredient list and the recipe: every step, whether a person or the tool does it, what goes in, what comes out, and what it needs. And every step has a customer, not just the last one. The associate is the paralegal’s customer, the partner is the associate’s, and the client is the partner’s.
Finally, a build-or-buy check: could the firm simply buy this? If so, what makes building it worth the effort? Usually the answer is personalization as in embedded a lawyer’s own judgment, behavior, practice reality, and idiosyncrasies.
Then building tools.
I had them split their instructions into separate files: who it’s for, what it reads, how it decides and what it hands back. When something goes wrong, you can fix one part without breaking the rest, and you can explain exactly what the tool was told to do. That matters when a lawyer must stand behind the output. I call this “defensibility and explainability” – which is being required by more and more clients and jurisdictions.
They also learned the difference between an instrument, which does what you tell it and an agent, which decides what to do next.
Their first real build was an interviewer, an AI that draws out how a lawyer actually works. They used it on themselves first, then on their partners.
They built in Claude, Copilot, ChatGPT and Gemini, and a few went further with Claude Code and LangGraph. When a platform’s limits got in the way, they switched. Choosing the tool is a judgment call too.
Then testing.
Every pod learned what evals are and how to build them into the workflow (few are doing this out here). They wrote down what good looks like before building, then wrote test cases designed to make the tool fail: an easy one, a hard one, the wrong input, and one where the tool should stop and hand the work back to the lawyer. If you can’t say what good looks like, you can’t tell when the tool is confidently wrong.
I told them the same applies to their own use of AI: before you ask it to draft a memo, know what a great memo looks like.
The six working prototypes:
-a market-terms coverage desk
-a contract tracker
-time capture that knows each client's billing rules
- practice interviews with simulated witnesses
- a matter timeline that remembers why each decision was made
- an AI adoption interviewer
That's the edge they leave with. Anyone can learn the tools. Few lawyers know how to take their own work apart and rebuild it.
AI-native starts with knowing this work. And my students just got a huge jump start.
Now I hope they follow my advice and keep staying up to speed with the developments in AI after this class finishes so they don’t fall behind and lose their edge.
Hernán retweeted
One of the most iconic rallies is coming back! 🇦🇷
See you in Argentina in 2028! ❤️
#WRC | #RallyArgentina
Hernán retweeted
RIP interbanking Argy
En teoría son 804k cuentas de banco. Ante la duda cambiaría la clave de todas.
Hernán retweeted
Jevons Paradox seems especially obvious in the field of law.
If you make lawyering easier, that produces more lawsuits and other legal activities, which means you need more lawyers.
a short history of software replacing lawyers:
1973: lexis puts case law on a computer, and the 70s become the fastest decade of lawyer growth on record.
2011: headlines say e-discovery will replace armies of lawyers, and the US goes from 1.2 million of them to 1.37 million.
2025: donotpay, the "world's first robot lawyer," settles with the FTC, whose complaint says it never had lawyers test its work.
2026: AI can draft the contract, check the cites and scan every new rule in every state.
if the last 50 years are any guide, a lot of what it makes cheap was never getting done in the first place.
Hernán retweeted
I used to be a language lawyer. In the late 80s I _studied_ the ARM (Annotated Reference Manual for C++). I knew the language cold, inside and out. I could make it do backflips. I shipped a lot of code using it. A LOT.
I’ve attacked other languages with a similar intensity and depth. Until recently, that was simply the only way to become adept at building software systems.
I now consider the deep understanding of a language to be a supreme waste of time and effort. I don’t want to learn Rust, or Golang, or Typescript, or any of the thousands of other languages out there. I’ll leave that to my agents.
What I need now is to understand the high level gestalt of the system. I want to see the systems and subsystems. I don’t need to inspect the cogs. If I truly need to understand a detail at the cog level, I’ll interrogate my agents. I’m not going to pretend to be a language expert anymore.
Hernán retweeted
BREAKING:
This morning’s cyberattack against supermarket fridges and freezers has now spread to Belgium, Sweden, Luxembourg and Estonia.
The hackers increased the temperature up to 50C. Food worth millions of euros has been spoiled.
Experts warn that such cyberattacks on supply chains and logistics can constitute a threat to national food security in many Western states.
Hernán retweeted
Muchas municipalidades del país "hackeadas", hace 3 días se robaron el dinero de todos los sueldos municipales en Santa Isabel, La Pampa, la provincia de Formosa hackeada.
No se ve reacción para lo que se viene. Esto recién empieza.
Never let Claude run wild on the privacy policy or legal pages on your website.
Opus 5.5, and other Claude models in general, are so knee deep into moral policing it can genuinely land you in hot water.
> it will get way too specific about how your work happens internally
> it will add promises and guarantees you didn’t want enforced
> it went as far as adding my name in the policy page
> it listed out all the third party tools that we have integrated too, basically showing our entire stack
It is built to write from the customer side by default, it said so when I pressed it. Just be careful.
Hernán retweeted
BREAKING:
The town hall of Limoges, built in 1873, didn’t survive today’s riots.
The building’s spire has now collapsed
Hernán retweeted
We always wanted more access to justice, more access to government.
My paper, Judicial Economy in the Age of AI, is about the impending strain on government resources, specifically courts, created by AI tools.
It predicts that courts will have to put up barriers to weather the AI storm, and that these barriers will harm individual rights.
We now have evidence of huge pressure building up on courts and government services.
There is a solution. You may not like it, but you will surely hate the alternative. We need to integrate AI quickly into our courts and agencies.
I’ve been talking to court administrators, judges, and others about solutions. If you are working in this space, I’m happy to talk more about how to build these new institutions.
Hernán retweeted
Replying to @mmbarrionuevo
Martín, si te interesa el tema armé un monitor visual, simplón pero funcional, de la cuenca del plata cuenca-del-plata.vercel.app que basicamente permite seguir cómo evoluciona la crecida de norte a sur, compararla con umbrales de alerta y entender los riesgos reales para diferentes localidades. Se actualiza cada 6hs con datos oficiales.
Hernán retweeted
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
Hernán retweeted
There is an optimal amount of courage x intelligence. You can't be so intelligent that it is worth it for you to spend time thinking. You need to be just a little retarded. And have unrealistic expectations of your ability. You also need to know that you're retarded, so you use simple stupid tricks to meet the expectations of your abilities































