On-premises for legal is not a good business - robertkarl.net

On-premises for legal is not a good business

May 25, 2026

AI disclaimer. The writing on this page is mine. The essay below, "Invalidating a business idea...." was originally written March 17, 2026. Opus gave me a few rounds of feedback on it. All the text emitted was from my keyboard though.

I tried to build a company called Curtilage. It was intended to be offline AI for legal.

As I reflect, I'm most impressed by the generosity of folks I spoke to.

I called lawyers and startup folks within my circle, and also people I don't know, warm intros, and said "Hey. I'm working on this idea. I don't have anything to sell. Will you talk to me about your AI tools? Will you talk to me about your process?" I expected people to be mean. It's uncomfortable to reach out, as an engineer and introvert.

And you know what? In every. single. case, people were generous with their time and willing to speak with me. My uncle actually warned me not to use this approach with lawyers. "What are you going to do," he said, "call up a lawyer and ask for free advice?" It took me a few seconds to get the joke; but then I had perhaps my loudest laugh of 2026.

All-in-all I spoke to about 15 lawyers and got universally negative or indifferent attitudes towards the idea.

So if you're one of the kind folks who spoke to me about this idea, thank you!

Anyways I killed the idea. It felt bad. It felt like a failure. Maybe for a week or so. I even printed business cards shortly before killing it. I never picked them up. But now with two months of distance, I view Curtilage as a roaring success. I felt vindicated when Claude released their legal package recently.

The opportunity cost has never been higher with how powerful the AI tools have become for technical founders. I invalidated an idea really quick. I'll say it again: killing Curtilage was a success!

I used AI heavily. I estimate it sped up my entire process 5-10x.

But the data points required for decision making were ultimately gathered in conversations with humans. I biased heavily towards spending my time talking to humans who could help me prove or disprove the idea. I used Claude for brainstorming, prototyping, market research, branding, outreach. I used it for idea generation, not delegating decision making. I will write more about my experience using AI in a startup context.

I'm not going to edit the following post much even though my opinions may have changed a bit in the intervening two months. What follows reflects my mid-March attitude, from being in the thick of it, working on this project all day every day for a month or so.

I think I missed two critical things: the moat aspect (Anthropic dropping Claude for legal nuked a lot of the value from orbit) and the team aspect. I'm not a lawyer; me as a solo founder in this space is preposterous.

The original piece post follows.

Invalidating a business idea in three weeks of market research

March 17, 2026

After using frontier models for coding, and setting up offline inference with open weights models at home, I decided to try to build a company around on-premises AI for regulated industries.

I wanted to start with a single use case that required offline inference, not just benefited from it. I started iterating on three things: 1) establishing some falsifiable criteria for the business. 2) doing deep market research on existing products and requirements and 3) building prototypes.

One court case, US v. Heppner, influenced me heavily. After reading it, I thought: lawyers must be itching to get some on-premises AI going.

Heppner, and Warner v. Gilbarco: two different takes on privilege

These two cases conflict on attorney-client privilege and cloud AI tools.

In Heppner, Judge Rakoff rules that using Claude waives privilege. Rakoff's reasoning: there are three criteria for identifying privileged information, and one of them is confidentiality. If Heppner sent the data in question to Claude for processing (and training, and sending to third parties, per the EULA), it's no longer confidential!

A month later in Warner v. Gilbarco, under the work product doctrine, client chats with Claude or ChatGPT can be protected and privileged. The Seed IP law firm has a good blog post about the issue. But basically, the law isn't solidified.

So what started as positive signal for the business idea became mixed at best, a few weeks later with Warner v. Gilbarco.

Three falsifiable hypotheses

1) Law firms care deeply about data privacy. They can't share it with third parties due to legal obligations. For particularly sensitive data, cloud API use is off the table.<br>2) There is wasteful, repetitive work that can be automated to save time and money for businesses.<br>3) Offline AI solutions are good enough to solve 1) and 2) -- that is, performing the useful tasks while the data remains entirely offline.

The first was the crux. But all three need to hold for a healthy...