Archives

The most subtle failure of AI document review isn't that it gets facts wrong. It's that it builds its entire analysis on the wrong question, because the right question was never in the documents.

When people talk about AI replacing lawyers, they imagine the disruption hitting big-firm partners. The reality is much sadder and much more advanced: a generation of solo and small-firm lawyers are quietly running out of work, and almost nobody is writing about it.

Before AI, clients sought legal advice and accepted whatever advice they got. Now they shop. They ask three friends, two YouTube videos, and ChatGPT until something tells them what they wanted to hear. Then they bring that answer to me.

Inside any law firm, there are two parallel conversations about AI: the official one, conducted at partner meetings, and the real one, conducted privately. The gap between them is the most expensive thing the legal industry isn't talking about.

After a year of testing both, I still cannot definitively answer the most basic question in legal AI: should lawyers use general-purpose tools like Claude, or specialized legal AI SaaS like Harvey? Here's what I've learned, and where I'm still uncertain.

Every meaningful conversation with a client used to be between two people. Now there's a third participant in the room: the AI the client has already consulted. Lawyers who don't learn to manage this third presence will lose clients to those who do.

AI hallucinations have stopped being a curiosity. They have started showing up in court orders, government notices, and official publications—the documents we used to be able to trust as ground truth. The damage is harder to undo than people realize.

A client walked into my office last month carrying an AI-generated legal opinion that told him he didn't owe ten years of unpaid capital contributions. The AI was confident, structured, and completely wrong. Here's what happened next.

Founders are taught that going first is rewarded. In the AI era, the opposite is becoming true: early innovators do the market validation that platforms then absorb, often for free. By the time a big platform announces a feature, the pioneers who proved it works are already obsolete.

Most lawyers use AI the way a junior associate uses a senior partner: they ask, they accept, they implement. That posture is wrong. The lawyer who uses AI well treats it like opposing counsel—as something to interrogate, contradict, and beat into shape.

Vertical legal AI companies have built their pitch around three competitive moats: industry expertise, information asymmetry, and training data. After reading their materials like a due diligence report, I don't think any of these moats actually exist.

The most dangerous thing AI has done to legal work isn't producing bad documents. It's producing documents that are good enough to look great to someone who doesn't know what great looks like.

AI can produce a legal analysis in twenty seconds. It cannot produce the feeling of having been correctly held through an uncertain moment by someone who actually knows what they're doing. That feeling is what clients are buying.

Every article about Anthropic's Claude For Legal led with the same number: twelve plugins. The plugins are real. But the lawyers paying attention should be looking at a different number: the twenty connectors that ship alongside them. Here's why.

I scheduled an article for June 8th admitting I couldn't decide whether lawyers should use general-purpose AI or vertical legal AI. Three days ago, Anthropic launched Claude For Legal. Here's how my view has moved—and why I should have seen it coming.

Most analysis of AI in law focuses on lawyers. But the judiciary is quietly adopting these tools too—and the consequences for how cases get decided are larger than anyone is publicly acknowledging.

The dominant narrative says lawyers should embrace AI to stay relevant. But there is a quieter, uglier pattern emerging at law firms: young lawyers turning into operators of tools they don't understand and won't question.

Most predictions about AI and the billable hour treat the legal industry as a single market. It isn't. Transactional and litigation work are different businesses, and AI breaks them at very different speeds.

The real danger of AI in legal work isn't hallucination. It's that the model agrees with whoever asked the question—and most clients don't know how to ask.

An M&A lawyer with 15 years of experience on what AI is actually doing to legal work, why most senior partners are wrong about their irreplaceability, and what the legal industry won't admit out loud.

A practicing M&A lawyer's honest review of using Claude Code for legal work. Why I stopped paying for "AI legal assistants" and what actually works.