The Partner Who Teaches AI to Think Like a Lawyer

By Tom Richardson, JD — 2026-02-14

Tom Richardson spent seventeen years closing deals. M&A transactions, cross-border acquisitions, complex restructurings — the kind of work where a single misread clause can unwind months of negotiation. He left a senior partnership at one of the country's leading M&A practices three years ago to run his own boutique advisory firm. He joined IXO six months later, drawn in by a problem he recognized immediately.

The gap between language and legal reasoning

"I was watching legal AI tools enter the market and I kept having the same reaction," he says. "They were fluent. They sounded like lawyers. But they didn't reason like lawyers. There's a difference between producing text that looks like a legal opinion and producing analysis that would actually hold up."

Richardson's specialty is corporate law — specifically the intersection of contract language and commercial intent. It's a domain where context is everything. The same clause means different things in different jurisdictions, different industries, different deal structures. A model trained on legal text can reproduce the language without understanding the significance.

"Take indemnification. A model can identify an indemnification clause and summarize it correctly. But can it tell you whether the carve-outs are standard for this type of transaction? Whether the cap is appropriate given the deal size? Whether the survival period creates undue exposure? That's where legal expertise lives — not in the words, in the judgment about what the words mean in context."

His IXO work centers on four areas. Contract analysis evaluation — testing whether AI outputs on M&A documents identify material issues correctly. Clause-level annotation — creating structured training data that encodes commercial reasoning, not just legal definitions. RLHF for legal assistants — ranking model responses to complex transactional questions by accuracy and risk awareness. And adversarial review — deliberately probing models with edge cases designed to expose systematic misunderstanding.

The annotation that changed how he thinks about AI

Six months in, Richardson was asked to evaluate a model's analysis of a merger agreement. The model's output was, on its face, impressive — well-structured, comprehensive, appropriately cautious. But he found three errors that a junior associate wouldn't have made. None of them would have been caught by anyone without deep transactional experience.

"That was clarifying. These models are genuinely capable — more capable than most lawyers want to admit. But the failure modes are invisible unless you know exactly where to look. The model doesn't know what it doesn't know. That's why practitioners have to be in the training loop."

"Legal AI that hasn't been trained by practicing lawyers is just pattern matching on language. And in law, pattern matching on language is how you get confidently wrong."

Read The Partner Who Teaches AI to Think Like a Lawyer on the IXO blog