Alex Kardos-Nyheim is the founder of SafeSign Technologies and now co-lead of AI research at Thomson Reuters. Alex built a legal-specific large language model while still a trainee at A&O Shearman, ultimately selling the company in 2024. He explains why Safe Sign focused on building the “engine” rather than the interface and how a legal LLM was able to outperform some of the biggest AI labs on legal tasks.

A central thread of the conversation is how the competitive landscape in legal AI has shifted. What began as a wave of “wrapper” products around foundation models is now evolving into a deeper contest over models, data and defensibility. Alex argues that the real moat lies not in user interface features, but in training models on high-quality legal data and building robust retrieval systems that reduce hallucinations and improve reasoning.

We also explore how Thomson Reuters is thinking about the full stack: combining proprietary legal data, live market intelligence and a legal-trained model into a cohesive platform.

Finally, Alex shares what he’s seeing inside law firms. Smaller firms should be using AI as a force multiplier to “punch above their weight”, while larger firms face a strategic imperative to convert their institutional know-how into machine-readable data to power their AI tools. His advice to managing partners is direct: hire data engineers, structure your data and lean into expertise.

The Non-Billable Podcast is sponsored by Legora. To find out more about Legora, visit: https://legora.com/

Chapters

00:01 Introduction

01:20 Building a legal LLM while at A&O

03:40 Engine vs wrapper: why Safe Sign focused on the model

04:41 Life after the Thomson Reuters acquisition

06:24 Inside Thomson Reuters’ “skunk works” AI team

08:42 From point solutions to platform wars in legal AI

11:49 The perfect stack: models, data and RAG

15:07 Why retrieval engines matter more than you think

18:33 Data as the moat

23:56 Can small data beat big data?

27:14 Making legal AI “market aware”

31:05 The evolution of Co-Counsel

33:57 What law firms are getting right and wrong about AI

36:50 “Hire 30 data engineers tomorrow”

38:22 Productising law firm know-how

43:10 Tips if you started a legal AI company today

About Non-Billable

Non-Billable is the media company for modern legal professionals across private practice, in-house and legal tech.

Visit our website: ⁠⁠https://www.nonbillable.co.uk