On the 67th episode of Enterprise AI Innovators, host Evan Reiser (CEO and co-founder, Abnormal AI) is joined by Joel Hron, Chief Technology Officer at Thomson Reuters, to talk about what AI looks like inside a 150-year-old information company that is past the adoption phase and into restructuring. With more than 5,000 engineers and franchises spanning legal research, tax preparation, and audit and compliance software, Thomson Reuters is one of the larger test beds in the enterprise for how agentic AI changes deep knowledge work.
The clearest example sits inside the company's largest product. Westlaw is the world's leading legal research platform, and Joel's team has rebuilt the experience around an agentic, deep-research paradigm. Westlaw Deep Research, released midway through last year, reads through case law the same way a web deep-research agent reads through hundreds of websites and synthesizes it into a coherent, citation-backed study. A next generation of CoCounsel, which already serves more than a million users, is in beta and extends the same pattern across a broader set of legal workflows. The legal industry, long described as a technology laggard, turned out to be one of the fastest movers once it had a tool worth moving on.
Inside the engineering organization, the structural shift is what happens when AI crosses 50% of the code being shipped. One of Joel's engineers flagged that threshold to him at the end of last year, and the framing stuck. Once AI is the primary contributor to the codebase, the engineer's job changes shape. As Joel puts it, "the lines of code were just the way for you to distribute and express your ideas and judgments." Last year, more than 90% of the engineering org used AI tools every day. This year, they aim to be AI-first in engineering, with AI shipping pull requests independently while humans focus on review, direction, and the systems that keep the work on rails.
The same pattern is showing up in tax. Thomson Reuters has built agents that take W-2s and 1099s, read each document, extract the values, interpret how the tax code applies, decide which deductions are most advantageous, and compute the actual return. The job that used to consume the bulk of a CPA's time is increasingly done by the agent, and the human shifts toward advisory work. That shift is landing in a market where the supply of new CPAs is shrinking each year, which makes the agentic capability a labor-supply story as much as a productivity story.
None of this is friction-free. Joel is direct about the two things that still hold enterprise agent adoption back. The first is access control: what files an agent can read, what applications it can reach, whether it can touch the password manager, and what it is allowed to do with any of it. The second is process: building the workflows in which an agent actually does work rather than just acting as a thought partner. His mental model for closing both gaps is treating AI as a colleague rather than a brainstorming tool. A colleague is someone you delegate to, which means giving them enough context to succeed but not so much that you would just do the job yourself. The same discipline applies to AI: design the context, design the guardrails, then hand the work over. Joel uses the model on himself, as well. Before a Friday product review, he will clone the repo and spend an hour with Claude or with Codex understanding what changed, which is the kind of executive depth that was not available at 5,000-engineer, 100-product scale before.
The principle Joel keeps returning to is about the kind of engineer who compounds in this environment. The horizontal slice of the T-shaped engineer now matters more than the vertical slice, and curiosity alone is not enough; what separates the people with outsized leverage is the ability to apply curiosity to learn fast, fail fast, and identify as a problem solver rather than as a front-end or back-end specialist. For CTOs, his advice rhymes with that: velocity is non-negotiable, but the work that lasts is the work that ships durable primitives, not the flashy thing with a three-month shelf life.
