Episode 127: How AI Agents Could Shrink Law Firm Teams by 60% with David Wong Thomson

The question is, can you do the same work, but instead of having a manager and a team of five, can you have a manager and a team of two? And then those two associates have an army of AI agents to help them to do a lot of the work.

We're starting to see that in legal within transactional work to do due diligence, you can now offload a lot of the document work to AI. In tax and audit, you can start to offload some of the grunt work to AI and that starts to shrink those teams a little bit.

Hi, my name is Demetri Panici and I'm a content creator, agency owner, and AI enthusiast. You're listening to the AI Agents podcast brought to you by Jotform and featuring our very own CEO and founder, Aytekin Tank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello and welcome back to another episode of the AI Agents Podcast. In this one, we have the chief product officer at Thompson Reuters, David Wong. How you doing today, David?

Really great. Thanks for having me here.

Thanks for being on. First question, you know, as fellow people who embrace the cold, how bad is it right now in Toronto?

It is properly winter in Toronto right now. Low of minus 10 Celsius. Ice everywhere. I almost fell down on my way outside this morning. It's properly winter.

That's what I like to hear. It's going to be similarly bad Sunday. I think we're going to get like negative 15 Celsius as well. So rough out here, but we're really warm and excited about everything you guys are doing over there at Thompson Reuters.

Just to kind of kick things off, would love to learn a little bit more about you, David. You're currently leading product for a massive content-driven technology conglomerate. I'd love to hear about your personal journey. What was the path that kind of led you to this spot right here that you're in talking about AI and being the chief product officer at Thompson Reuters?

I really appreciate that. It's a bit of an unexpected journey. I didn't think I would be working in legal and tax and regulatory tech at this point, but my history is I've been working in product for over 15 years. I studied engineering, did some business consulting, and then like many, I went back into tech.

I've been building B2B software for about 15 years, focused first in marketing, then at Facebook working on advertiser systems. Then I got the call about six years ago to join Thompson Reuters. At the time, my first question was, what is Thompson Reuters?

Because I think a lot of people know Reuters as the news agency. Everyone kind of reads the news. Reuters reaches billions of people every month. They do polls during election seasons. They cover all the major world events, elections just being one of them.

What I didn't realize was that Thompson Reuters' bulk of business, more than 90% of the revenue, is software and research systems. These are software and research systems used by professionals—lawyers, accountants, tax professionals, auditors, risk professionals. It's the essential knowledge and software that underpins everything else: the justice system, the tax system, a lot of economic activity.

As I discovered more about what Thompson Reuters did, I realized this is something really important, fundamental to how the economy and society works. The opportunity was huge.

Six years ago, AI was just one technology among many that could be used for these use cases. Generative AI had not hit the scene in the way it has now. There was a hypothesis that tech and AI would transform these industries, but it was pretty slow going.

I made the bet that something would happen and that there would be some change. It turned out there was a huge change, just faster than expected.

You said 90% of revenue comes from software. Maybe that's a misconception about how companies work and where they make money. That's actually more common than you'd think for many companies.

Why do you think that misconception exists? Maybe because people see the news part and all the cool stuff they do there and don't realize the software side?

The news part is so well known. Everyone reads Reuters News. We fly under the radar. We're pretty quiet about what we do, quietly making progress and supporting customers in many professions.

As people learn more about Thompson Reuters, they realize there's some incredibly cool stuff happening. On top of that, with the AI transformation, pundits and analysts say the work most affected by generative AI tools is professional work done in back offices by legal, finance, and accounting teams.

There's a lot of hypothesis about that, and guess what? Thompson Reuters is where a lot of that happens already. We have a lot of data and tools, and that's what I'm excited to share more about today.

That's what I'm really excited to talk to you about as well. Tell us a little bit about how your role has evolved over the last few years. When did it really start to incorporate and work with AI in the sense that it is now?

I'm also curious, before you started working at Reuters, what was your initial exposure to AI? What was the moment when you interacted with it and thought, 'Oh crap, this is pretty incredible. I got to get my hands on this and use it more often'?

If you want the real story all the way back, I remember in high school doing a neural nets project for computer science. That was when it was the coolest. But then I put it aside because the common belief was it was just hypothetical and not practical. It took decades for the tech and findings to arrive at where we are today.

All the work I've done has been related to machine learning or predictive modeling. When I worked in marketing and advertising, it was about systems to estimate audiences and predictive parameters for marketing and sales.

At Facebook, the entire advertising system is basically one gigantic AI system to predict outcomes. It's one of the most powerful in the world and helped drive Facebook's revenues.

When I joined Thompson Reuters and switched to the professional space, it was a new domain, but I already believed AI was powerful and capable. The question was, did it work on the problems and use cases we have here?

Our journey at Thompson Reuters goes back almost 30 years. We were pioneers in natural language processing techniques. We had a natural language question answering query system released in the late 90s, a precursor to many search engines.

We've experimented with all precursors to current generation LLMs like transformer models and BERT to improve search and queries.

When generative AI hit the scene, it was a natural application for us. I remember August 2020, a couple months after I joined, a research scientist emailed me saying they got exposure to GPT-3 and asked, 'Is this our Kodak moment? Is this going to replace legal research?'

The team tested GPT-3 on legal research with grading systems. GPT-3 failed miserably, worse than an F student. The tech was getting better but not at the level lawyers demand.

Fast forward, GPT-3.5 and GPT-4 came out. We ran the same tests and saw the F student become a C++ student and then a B+ student with GPT-4. We realized this was the moment technology reached a stage to apply it.

At Thompson Reuters, we geared up investment in fall and winter when GPT-3.5 and 4 became available. We pivoted and invested about $150 million initially towards internal AI development using generative AI, starting with legal research use cases in 2022 and 23.

What are some examples of those legal use cases? How does this save time and do more?

There are entire podcasts dedicated to legal tech, but I boil it down to a few places. Lawyers and professionals generally demand either research and information retrieval to get answers to nuanced questions or help producing written work products like contracts, research briefs, court submissions, or tax returns.

The core use cases are information retrieval and producing written work products. We've had great success in the research step, helping answer questions and taking the grunt work out of legal and tax research.

Legal research for complex litigation can take days of grinding research to figure out nuances, arguments, statutes, and case law. We've been able to take hours out of that process and improve research outcomes.

What kind of companies do you work with? My girlfriend does legal research and would appreciate this.

We serve three major customer segments: law firms of all sizes from solo practitioners to the biggest firms; tax accounting firms from small to big, including the big four; and corporations with legal, finance, and risk teams.

With tax, what does that look like? People might assume tax codes and laws are just off the top of the head, but there's new research and laws coming out. How do you help with tax?

The law is always changing with new decisions and laws. You need to stay on top of it. We are one of the largest providers of research systems. On the tax side, a lot of work is about application and completing tax compliance.

We produce software to help corporations and tax firms complete taxes. If you work with an accountant, there's a good chance they're using Thompson Reuters software to do your taxes, not TurboTax or Pen software.

Enterprise companies don't use QuickBooks for taxes. I was working at a marketing agency running paid search for QuickBooks and we laughed at that.

The ultimate use case is helping tax professionals do taxes on behalf of clients. The AI opportunity is automating work within the tax prep process. Many applications are well suited to current agentic tech in tax.

Completing tax returns is routine but complex. You can break it down into steps: submit documents, process, interpret, fill forms, review, submit, and payment. These steps are complicated and ambiguous.

Until now, technology could only help with OCR and data extraction, not with interpreting tax positions or completing returns. Now, systems can make reasoned decisions to draft and create tax returns, thanks to advancements in LLM and agentic tech.

I like where this can help across the board. Healthcare removes paperwork effort, and legal and tax are huge categories for following complex rules repeatedly.

Would you say tax and law are opportunities for improvement because many components are if-then logic with numbers or rules? Maybe I'm spitballing, but I feel like at a pared-down level, especially tax, that's true.

Certainly on tax, it's a good use case because you can boil down the problem into subproblems guided by written instructions. When completing taxes, you get guidance from the IRS as blocks of text, not formulas.

Companies like ours interpret IRS guidance and turn it into formulas, wizards, and calculations. That assists tax completion if you know how to enter data into forms.

The labor situation in tax is well suited to automation. People say AI won't take jobs, but leaders want labor disruption because there's a labor shortage in tax and accounting.

Not enough people are entering accounting roles to fill new positions. Forecasts show a huge talent gap due to retirements. The industry has looked to tech for a decade to offset this gap. AI offers the greatest promise to make the profession better and more sustainable.

I could imagine that in these spaces because it's complicated but repetitive. There are interesting problems and solutions to save money, but also concerns about removing people since creativity isn't lost in accounting.

There's a human element in every job. I struggle to find arguments that accounting needs more creativity than creating images or video.

Parts that make tax and accounting unattractive are the drudgery, which AI promises to remove. There's a lot of creativity in tax and accounting, but you can only do creative work after the drudgery.

People want advice on tax positions, but can't get there until taxes are done. There's barely enough time in tax season to complete taxes, so not enough time for advisory or creative parts. That's the promise of AI systems.

I'd love to share what we've learned about building those systems and what we think will happen eventually.

Tax was viewed as a bad use case for LLMs because they sometimes get math right and sometimes wrong. It wasn't a given that LLMs could be applied to tax problems.

We solved this with agentic frameworks and techniques for constructing AI systems. Instead of teaching LLMs math, we taught them how to use a calculator.

We created an agent with automated API access to the entire tax calculator. Imagine TurboTax available completely via API with instant computation as inputs update.

This way, you don't worry about math accuracy, just whether the position is interpreted correctly and data input properly into the calculator. This problem suits AI systems better.

To build great AI agents for specialized work, you need a few ingredients: software tools, access to authoritative reference information like tax law, and training from experienced professionals.

You can't just have book smarts; you need street smarts from people who've done the work. Much professional work involves judgment based on prior art and industry consensus on applying laws or tax positions.

When you combine the right software, reference information, professional training, and mastery of LLM technology, you're more likely to create an agent that performs at a substantive professional level rather than a small assistant.

I like how you broke that down. Teaching AI to use a calculator rather than learning math shows how you're teaching it to functionally do a task. Many see AI as a magic bullet to solve all problems without much training, but it's more about fine-tuning models or agents around specific tasks.

You can have an army of specific task completers, which is what's doable now. General-purpose AI isn't there yet. Even perceived general-purpose AI in companies is probably a front-end interface delegating tasks to focused agents.

That's right. There's a parallel to human intelligence evolution: using tools made us more capable. People expect AI systems to become more capable, able to do more things for them.

That capability connects software integration with knowledge and experience. It's like training a smart team member to use tools and apply both book smarts and street smarts to do the work as expected.

What's your favorite part about what you do?

I love building new things. This is a moment of incredible invention for the entire industry. In the past two years, state-of-the-art in tax, accounting, risk, and audit has advanced more than the past 10 years.

Every month there's a new breakthrough or product release. Being part of that is a real privilege. My teams often say they feel like they're doing the best work of their careers right now, which is awesome.

I'm no prouder leader than to lead a team that feels like they're doing incredible work.

What drives people on your team day-to-day to feel like they're doing their best work?

There's a strong view at Thompson Reuters and similar companies that they finally see solutions to long-standing industry problems, like the talent shortage in tax and accounting.

Every year at industry events, people talk about solving the talent shortage. The hypothesis has been technology for a decade, but it's been incremental—reducing time on emails or engagement letters.

Now there's the promise of a big step change, which is motivating.

On the legal side, it's about advancing the justice system. Technology isn't just to make lawyers' lives easier, though that's good. It's also to make the entire justice system work more effectively.

That includes reducing court backlogs, addressing talent shortages like not enough judges or clerks, and getting better information out. Access to justice is a big opportunity.

We've made AI systems available to pro bono and access to justice organizations, improving outcomes and helping more people.

General techniques we're developing will be applied across the legal industry.

You mentioned tax firms want to implement AI to the point of workforce adjustment. What's the timeline for major changes like that? Where do you think jobs will be affected?

It's a big question. In corporations, there's pressure now. A recent MIT study said about 95% of AI programs aren't delivering ROI, defined as financial returns like cost savings or top-line improvements.

CFOs and Wall Street are demanding to see if AI investments lead to actual savings or cost structure changes.

That's why we're pushing hard on agentic technology. Broad horizontal tech gives small efficiency gains across many use cases, but that doesn't easily translate to bottom-line impact.

If someone is 5% more efficient, they usually do 5% more work, not lose their job. They become more effective, but it doesn't show directly.

When systems can do substantive work, like a paralegal or junior tax preparer, you can change team structure, needing fewer junior people.

That shows up in organizations' bottom lines. In professional services, the pyramid structure is getting squished or becoming torpedo-shaped, flattening with fewer juniors needed to leverage work.

Junior team members can be more productive with new tools, getting more work done for clients.

I agree there's big opportunity for AI. I like the torpedo analogy. It means fewer juniors are needed, not none.

In marketing agencies, many associate managers exist to manage subordinates. If subordinates reduce, fewer associate directors are needed.

The torpedo shape applies more in professional services, where work is delivered by a manager and associates.

Can you do the same work with a manager and two associates, each supported by an army of AI agents? We're starting to see that in legal transactional work and tax and audit, offloading grunt work to AI and shrinking teams.

What about things you're doing personally? What's your favorite AI tool right now?

My favorite is Text Cortex, a great mini agent creator, better than main model interfaces.

I don't get much personal time to explore because I focus on our products, but I'm a rabid user of horizontal providers like Anthropic, OpenAI, Google, and Microsoft.

Exploring Gemini 3 Pro's limitations and capabilities and Google's new workflow studios and integrations has been interesting. Their product strategy is coming together, and I see why OpenAI has code red now.

Code red was interesting because version 5.2 dropped recently, making big improvements in making decks and spreadsheets per request, which many people care about.

Last questions: Have you guys taken advantage of MCPs? What are your thoughts on them?

I think MCPs bring cool efficiency improvements, making people 5%, 10%, or 20% more efficient. The key is teaching someone how to properly use MCPs.

We see MCPs as a way to expand our systems' capabilities for customers. Legal and tax AI systems need to interact with other software customers use.

There's a lot of proprietary software in tax, accounting, and legal. Our opportunity with big enterprise customers is creating platforms to connect and integrate our intelligence and data with their software and data, and possibly their intelligence.

We're creating MCP integrations across systems for more capable workflows and AI systems. Much of this involves exposing private or proprietary data selectively and carefully, not dumping data indiscriminately.

Many MCP servers expose the right data as needed for workflows.

That makes sense. I had my CRM and cold outreach tool release their MCP recently. I was working with a buddy doing sales and told him he could ask questions about campaigns using Claude now, which is crazy.

The diversity of tools used by firms is wide. There's big opportunity to make AI assistants like co-counsel or generic assistants like Claude or ChatGPT access much of that data.

APIs are great for specific data endpoints, but MCPs let AI access many points at once, parse everything, and do things practically.

For example, I ran this podcast and wanted to help people get on more shows and do PR work. I used Claude and MCP to find the last six PR people, their websites, and what they do to offer good podcast PR services in five minutes, which would have taken hours manually.

MCPs are great. They tie back to ingredients needed to create capable systems: tools, knowledge, and training. MCPs are plumbing to access applications and data effectively for AI systems.

Exposing information is the first step. Stitching together and using all systems is where software companies spend time. Just MCPD tax law and a tax calculator to Claude wouldn't do your taxes; creating an agent around that is the next step.

Where can people go to learn more? Just go to www.tr.com. You can see everything Thompson Reuters does and learn about our AI technology under the brand co-counsel, serving legal, tax, audit, accounting, and risk professionals worldwide.

Thanks for your time and insights. Make sure to visit tr.com or thompsonreuters.com to learn more.

Thank you for listening. Please like, subscribe, and review the podcast on all platforms to help David and everyone at TR get their cool stuff out there that I wasn't aware of compared to their news forefront work. Thanks for watching. See you next time.