Episode Summary
In this episode of Leap to Scale, Justin Davis and Greg Ross-Munro are joined by Dr. John Licato, CEO and founder of SquarePak and Associate Professor at the University of South Florida, for a deep discussion about one of AI’s most important limitations: trust. Rather than focusing on flashy demos or speculative hype, the conversation explores what it actually takes to build AI systems professionals can rely on in high-stakes environments like legal contracts, compliance, and business decision-making. Dr. Licato explains why his company prioritizes “slow, reliable AI” over speed, how layered verification systems reduce hallucinations, and why accountability still has to remain human-centered.
The discussion also explores broader implications for professional services firms as AI adoption accelerates. Justin, Greg, and John unpack what happens when AI replaces junior-level tasks, why that could create long-term leadership gaps inside organizations, and how firms should rethink hiring and talent development over the next five years. Along the way, they dive into adversarial reasoning systems, AI agents debating each other, and the surprisingly difficult challenge of teaching machines when to say “I don’t know.” It’s a thoughtful and practical conversation about where AI is genuinely useful today, where it still falls short, and what business leaders should be preparing for next.
Episode notes
- Introduction to Dr. John Licato, CEO of SquarePak and Associate Professor at USF
- Why trust and reasoning are becoming AI’s biggest challenges
- How SquarePak approaches hallucination-resistant AI for contracts and legal documents
- The tradeoff between AI speed and reliability
- Why redundancy and layered verification reduce hallucinations
- The difference between replacing lawyers vs. augmenting legal work
- How AI changes the value proposition of professional services
- The role of accountability and human oversight in AI systems
- Why adversarial debate systems can improve AI reasoning
- Lessons from legal reasoning and court systems for AI design
- The challenge of building AI systems that can confidently say “I don’t know”
- Why optimizing AI for uncertainty can sometimes reduce performance elsewhere
- AI agents, debate structures, and truth-oriented decision making
- Risks of replacing junior employees too quickly with AI automation
- Why firms may face future leadership shortages if entry-level learning disappears
- The future role of universities and practical experience in developing AI-era professionals
- Advice for service firms adopting AI over the next five years
- Why customer discovery and finding product-market fit still require humans
- How AI is reshaping programming, project management, and professional expertise
- The importance of balancing efficiency gains with long-term talent development
Episode Transcript
Greg Ross-Munro: We’ve got our first guest on the show, everybody. It’s exciting. And we didn’t want it to be somebody who just has opinions about AI or software. We wanted somebody who was actually building, shipping, and taking accountability for a tool that’s getting used in the real world.
Today we have Dr. John Licato. He’s the CEO and founder of SquarePak, a platform designed to help professional services teams make better decisions inside high-stakes documents like contracts. Not just summarize them, but actually surface real risks, tradeoffs, and questions you should ask before you sign something you cannot unsign.
He’s also an Associate Professor at the University of South Florida’s Bellini College of Cybersecurity and Computing, where he directs the Advanced Machine and Human Reasoning Lab.
John is also a personal friend, so we know he’s not here to sell us a magic trick. He’s here to talk honestly about what works and what doesn’t work when building systems professionals can rely on. John, thanks for being here and being our guinea pig.
John Licato: Thank you for having me. Although I do in fact have strong opinions about AI and associated topics, so I’m not afraid to share them.
Greg Ross-Munro: That’s exactly why you’re here. Let’s start with SquarePak. What are you building right now, and why did you choose this space?
John Licato: We describe SquarePak as a layer for reliable, hallucination-resistant complex document intelligence.
Anybody who has interacted with ChatGPT or general AI tools understands hallucinations are a fact of life. They’re probably never going away entirely. But when accuracy matters, like contracts, proposals, or highly regulated documents, you can’t afford hallucinations.
So we’re intentionally going in the opposite direction from consumer AI tools. We’re completely okay trading speed and scalability for slower, more reliable AI analysis.
Greg Ross-Munro: What does “slow” mean? Seconds? Minutes?
John Licato: Mostly redundancy.
One way to reduce hallucinations is layered verification. If one AI generates citations, another layer checks whether those citations actually say what the AI claimed they said. Then another layer checks the reasoning holistically from different perspectives.
Each layer adds time and cost, but reliability matters more than speed for these use cases.
We also want to float on top of whatever the best models are at any given time. OpenAI will release another model tomorrow, then Anthropic will respond, and the cycle keeps going. We don’t want to compete directly with foundation models. We want to combine the best available tools and provide a trust layer over them.
Greg Ross-Munro: Out of all the AI problems you could work on, why this one?
John Licato: I’ve always been obsessed with correctness and reasoning.
What fascinates me about legal reasoning is that it’s essentially a system designed to get the best outcomes from imperfect human thinkers. Humans are biased. We use shortcuts. We make mistakes. But the legal process, courts, arbitration, adversarial debate, all of it is designed to produce better reasoning despite those flaws.
I’d been researching this for years before ChatGPT existed. Then ChatGPT arrived and suddenly we could actually build some of these ideas.
Greg Ross-Munro: So if I go to SquarePak today, can I replace my lawyer?
John Licato: No. Absolutely not.
But you can replace the pre-lawyer step. You can analyze contracts, identify risks, and generate better questions to bring to your lawyer.
And honestly, it’s useful post-lawyer too. Most lawyers don’t have hundreds of hours to deeply analyze how every clause aligns with your long-term business goals. They simply can’t do that economically.
These tools can supplement legal services and help organizations reason more deeply about contracts.
Justin Davis: I’m curious how this changes the value proposition of professional services.
If AI starts handling reasoning tasks that used to define legal work, what remains uniquely valuable about the lawyer?
John Licato: There’s still enormous value in strategic thinking and relationship context.
A lot of legal work today is extremely mechanical. Lawyers spend huge amounts of time jumping between defined terms in Word documents. Programmers solved this years ago with IDE navigation tools, but lawyers still don’t have equivalent workflows.
We can automate a lot of that low-level friction while still freeing lawyers to think strategically about long-term consequences, client relationships, and risk management.
Justin Davis: Can LLMs do that strategic thinking today?
John Licato: They can produce outputs that sound strategic. But I wouldn’t say they do it at a trustable level yet.
Greg Ross-Munro: At the end of the day, somebody still has to be accountable for the work.
John Licato: Exactly. Even if a lawyer uses AI, they remain ethically and legally responsible for the outcome.
The lines are going to get blurry over the next few years as autonomous agents become more capable, but accountability still matters.
There’s also a joke that law will be the last profession replaced by AI because lawyers won’t allow it.
Justin Davis: They control the rules.
John Licato: Exactly.
Greg Ross-Munro: Let’s shift to your research work. Your lab focuses on machine and human reasoning. What does “better reasoning” actually mean in practice?
John Licato: Originally, my work focused on formal logic and theorem proving. Everything was measurable and precise.
But after the 2016 election, I became much more interested in informal reasoning, discourse quality, bias, and how humans actually communicate.
I realized AI didn’t necessarily need to force people to think correctly. It just needed to act more like a spellchecker for reasoning. A small nudge. Something that says, “Hey, this argument may not be grounded in evidence.”
That became the direction for the lab. Making AI reason better in service of helping humans reason better.
Justin Davis: How do you define “better”? More truthful? More aligned?
John Licato: In a lot of our work, it comes down to discourse quality.
Most debates aren’t truth-oriented. People judge arguments based on whether they support their side, not whether they move everyone closer to truth.
That applies directly to AI agents too. A lot of people think you can just throw multiple AI agents into a debate and truth will emerge automatically. But the same problems humans have show up in AI systems too. Peer pressure. Dominant voices. Groupthink.
So a huge part of the work is designing systems where disagreement is structured productively.
Greg Ross-Munro: That ties into something we focus on a lot: auditability and traceability.
The danger with these systems is confidently wrong answers. Sometimes no answer is better than a bad answer.
John Licato: Humans have the same problem. We hate saying “I don’t know.”
Even trying to optimize language models to say “I don’t know” more often can create downstream performance issues in other tasks.
There may actually be a deep reason for that. It could be that the ability to deeply commit to an idea and explore it fully is fundamental to reasoning itself.
Really smart people still dive deeply into ideas. They just know when to back out and try another path.
Justin Davis: Maybe there’s also a persistence problem there. If an AI gets too comfortable saying “I don’t know,” maybe it also becomes more likely to give up on hard problems.
John Licato: Exactly. Humans work that way too. If you give up every time something gets difficult, you’ll never solve hard problems.
The challenge is knowing how long to struggle before changing approaches. That may actually be impossible to solve perfectly.
Greg Ross-Munro: And there’s also a difference between “I don’t know yet” and “this is unknowable.”
John Licato: Right. Which is why you need multiple approaches running in parallel. Different perspectives. Different persistence levels.
There’s never going to be one perfect AI agent that solves everything flawlessly.
Justin Davis: As you look ahead a few years, what should service firms realistically expect from AI?
John Licato: One thing AI absolutely has not solved is product-market fit or customer discovery. Nothing replaces talking to customers and developing intuition through real conversations.
But what we are seeing is that industries are shifting toward higher-level coordination and strategic work. Programmers become more like product managers coordinating AI systems.
The problem is that AI is replacing many of the junior-level tasks that traditionally trained future leaders.
I think we’re going to face shortages of experienced professionals because people won’t have had those formative years making mistakes and learning the fundamentals.
Greg Ross-Munro: That creates an interesting tension with universities too. For years people said higher education was becoming less important in tech. But if junior jobs disappear, where do future experts come from?
John Licato: I’ve always believed education and industry experience should be more integrated.
But I’ll also say this: building a startup taught me things academia never could. There’s still a massive gap between theory and operating in the real world.
I don’t know exactly how we bridge that gap yet.
Justin Davis: It’s something firms need to think about now, not five years from now.
The people leading organizations in five years are already inside those organizations today.
John Licato: And if future managers never learned how to actually do the jobs they’re coordinating, that creates a serious problem.
Greg Ross-Munro: John, thank you. It’s rare to find someone who’s both deeply technical and actively building a business in the real world.
You’re doing hard work in both spaces, and it’s incredibly impressive. Thanks for being here, for being thoughtful, and for being willing to talk honestly about all of this.
John Licato: Thank you, guys. And yes, we absolutely need to start playing squash again. That’s one thing AI still can’t do.
Justin Davis: Yet.
Greg Ross-Munro: Thanks again, John.