How this CTO builds the foundation for AI

Key takeaways

• Robert Lyons, CTO of Katz Media Group, takes a sequenced approach to AI—building infrastructure, clean data, and governance foundations—before tackling harder AI challenges.

• Before deploying any AI, Lyons maps initiatives on a "value/effort matrix" that ranks ease of implementation against business value, starting in the high-value, low-effort matrix.

• To drive adoption, Lyons ran a company-wide AI primer webinar before rolling out any tools—presented by a third-party research firm, not IT.

Robert Lyons has seen the AI frenzy up close. As CTO of Katz Media Group, the largest media representation company in the U.S. and an 800-person business unit inside the 10,000-employee iHeartMedia enterprise, he's responsible for the systems, infrastructure, and innovation agenda that keep the company running and growing. 

He's also watched peers make the same avoidable mistakes: feeding messy data to AI and wondering why results disappoint, chasing hard problems before proving out wins, and rolling out tools before the organization is ready to use them. Lyons takes a different approach, grounded in a simple conviction: If your foundation isn't ready, neither are you. The following is an edited transcript of a recent conversation.

You’ve been leading your organization through a digital transformation and now looking ahead to an AI transformation. What move has had the biggest impact on IT’s ability to deliver on that?

Robert Lyons: It was when we decided to go from on-prem to cloud, reducing the amount of manual work we were doing, and accelerating our ability to implement change so we could better control and optimize cost spend. Cloud allowed us to migrate to SaaS solutions, which let us focus more on what our core competencies were inside of our own organization. My core competency isn't building ITSM software or my own ticketing system. Our business is selling radio, TV, and digital audio/video advertising. Transitioning to SaaS platforms meant we spent less time running IT and more time directing our efforts around innovation, customer experience, and revenue-driving work. 

A lot of IT leaders are under pressure to "do AI" right now. You have a more measured view of that, focusing on getting your house in order first. What’s the most important place to begin?

Data fundamentals are huge before tackling any AI—cleaning and getting all our data well-labeled—so that we're not just feeding a bunch of data to AI, but deliberately providing good, clean data to get a good outcome. You also must make sure your infrastructure and plumbing is ready: a reliable API, optimized integration processes, solid event flows. A lot of work was spent prepping for that before we tackled some of the harder problems. And then security and governance: data privacy controls, identity access management. What AI is allowed to do versus what it's not allowed to do is an important part.

How did you sequence your AI initiatives, so the foundation work didn’t just become a reason to stall?

My decision-making process for what we were going to do now versus what we were going to do later was working through use cases that make sense for the business and then determining what the success outcome would be for that particular use case. I really prioritized these things by ranking ease and effort of implementation against delivered value, creating kind of a value/effort matrix. The whole goal was to start with what is easy and delivers the most business value, get a solution out in front of people, and show deliverability and actual success before tackling the harder challenges.

Read also: Author Bob Sutton on when business leaders need to add 'good friction' to critical workflows

How did all this preparation pay off?

The data optimization piece was relevant as it pertains to implementing Freshworks’ Freddy Virtual Agent inside of our ITSM platform by first optimizing our knowledge base and service catalog. Most importantly, we wanted to define and establish a clear, self-service outcome. Before touching any content, we wanted to define what success looked like. And for us, that was ticket deflection, that was the KPI we were going to go after. We didn't want to just optimize all content for completeness; we wanted to target which tickets had the most opportunity for deflection and then optimize the data around that to get success.

How do you get people to use AI once it's live?

It depends on the audience. For agents using Freddy Copilot inside the platform, there's no resistance. For the broader organization, a lot of it is marketing. We're not promoting self-service as "we're never going to be here for you again." It's another channel to get assistance faster. We're also not a 24/7 shop, so we can market it as giving people options after business hours. But before any of that, we ran what I called an AI primer webinar for all employees. The purpose was to give everyone a fundamental understanding of how AI works, so that after this, they could hold their own in any conversation about AI. And I had our research advisory firm present it, not IT, so it's not IT barking at you. A neutral party socializing this makes it land differently.

You're also expanding AI-powered service delivery beyond IT to other functions. What does that look like?

We've got IT, facilities, accounting, and finance all living in the same workspace. I'm excited about the AI-powered agent possibilities for determining workflows that cross these different functional areas, but the key is that domain-specific stuff needs to stay within that domain. We must deliberately define what crosses a business function boundary before we build it out.

The first use case is onboarding and offboarding, because it touches every one of those functions. A new employee needs equipment and apps from IT, building access from facilities, and if they're in a sales role, a company credit card from finance. It's all about finding those cross-functional workflows and enabling them deliberately.

What do you tell other CTOs who are just starting out, especially if they're starting from a messy operational foundation?

That's where I gravitate right back to that matrix. Focus on solutions with easy implementation that deliver immediate value. Keep delivering that way inside of the organization, while you then work on improving those operational foundational changes—data governance, tackling the hard custom AI challenges. Don't start with the worst problem first, because you're not going to deliver the value. Focus on ease of implementation with immediate payback.

When you imagine Katz's operations 18 months from now, what's the most meaningful difference from today—the thing that will make you say, "this worked"?

Operations running on AI-assisted decisions by default, not as an exception. Most of our operational decisions are humans choosing from AI-generated recommendations, not starting with a clean slate. If I can get there in 18 months or sooner, I'd say this thing is working. You can't get there by chasing perfection. But if you can immediately deliver 80% of the way there and then iterate, there's business value in that.