Matt Scavetta, CTIO at FutureTech: “Agentic AI is different. It forces organizations to think more like software companies.”

Agentic AI promises revolutionary gains. But Matt Scavetta urges organizations to proceed with caution and look beyond the hype. The potential is there, but many leaders are focused on the agentic AI itself while overlooking the foundation required to make this technology work at scale.

Rather than relying on industry buildup, Scavetta, the Chief Technology Innovation Officer at Future Tech, built his own AI agent. In doing so, he better understands what it actually takes to make these systems work. This hands-on experience has reinforced his belief: the AI model is no longer the biggest obstacle to scaling. The challenge is that many companies lack the engineering, integration, and culture needed to deploy agentic AI safely at enterprise scale.

In this Q&A, he explores why scaling agentic AI isn’t a technology problem; it’s a cultural one. 

As companies start using agentic AI, what do you think they’re getting wrong about what it actually takes to make it work at scale?

Companies aren’t necessarily getting it wrong because the technology is so new. But many haven’t gone through the exercise of making sure they have the right teams and skills in place to build these systems. It’s so much more complex than traditional technology adoption.

With something like ChatGPT, you purchase licenses, roll it out, and maybe do some change management. The same is true with a new ERP, CRM, or any other enterprise system. Even if you don’t have all the expertise in-house, you can bring in a systems integrator, hire a few experts, or build an application support team to handle basic customizations.

Agentic AI is different. It forces organizations to think more like a software company. You can’t simply buy an AI agent and expect it to automate your business. You have to build the underlying systems and platforms, or at least engineer multiple applications to work together. 

I do think many CEOs, CIOs, and CTOs underestimate what agentic AI adoption means from an organizational perspective. Building and operating true DevOps and software engineering teams is a very different capability than implementing off-the-shelf enterprise software. It requires a different culture, skills, and a much deeper investment in engineering than many organizations are prepared for. There’s a big culture shift that needs to happen to make this work at scale. 

AI models are improving so quickly that they’re no longer the primary challenge for most enterprises. What do you think today’s challenge is? 

Today, the real challenge is everything you build around the model.

A language learning model (LLM) is the brain, but it needs tools, integrations, and structured inputs to do useful work. Some capabilities also need to be deterministic. For example, when I built a personal AI agent for investing, I didn’t ask the model to calculate things like capital runway or cash burn. Those are mathematical equations, or data you can pull from an API, so I handled that logic in Python and then fed the structured results into the model for reasoning. That approach produced much more reliable outcomes because I wasn’t asking the LLM to solve problems it isn’t best suited to handle.

That’s why I think the models themselves are becoming less of a limiting factor. They’re all incredibly powerful and capable, but you have to engineer everything around them. You have to think about how you connect it to systems, structure the data it receives, and determine what logic belongs inside the model versus outside of it. 

What sits between an AI agent and an enterprise system?

There’s a middle layer of software known as middleware that connects agents and systems. Its one task is to make sure the AI agent can securely access the right data it needs to do its job. 

Middleware is what controls integrations. Integrations are what enable an Agentic system to use tools. This is one of the key differentiators between a more common AI assistant and an AI agent.  

Organizations that have already invested in mature integration systems and practices are at a significant advantage because their systems are already designed to communicate with one another. If they haven’t, and have weak integration and middleware practices, they’ll quickly find themselves running an integration modernization initiative instead of an AI initiative.

Why is it so much harder to deploy AI agents inside a large company than it is for an individual user?

For individual users, you’re typically working in a single environment where you already have permission to access everything you need. The blast radius is small so that you can build, iterate, and deploy quickly with little risk. When you move it to the enterprise, everything changes. 

Now you’re introducing bureaucracy. You’re injecting security controls, governance, change management, and organizational processes. Your cybersecurity team has to ensure the right safeguards are in place, and every action has to be evaluated for risk.

That’s why I keep coming back to culture. If companies want to be successful with agentic AI, then they need a little more Silicon Valley in them. They have to be willing to move fast, iterate, gather feedback, and take calculated risks to harness this technology fully.

How important are identity, permissions, and governance when it comes to letting AI agents actually take actions in an enterprise?

It’s a non-starter. You can’t deploy AI agents at the enterprise level unless you can ensure that the person requesting an action actually has permission to perform it. Identity, permissions, and governance are the number one blocker. I’d argue they account for 40% to 50% of why it’s so much harder to deploy an AI agent in the enterprise than it is for an individual user.

An individual vibe coding from home, or an engineer working in a sandbox, has full access to the environment and applications within it. At work, that same person likely doesn’t have the permissions needed to install Python or other tools without going through the Help Desk or providing a business case. 

Every action an AI agent takes has to answer the same question: Does this user have permission to view this data or perform this action in the downstream system? Until you solve that problem, you can’t safely deploy enterprise AI at scale. 

People expected AI to drive big productivity gains by now. Why do you think that hasn’t really happened inside most organizations yet?

It’s because most organizations aren’t deploying agentic AI at scale. In fact, a recent Capgemini study found that only 2% of organizations have implemented agentic AI across their business.

Part of the confusion is that many companies are calling AI assistants “agents,” but they’re not the same category of AI. An assistant can answer questions or summarize information, while a true agent can reason, use multiple tools, incorporate feedback, and take action across systems.

If a company wants to move beyond pilots and experiments with AI, what do they actually need to get right first?

The biggest hurdle is that organizations don’t act. Leaders get caught in a spin cycle where it takes so long to make decisions and deliver something that projects end up over budget or lose momentum before they ever create value.

To me, that’s not really a change management issue. It’s, as you may have guessed, a culture issue. Many organizations aren’t used to operating like technology companies, yet they’re being asked to adopt increasingly complex technologies. And that requires a significant mindset shift.

If they want to move beyond pilots and start realizing the benefits of agentic AI, they have to be willing to move with speed, take calculated risks, and trust the people you’ve tasked with leading the effort. If every decision gets bogged down in bureaucracy, you’re going to spend more time planning than deploying. At some point, you have to start building. 

If you don’t feel comfortable giving your technical leaders the autonomy to do that, then either you hired the wrong people or your organization isn’t as ready for AI as it thinks it is.

Looking ahead, what separates the companies that will actually scale agentic AI successfully from the ones that won’t?

Companies need a dedicated team, mature integration practices, strong technical leadership, and a culture that understands software, data, and calculated risk. Those that will successfully scale agentic AI won’t excel in just one of these areas. 

Scaling agentic AI isn’t about getting one piece right. It requires the right people, the right technical foundation, and the right culture all working together.

More interviews

• David Benskin, Founder and CEO of Wealth Access: “You modernize by layering, not by high-risk wholesale system replacements”
• Adam Khan, VP, Global Security Operations at Barracuda: “In 2008 the enemy had a heartbeat… now it runs as code that never sleeps”
• Boston Dynamics VP: “Our goal is to deliver an intelligent, adaptable humanoid robot”