Taking A 'Lego-Like' Approach: Why Enterprises Should Prioritize Stackable AI Agents

Enterprise AI is no longer a buzzword; it's an imperative part of business operations. Yet, for many organizations, the promise of intelligent automation is colliding with the hard reality that AI projects aren't generating ROI. In 2025, the grace period of "just figuring it out" is over. Many executives are now confronting a quiet truth—the ROI showcased in demos and marketing decks simply isn’t showing up in most applications. To combat the potential misallocation of capital, time, and resources, some organizations are still cautiously kicking the tires in limited, internal sandboxes.

The central challenge to ROI isn't a lack of vision, but the absence of a safe, scalable, and governable bridge from today's hype to tomorrow's ROI. The path forward, it turns out, may not be a revolutionary leap but a disciplined, architectural strategy that looks surprisingly familiar.

We spoke with Al Liubinskas, Vice-President and NA Cloud Integration Practice Lead at Capgemini, an executive who has spent over two decades architecting enterprise systems at scale. With his deep experience managing complex integration platforms like MuleSoft, TIBCO, and Apigee, Liubinskas argued that to understand the future of AI, we must start small and utilize a structured framework for turning AI into manageable, value-driven components. In other words, success comes from a Lego-like approach: assembling small, governed, stackable AI agents that can scale safely, deliver measurable value, and avoid the chaos of agent sprawl.

• The Lego block approach: "Companies will not hit a grand slam with one pitch," Liubinskas said. "I'm envisioning a Lego block model where you'll have an agent to create orders and another to offer different product options. It's a Lego approach versus having a sales agent that does everything from prospect to lead to order as one gigantic agent."

• Component-based architecture: According to Liubinskas, this modularity is also the key to accountability. "When something goes bad, there has to be a way of triaging what exactly happened. People want answers." He further argued that by putting it into a component-based architecture, it lets you scale when you want, how you want, without worrying about ruining the whole project in the process.

This Lego-like construction is a direct response to the primary hurdle facing enterprises today: scaling safely. While a proof-of-concept in a lab or sandbox is one thing, a production system that touches customers or sensitive data is another. For industry adoption, Liubinskas noted, there has to be a "guarantee of accuracy," which is precisely where enterprises are struggling. The delta between a demo and a deployed solution is vast.

• In the blind spot: "Everybody's focused on how to build the use case and get the use case working," Liubinskas said. "But to put it into production, it has to be bulletproofed, it has to be tested, and it needs to be effectively protected for the company that's providing that service."

If every department begins building its own AI solutions with dozens of modular agents, the natural outcome is chaos. Liubinskas said that enterprises are about to face a familiar challenge, one they just spent the last decade solving.

• API 2.0: Liubinskas echoed the sentiment of experts who predict AI's adoption wave will be no different than APIs; and on the path, many enterprises will have to combat agent sprawl. In his words, "We’re in the same situation we were when companies were juggling multiple gateways — whether it was Azure API Gateway, Apigee, or others. It feels like déjà vu."

This disciplined approach also provides a pragmatic path forward for companies paralyzed by imperfect data. Liubinskas pushed back against the idea that organizations must complete massive, multi-year data cleansing projects before they can even begin their AI journey. The key is to start in a domain where data is "good enough" to operate because if the data is bad, agentic AI is not going to solve it; it's just going to accelerate the problem.

• The data dilemma: Liubinskas clarified that enterprises don't need to wait for the perfect data state to get started, and you don't even need a PIM if you're utilizing the component-based approach. From his perspective, "By building in small stackable components you're able to build an order agent for one business unit and then a second one for another without starting from zero. So even if the data doesn't match, you're still able to build off similar components while you work with the data team to get the right information in place."

Ultimately, connecting all these Lego blocks, governing their use, and preventing agent sprawl requires a dependable central nervous system. For Liubinskas, the answer isn't a flashy new AI product, but the established, reliable workhorse of the enterprise—the iPaaS.

• The linchpin: Liubinskas and other experts believe leveraging an enterprise integration platform as the new enterprise agentic orchestration layer is the right fit for most enterprises. "I'm envisioning that integration platform vendors should be the linchpin that ties that whole orchestration layer together," he said. "It won't be a GenAI product; it will be an integration product, because the essence of discovery, logging, and trust was already baked into their platform."

For enterprises to finally move projects out of the ROI rut and turn AI hype into lasting results, leaders must embrace a more deliberate AI strategy. Success won't come from a single, revolutionary breakthrough, but from the careful and methodical assembly of trusted, well-governed components. As Liubinskas put it, "There's lower risk because you are cutting your teeth on the technology by learning and applying the guardrails. It's an evolutionary process, not a revolutionary process."