How AI Leaders Are Reining In Bottom-Up Sprawl Through Unified Integration Layers

The views and opinions expressed are those of Nithin Ramachandran and do not represent the official policy or position of any organization.

Enterprise GenAI is entering its orchestration phase, where context and connectivity have become the primary discussion points between forward-looking AI builders. After a year of pilots and point solutions, leading organizations are moving away from isolated demos and toward coordinated systems that can operate across workflows, data, and teams. Catalyzed by early GenAI stumbles, the shift is forcing CIOs to confront the fact that scale comes not from more experimentation but from engineering discipline, operating structure, and decisions about where AI actually belongs in the business.

Nithin Ramachandran is the Global Vice President of Data Analytics, MDM and AI at industrial machinery manufacturing leader 3M, where he heads enterprise-wide data, analytics, and AI strategy. A senior data and AI executive with leadership experience at companies like Kohler Co. and Direct Supply, he has spent his career operating at the intersection of platform engineering, business transformation, and executive decision-making. That vantage point informs a clear, opinionated view on why scaling generative AI now demands far more discipline than experimentation.

"AI doesn’t create value in isolation. Value shows up when systems, data, and decisions are orchestrated end-to-end," said Ramachandran. That principle underpins his approach to enterprise GenAI evaluation, starting with a clear distinction between foundational capabilities and those that deliver real business impact. From there, Ramachandran advised breaking AI initiatives into three levels, each with a different role in how value is created, governed, and ultimately scaled.

• Three levels of impact: "Level one is pervasive generative AI that applies to every employee, supporting basic tasks like document search and email and delivering AI-assisted productivity that now amounts to competitive parity," he explained. "Level two is domain-specific, because the AI needed in marketing looks very different from what’s required in finance, legal, or supply chain, and this is where measurable business value starts to emerge. Level three is the most specialized, built on proprietary company information that truly differentiates the business and creates a sustained competitive edge."

With a strategy in place, the focus turns to execution. Ramachandran views stability and reliability as outcomes that must be deliberately engineered through structure and discipline. To avoid the day-one failures that quickly undermine trust, he advocated a three-tier operating model: a platform team responsible for core technical reliability, product teams accountable for translating AI capabilities into business value, and business teams focused on embedding those tools into real workflows and processes.

• Unchecked expansion: That discipline starts by recognizing that leaders are often debating the wrong issue. "The build versus buy conversation no longer holds, because nobody truly builds models from scratch anymore. They configure baseline services with their own data, while the real problem goes unaddressed," he said. "Most companies take a bottom-up approach, collecting ideas inside functional silos, which leads to a rush of disconnected copilots and tools. That’s how AI sprawl takes root across the enterprise."

The solution to this fragmentation is to invest engineering resources in creating a connective tissue that sits above these disparate applications, using APIs and emerging agent-to-agent protocols. This is how an organization makes the leap from simple information retrieval to orchestrated, task-based automation. And, Ramachandran argued, it’s where the true enterprise ROI is found.

• The wrapper layer: "Where you want to really invest your time is in creating that 'wrapper layer' over everything that connects to all of these systems and could orchestrate AI workflows across systems. That is another area, especially in large companies, where it's worthwhile to invest."

• Orchestration makes the impossible possible: "Companies can now manage an agentic interface to connect all agents across systems. Most of the industry is aligning to this because without that, productivity gains from generative AI in the enterprise are impossible."

As the pace of technological change begins to stabilize, Ramachandran noted that the real work is shifting to areas that move far more slowly: human adaptation, data readiness, and the redesign of core processes. These foundational efforts, rather than new model releases, are what ultimately determine whether AI delivers lasting value inside the enterprise.

• Slower than software: "The change management cycles take longer than the technology release cycles. Our ability to drive change, train users, and redesign processes is what really impacts value. Working with the workforce and enabling change continues to be the source of greater value." It shifts the leadership challenge away from adoption and toward patience, sequencing, and sustained follow-through across the organization.

• The cost of an error: Ramachandran applied a pragmatic risk lens to data readiness, arguing that governance should scale with the real-world impact of getting an answer wrong. "The level of data quality you need depends upon the cost of an error. If a model is responding to a customer service query about a product and the information is wrong, that’s a problem. In cases where an error can undermine a customer experience, you must maintain high levels of data quality and put real-time data observability in place."

• Context is king: "A bottleneck that isn't given enough consideration is metadata. It’s the data about the data. When an agent works, it needs to know the context, like what part of the business process a customer order is in right now," he continued. "You need a knowledge graph of metadata over your repositories to actually advance with AI."

Taken together, Ramachandran’s approach offers a way out of pilot purgatory and into execution at scale. Rather than rewarding enthusiasm and experimentation, it forces hard choices about where AI is actually ready to deliver value and where it is not. "I always advocate a top-down strategy process, not a bottom-up one," he concluded. "Go through a simple checklist: do you have process readiness with a documented business process, people readiness with teams willing to make a change, and data and technology readiness? Without that, it’s impossible to move forward with discipline or focus investments where they matter."