Most enterprise AI initiatives don’t fail on the model — they fail on the meaning. Data exists in abundance across modern organizations, but the business context needed to make that data reliably usable for AI rarely does. The result is a familiar cycle: promising pilots, stalled production, and months lost rebuilding the same foundational groundwork for every new initiative.
Niraj Kumar, Chief Technology Officer of Onix, argues that the missing piece is a semantic layer — one that understands enterprise data before a question is even asked. With Wingspan 2.0 and its Semantic Twin architecture, Onix is making the case that AI execution doesn’t need to be hard. It needs to start from the right foundation.
Chief Technology Officer
Onix
CIO&Leader: Wingspan 2.0 promises a “unified, context-aware intelligence layer” — what does that actually mean in production, and how does it differ from a traditional data platform?
Niraj Kumar: In production, the difference shows up in how much effort it takes to make data usable for AI. Traditional platforms are effective at storing and moving data, but they do not understand what that data means in a business context. That missing layer forces teams to manually define relationships, map ontologies, and maintain context over time, which is often where AI initiatives stall.
Wingspan 2.0 removes that dependency through a living intelligence layer across the enterprise. It automatically maps data relationships, tracks lineage, connects processes, and aligns everything to business KPIs without requiring manual intervention. This context is continuously maintained and made available to every AI agent and workflow, so each operates from a shared, up-to-date understanding of the business.
That shift changes the architecture entirely. Instead of querying systems to assemble meaning, organizations work with a platform that already understands their data landscape before a question is even asked. As a result, the path to enterprise AI readiness compresses from 6 to 18 months down to just 4 to 6 weeks while also avoiding the rigidity and lock-in that come with manual ontology-driven approaches.
CIO&Leader: The Semantic Twin is a bold concept — how does it bridge the gap between raw enterprise data and real business context, and what does high-accuracy decision-making look like in practice?
Niraj Kumar: Most enterprise data is not short on information; it is short on meaning. It captures what happened across systems, but it does not naturally explain how those events relate to the way the business actually operates. That gap is where most decision friction begins, because the same data point can be interpreted differently depending on who is looking at it and where it is accessed from.
The Semantic Twin addresses this by grounding data in business reality rather than leaving it as disconnected records. It brings together how systems behave, how processes connect across them, and how the enterprise defines and applies core business concepts. In doing so, it removes the need to repeatedly reconstruct meaning before data can be used, as the interpretation is already embedded in how information is organized and understood.
This changes the starting point of decision-making. Instead of beginning with raw signals that need clarification, decisions begin with information that already reflects how the business functions in practice. That reduces variation in how different teams interpret the same situation and brings more consistency to how actions are taken across the organization.
High-accuracy decision-making, in this sense, is not just about the correctness of output. It is about whether decisions remain aligned with the enterprise’s actual operating behavior, even as inputs and conditions evolve.
CIO&Leader: Agentic AI is generating a lot of noise right now — where does autonomous operation end and human oversight begin in Wingspan 2.0’s architecture?
Niraj Kumar: What tends to get lost in the current noise is that autonomy on its own is not the goal, especially in environments where decisions have real consequences. Enterprises are not looking for systems that act faster; they need systems that can act correctly within the boundaries they already operate under.
Wingspan 2.0 reflects that reality by allowing agents to take on complex tasks, but only within a clearly defined operating context shaped by compliance requirements, governance policies, and business priorities. This means agents can move independently across workflows, but they never operate in isolation or make assumptions beyond what the enterprise has already validated.
The point where people step in is not during execution but when outcomes need to be understood and trusted. Instead of opaque results, the platform surfaces how a conclusion was reached, with supporting evidence that can be reviewed and challenged if needed. That shift makes oversight more meaningful rather than more frequent. It was evident when a risk and compliance agent reviewed FCA, HIPAA, and PCI DSS frameworks in 90 seconds, producing findings that were both fast and fully traceable, without requiring constant human intervention.
CIO&Leader: Multi-agent orchestration at enterprise scale is notoriously complex — what are the biggest failure points you’ve seen, and how does Wingspan 2.0 address them?
Niraj Kumar: What tends to break first in multi-agent environments is not the agents themselves, but the lack of a shared foundation they can rely on. When each agent pulls from slightly different data definitions or outdated context, coordination quickly falls apart. Outputs start to conflict, dependencies are missed, and teams end up spending more time fixing inconsistencies than benefiting from automation.
This usually traces back to how work gets started. In most enterprises, every new initiative begins with rebuilding context from scratch, often taking six to eight weeks to align on definitions and dependencies. Even then, gaps remain, which is why large programs frequently run about 40% over timeline with constant rework, leading to significant cost leakage across engineering effort, duplicated work, and delayed delivery cycles.
Wingspan 2.0 removes that starting point altogether. Instead of recreating context for each project, all agents operate from a continuously maintained semantic twin that serves as a shared, up-to-date knowledge graph. This keeps every agent aligned from the outset, so coordination becomes natural rather than forced. The result is fewer breakdowns, no repeated groundwork, and a much faster path from execution to measurable outcomes without the usual cost leakage.
CIO&Leader: You’re positioning the “Enterprise Intelligence Fabric” as the next evolution beyond data infrastructure — what does an enterprise need to retire or replace to get there?
Niraj Kumar: What stands in the way of an Enterprise Intelligence Fabric is not technology maturity, but how enterprises still think about change. Transformation is often treated like a sequence of contained projects, where work begins, ends, and then resets. The problem is that every reset wipes out the operational memory needed to move faster the next time, so progress always feels harder than it should.
To break that cycle, several long-standing patterns need to disappear. One-time migration programs have to give way to continuous modernization, in which systems evolve as part of ongoing operations rather than as fixed delivery milestones. The same applies to operational oversight models across DataOps, AIOps, and FinOps, where human review remains the primary control mechanism, even in environments that demand real-time response.
Compliance follows the same shift. Instead of waiting for audits to uncover issues, regulatory frameworks can be continuously evaluated with traceable outputs that surface risks as they arise, not after the fact. When these shifts come together, intelligence stops being something added onto infrastructure after deployment. It becomes part of how the enterprise runs, turning transformation from a series of projects into a continuous state of execution.
CIO&Leader: Studies show most enterprise AI pilots never reach production — how much of that failure is a technology problem versus a missing semantic layer problem?
Niraj Kumar: Most enterprise AI initiatives fail to reach production, not because the technology is insufficient, but because the underlying data environment is not ready for it. While models and infrastructure can perform well in isolated pilots, they are quickly exposed when deployed in real enterprise environments.
The core issue lies in data modernization, which remains incomplete or uneven. Enterprises often have modern platforms in place, but data still lives across legacy systems, cloud stores, and disconnected tools that were never designed to operate as one environment. Even after modernization efforts, business meaning is not consistently aligned across these systems, leading to fragmented interpretations of the same information.
In pilots, this complexity is usually hidden because the scope is narrow and the data is carefully curated. But once AI is connected to broader workflows, gaps in lineage, inconsistent definitions, and unclear dependencies begin to surface. That is when outputs stop being reliably usable in day-to-day operations.
So the failure is not about AI capability itself. It is about incomplete data modernization combined with a missing unifying layer that connects fragmented systems into a consistent, usable enterprise context. Without that, production-scale execution cannot be sustained, regardless of how advanced the underlying technology stack becomes.
CIO&Leader: What is the single hardest engineering challenge in moving an enterprise from AI experimentation to full-scale AI execution — and how did your team solve it?
Niraj Kumar: The hardest engineering challenge is not building AI agents that perform well individually, it is making them behave predictably when they operate together across a real enterprise. In controlled environments, each agent can be tuned, guided, and corrected by engineers who understand the full system context. That hidden human layer quietly resolves ambiguity and keeps outcomes aligned.
The difficulty begins when that layer can no longer exist. At scale, no team can manually supervise how every agent interprets data, applies logic, or interacts with other agents across hundreds of workflows. Without that coordination, even well-built components start producing inconsistent outcomes simply because they are working from slightly different assumptions.
The engineering solution is to design agents that are not isolated performers but parts of a coordinated system. Eagle handles discovery and maps dependencies across environments so agents are not working blind. Raven manages legacy transformation so older systems can still participate in execution. Pelican ensures validation as data moves through pipelines so outputs remain stable. Kingfisher generates synthetic data where real inputs are incomplete, so workflows do not break. Phoenix converts the combined output into usable intelligence for decision-making.
What actually solves the scale problem is not the capability of any single agent, but the way they stay synchronized through a continuously evolving shared operating layer. That coordination removes the need for repeated manual alignment and allows execution to remain consistent even as complexity increases.
CIO&Leader: AI-driven operating models claim to cut manual effort significantly — what’s a concrete, measurable outcome a Wingspan 2.0 customer has achieved that you can point to?
Niraj Kumar: IBX’s outcome reflects what changes when enterprises stop treating AI as isolated pilots and instead run it on a coordinated modernization and execution layer. In large healthcare environments, the constraint is rarely data availability. It is the amount of manual effort required to prepare, align, and operationalize that data across systems before it can reliably support decisions.
Wingspan 2.0 addresses this by bringing together agentic AI and a Semantic Twin approach that maps the enterprise’s data landscape, system dependencies, and business context into a unified structure. This reduces reliance on repetitive manual effort in data-heavy workflows and helps shift execution away from fragmented, tool-by-tool handling toward a more continuous operating model.
The measurable outcomes reported include a 50–80% reduction in manual effort across data-intensive processes and a 3× acceleration in modernization cycles, enabling organizations to move faster from legacy environments into AI-enabled execution.
What changes in IBX’s context is not a single workflow, but how work itself is structured. Effort previously spent on repeatedly preparing and reconciling data across systems is significantly reduced, allowing teams to redirect focus toward execution and operational outcomes rather than data enablement work.