Kanchan Ray explains how autonomous systems, governed by strong guardrails and human oversight, can accelerate decisions, unify enterprise intelligence, and transform retail.
Retail and consumer businesses are moving beyond AI-powered insights toward AI-powered execution. As enterprises adopt agentic AI, the focus is shifting from isolated automation to intelligent systems that can sense, reason, and act autonomously across complex business workflows while remaining governed by enterprise policies.
In this conversation, Kanchan Ray, CTO at Nagarro, explains how agentic operating systems are redefining enterprise decision-making. He discusses why governance is critical to autonomous AI, how enterprises can build a unified intelligence layer across fragmented systems, and why the future of retail lies in reducing decision latency while keeping humans at the center of strategy, oversight, and accountability.
CIO&Leader: How does the architecture of Mosaic OS fundamentally differ from a standard multi-agent orchestration framework to safely allow agents to execute actions without constant human intervention?
Kanchan Ray: Most AI systems today generate recommendations, humans make the decisions, and workflows pause until someone intervenes. While this approach improves productivity, it does not fundamentally transform how enterprises operate. Mosaic OS was built to address this. It enables safe, autonomous action, not just orchestration.
For an AI agent to act responsibly without continuous human oversight, it must understand the business context in which it operates, including priorities, policies, constraints, risk thresholds, and escalation boundaries. This is where our Conscious Core comes in. It serves as a shared intelligence layer that gives every agent a unified understanding of the enterprise. From there, the orchestration layer dynamically routes work to the right agents, while the agents collaborate and hand off tasks without requiring a lot of integrations. Governance policies are enforced in real time, and observability continuously feeds outcomes back into the system.
Enterprises are moving beyond isolated AI tools toward an Agentic Operating System, where intelligence is embedded into every decision flow.
The result is an AI-first operating environment where agents can act confidently within clearly defined boundaries, escalate when necessary, and leave a comprehensive audit trail of actions and decisions. This reflects our broader vision: enterprises moving beyond isolated AI tools toward an Agentic Operating System, where intelligence is embedded into every decision flow.
CIO&Leader: In retail and CPG, data is famously siloed between legacy ERPs, supply chain logistics, and front-end e-commerce platforms. How does Nagarro establish a unified ‘living context layer’ across these fragmented systems?
Kanchan Ray: Organizations typically operate across CRM platforms, ERP systems, dealer management systems, commerce applications, customer service channels, and a wide range of structured and unstructured knowledge sources. Valuable context often sits inside SOPs, manuals, documents, videos, and even in the experience of employees rather than in a single system of record.
Our approach is to build an intelligence layer that connects these distributed knowledge sources and operational systems, allowing context to flow seamlessly across the enterprise. Rather than forcing all data into a single repository, we focus on creating a unified understanding that agents can access and reason against in real time. This is what enables Fluidic Intelligence, an environment where systems can collaborate, share context, and act on insights without being constrained by organizational or technology silos.
We have implemented a similar approach for an industrial manufacturing client where historical sales data, promotions, festivals, elections, and local events were connected into a unified decision intelligence layer. The system did not simply forecast demand; it reacted to changing signals in real time and continuously improved inventory decisions through reinforcement learning.
We typically build this through a combination of canonical data models, API-led integration, event streams, knowledge graphs/vectorized enterprise knowledge, policy-aware access controls, and observability. The goal is not to centralize all data physically, but to create a governed semantic and operational context layer that agents can reason against.
CIO&Leader: Can you share a practical example of how a brand uses an agentic ecosystem to sense a sudden viral demand signal on social media and autonomously adjust manufacturing or fulfilment pipelines in real time?
Kanchan Ray: Yes. The shift you are describing is exactly what agentic ecosystems are built for. Seasonal forecasting assumes demand moves in predictable cycles, but a trend can now start on social media in the morning, spike by afternoon, and trigger stock-outs by evening. That speed breaks traditional planning. The alternative is a continuous “Sense-Think-Act” loop, where agents sense signals from social trends, search behaviour, sales velocity, campaign performance, weather, local events, competitor pricing, and fulfilment constraints, reason across them, and then recommend or autonomously execute actions like inventory reallocation, replenishment, dynamic pricing, or fulfilment re-prioritisation. The value comes from orchestration. Instead of disconnected bots, one agent monitors demand volatility, another inventory risk, another pricing, and another fulfilment, all coordinated as one system.
A practical example: imagine a major influencer or celebrity posts about a product overnight. Traditionally the brand has little visibility into how that attention converts to demand and reacts only once shortages appear. In an agentic ecosystem, the agents detect the engagement spike as it happens, correlate it with rising search and sales velocity in specific regions, check live inventory availability, and then adjust the pipeline before stock runs out. They can trigger replenishment orders, reroute fulfilment priorities toward the regions heating up, and flag manufacturing or supplier capacity that needs to move. The brand is acting on the signal within hours rather than reconstructing what happened after the fact.
For low-risk actions, agents may execute automatically — for example inventory reallocation or fulfilment prioritization. For higher-impact actions, such as supplier capacity or manufacturing changes, they escalate with recommended scenarios and supporting evidence to a human-in-the-loop.
We saw the same principle work in a more traditional setting with an industrial manufacturing company moving beyond static demand planning. Its forecasting relied on periodic reviews, so demand shifts were often caught only after inventory decisions were already made. We built a decision intelligence layer that combined historical sales with external demand drivers such as festivals, promotions, elections, and local events. Reinforcement learning models continuously evaluated changing conditions to optimise ordering, replenishment recommendations adjusted automatically as signals shifted, and planners could run scenario simulations before committing. The result was faster response to demand fluctuations, lower inventory carrying costs, improved coverage, and shorter invoice lead times. Whether the original signal comes from a factory floor or a viral Instagram, the loop is the same: sense, reason, act, and learn continuously.
CIO&Leader: When AI agents are given the autonomy to dynamically alter pricing strategies or reroute inventory across warehouses, how do you design governance to ensure these agents don’t inadvertently trigger margin-eroding price wars or supply chain bottlenecks?
Kanchan Ray: The short answer is that autonomy without boundaries is the actual risk, not autonomy itself. A pricing agent left to optimise purely on conversion can chase a competitor’s discount into a margin-eroding race to the bottom, and an inventory agent optimising for one warehouse can starve another and create the very bottleneck it was meant to prevent. Good governance solves this by giving agents room to act fast within limits they cannot cross, and by catching cross-agent conflicts before they execute rather than after the damage is done.
That is how we have designed governance into Mosaic OS, embedded directly into how agents operate from the start rather than bolted on afterward. Agents are assigned objectives within predefined operating boundaries. A pricing agent can optimise dynamically, but it cannot breach margin thresholds or commercial guardrails, so it is constrained from making actions that would breach margin, brand, or commercial guardrails. An inventory agent must optimise across enterprise-wide constraints rather than solving for a single location, which prevents it from clearing one warehouse at the expense of another and creating downstream supply bottlenecks.
The orchestration layer is what stops these risks from compounding. Before any significant action executes, it evaluates dependencies across agents, identifies conflicts such as a pricing move that would collide with an inventory or fulfilment constraint, and applies policy controls. Where an action exceeds predefined risk thresholds, it is escalated for human review rather than executed automatically. Every action is logged, auditable, explainable, and governed by runtime policies. The goal of governance is not to slow agents down. It is to give enterprises the confidence to let them operate at speed while maintaining accountability, transparency, and control.
CIO&Leader: When you look at real-world implementations of Fluidic Intelligence, what are the primary metrics that prove an agentic operating model outlasts a human-driven one?
Kanchan Ray: I would frame it slightly differently. It is not about replacing human-driven models but about enabling a level of speed, responsiveness, and scale that would otherwise be impossible.
The metric we pay closest attention to is decision latency—the time between a signal becoming available and an action being taken. In traditional operating models, that latency may be measured in hours or days. In specific decision loops, that latency can move from hours or days to minutes, and in some automated workflows, even seconds. In today’s retail environment, that compression creates a significant competitive advantage.
Beyond decision latency, organizations track metrics such as forecast accuracy, stock availability, inventory turns, replenishment frequency, carrying costs, conversion rates, repeat purchases, and fulfilment efficiency.
The business impact typically appears in three areas:
First, customer experience. AI-driven agents make customer engagement more relevant, timely, and contextual, improving loyalty because customers feel the brand understands their intent rather than simply reacting to transaction history.
Second, operational efficiency. The biggest gains come from reducing waste, preventing stock-outs, minimizing overstocks, reducing manual intervention, and eliminating planning latency.
Third, business agility. Agentic systems allow brands to respond faster to demand spikes, supply chain disruptions, fulfilment bottlenecks, and sudden changes in consumer behaviour.
We saw this in practice with a global beauty brand where combining AI forecasting with human judgment improved forecast accuracy by 20%, directly improving inventory efficiency and product availability.
Our goal is not to remove humans from decisions. The goal is to remove latency from decisions.
However, the real ROI comes from building an enterprise architecture with explainability, audit trails, escalation paths, and clear boundaries between what AI can decide autonomously and what requires human intervention. Our goal is not to remove humans from decisions. The goal is to remove latency from decisions.
CIO&Leader: Looking ahead toward the end of the decade, as predictive commerce matures into fully autonomous retail ecosystems, what will the role of the human workforce look like?
Kanchan Ray: The workforce will evolve, not disappear. By the end of the decade, many routine operational decisions will likely be handled autonomously. But that does not reduce the importance of people but changes where people create value.
The roles most likely to decline are those centered on repetitive coordination, manual workflow management, and transactional decision-making. The roles that will grow are focused on agent design, governance, exception handling, strategic oversight, and business interpretation.
As retail increasingly becomes an Agentic Operating System, with intelligence embedded into every decision flow rather than added to existing processes, someone must define the rules, monitor outcomes, manage exceptions, and ensure that autonomous systems remain aligned with business goals, regulatory expectations, brand values, and customer trust.
Humans will move from being the operators of every workflow to becoming designers, supervisors, and stewards of intelligent systems.
I believe the most valuable human roles will sit at the intersection of domain expertise, technology fluency, and judgment. Humans will move from being the operators of every workflow to becoming designers, supervisors, and stewards of intelligent systems. The future workforce will not be less human; it will require humans to focus more on creativity, ethics, strategy, empathy, and complex decision-making — areas where human judgment remains essential.