Indian enterprises have largely moved past AI experimentation, yet many still struggle to take applications into production. The main bottleneck is increasingly the data layer, not the model. Fragmented systems, stale context and technical debt are slowing the shift to agentic AI, even as agents begin to reason, retrieve and act on enterprise information with little human oversight. In this conversation with CIO&Leader, Aamir Sait, VP-India, MongoDB, explains why live operational data matters, where vector search fits in the stack, and how permissions, auditability and governance must be built into the data layer from day one. He also shares lessons from Zomato, Leena AI and Emergent Labs, and what CIOs should prioritize over the next 12–24 months.

CIO&Leader: Indian enterprises have moved past the AI experimentation phase, but many still struggle to take AI applications into production. Across MongoDB’s customer base, what are the biggest technical bottlenecks blocking this transition?
Aamir Sait: The data layer is critical to scaling AI into production. Enterprises need accurate retrieval, real-time data access, and persistent context to scale AI beyond pilot projects.
Many enterprises can run AI pilots but struggle to scale them across the business because their systems were not designed for changing AI workloads, real-time interactions, or growing data volumes. Fragmented data, disconnected systems, and the need to maintain real-time context are major technical bottlenecks.
This is where MongoDB sees the data layer as a key differentiator. Our approach brings operational data, search, vector capabilities, embeddings, and reranking together on one platform, reducing complexity and improving accuracy.
CIO&Leader: AI applications are only as good as the data and context they can access. Why is access to live, operational data becoming more important for enterprise AI, and what changes are required in the underlying data architecture to support this shift?
Aamir Sait: Live operational data and accurate context retrieval are becoming critical because AI applications need access to the right information at the right moment. Stale or incomplete context can lead to inaccurate responses and, as applications become more agentic, to poor decisions or actions.
Supporting this means moving away from architectures in which AI relies on lagging copies of data. Operational data, search, and context need to sit close together so applications can retrieve what is true now and keep context across interactions.
CIO&Leader: Many enterprises still have data spread across relational databases, data warehouses, data lakes, and legacy systems. Does this fragmented data architecture fundamentally limit the ability to build reliable AI and agentic applications? If so, how should CIOs approach this challenge?
Aamir Sait: Technical debt is directly affecting AI outcomes. MongoDB’s research with IDC found that organizations that fail to address technical debt could face 50% higher failure rates and rising costs for their AI initiatives by 2027. The answer is therefore not simply to add AI capabilities on top of existing infrastructure, but to modernize the underlying data foundation that limits scalability and flexibility.
For a large Indian enterprise, the practical path is to address the parts of the architecture that create the most friction for modern applications and AI workloads. That can mean reducing the number of disconnected systems an application depends on, bringing operational data and search closer together, and making it easier to work with both structured and unstructured data. AI-powered modernization can help enterprises accelerate this process. MongoDB’s Application Modernization Platform (AMP) combines AI-powered tools, a proven framework, and engineering expertise to help organizations modernize legacy applications and their underlying data architecture faster, while reducing technical debt.
CIO&Leader: With agentic AI, applications are increasingly expected to reason, retrieve information, and take action autonomously. How does this change application architecture compared with traditional AI or software applications, particularly around data access, context, security and governance?
Aamir Sait: Agentic AI is changing application architecture from systems that primarily respond to predefined inputs to applications that can reason over information, make decisions, and take actions. This requires the architecture to provide agents with reliable, real-time access to enterprise data, while maintaining context across interactions, including memory, task state, and retrieved knowledge.
Security and governance also become more important because agents can act on the information they retrieve. Enterprises need to define permissions, establish guardrails, and maintain auditability so agents can access only the data and systems they are authorized to use and their actions can be monitored.
CIO&Leader: Vector databases and vector search have become central to enterprise GenAI architectures. Where do you see vector search fitting within the broader enterprise data stack, and what are the limitations of treating it as a standalone capability?
Aamir Sait: Vector search has an important role in the enterprise data stack because it allows AI applications to retrieve information based on meaning, rather than relying only on exact keyword matches. But it is one capability within a broader AI application architecture, not a standalone solution. Production applications need to combine semantic search with operational data, metadata and business context to deliver relevant results.
Treating vector search as a separate system adds complexity to data movement and synchronization, particularly as operational data changes. The result can be stale or inconsistent results, plus another system to secure and govern.
CIO&Leader: Companies such as Zomato, Leena AI and Emergent Labs are building AI-powered applications on MongoDB. What are you learning from these implementations about the data and infrastructure requirements of AI applications that are actually operating in production?
Aamir Sait: Nugget, Zomato’s AI-powered customer support platform, uses MongoDB Atlas to bring together the data needed to support AI-driven interactions. Less than four months after launch, it had improved human-agent efficiency by 40%. With Atlas Stream Processing and Change Streams, Zomato can also transform data in real time without maintaining separate data-transformation jobs, allowing Nugget to scale without adding architectural complexity.
Leena AI had a different requirement: its virtual assistants needed to handle high volumes of employee queries while retrieving the right policy context in real time. MongoDB Atlas brings together its document store, hybrid search, vectors, embeddings, and reranking, helping the platform scale across more than 500 global enterprises.
Emergent Labs is using MongoDB at the application-development layer. Its platform allows users to build production-ready applications through natural-language prompts, with autonomous agents writing, testing, and iterating on the code. Moving from PostgreSQL to MongoDB’s flexible document model helped Emergent adapt faster without rigid schema constraints. The platform has since powered nearly 12 million applications across 180 countries.
Across all three, the common requirement is a flexible data layer that keeps pace with changing applications and delivers current information at scale, without adding architectural complexity.
CIO&Leader: Agentic AI is also changing the role of developers. As agents begin generating code, interacting with APIs, and accessing enterprise data, which parts of the application development lifecycle do you expect to change most significantly, and what new engineering skills will developers need?
Aamir Sait: Agentic AI is changing application development from fixed workflows to systems that can understand data, respond to changing needs, and take action. Developers now spend more time defining the data, business rules, permissions and guardrails that guide these AI agents. As more AI applications move into production, strong data-layer skills are becoming essential.
Builders need to ensure that applications can retrieve the right context, maintain memory, work with both structured and unstructured data, and integrate AI agents safely into existing workflows. Skills in data architecture, modern databases, AI application design, security, and governance will become crucial. When these capabilities work together, builders can spend less time connecting infrastructure and more time building useful applications.
CIO&Leader: MongoDB is positioning Atlas as a platform for building AI-powered applications, with capabilities such as Vector Search and real-time data processing. What technical advantage does having these capabilities within the same data platform offer developers compared with stitching together multiple specialized systems?
Aamir Sait: An AI-ready application needs reliable operational data, accurate search, and access to information as it changes. MongoDB’s document model provides the foundation for this by accommodating evolving data structures, including semi-structured and unstructured data, without the rigidity of traditional relational tables. This flexibility is particularly relevant for AI applications, where data structures and requirements can change as applications evolve.
MongoDB Atlas provides the operational data foundation, while Vector Search helps the application find relevant information based on meaning. Real-time data keeps that information up to date as business conditions change. Together, they allow an AI application to retrieve the right context from enterprise data and use it to respond or take action.
Keeping data, embeddings, search, and AI models synchronized can lead to delays, stale information, and inconsistent results. MongoDB’s approach brings operational data, embeddings, search, and retrieval capabilities together on a single platform, reducing data movement and integration complexity.
CIO&Leader: AI agents can make decisions or take actions using enterprise data without a human in the loop. From a database and application architecture perspective, how should enterprises design for controls such as permissions, auditability, data lineage, and the ability to stop or constrain an agent?
Aamir Sait: Controls need to be designed into the data layer from the start, not added once an agent is live. In practice, that means giving each agent only the access it needs, using role-based access controls. It also means recording what agents read and write, through auditing and change streams so that every action can be traced. And it means being able to narrow or revoke an agent’s access by changing its permissions.
Running agents against one operational record, rather than copies spread across systems, also leaves fewer places to govern. And because MongoDB can run across clouds and in customer-managed environments, enterprises keep control over where their data sits, which matters under India’s DPDPA.
CIO&Leader: Looking ahead over the next 12–24 months, which technology shift will have the biggest impact on how Indian enterprises build AI applications: agentic architectures, real-time data, vector search, multimodal AI, or something else, and why?
Aamir Sait: The big shift will not be any one of those individual components but how companies bring them together in a cohesive way.
The long-term value in the AI economy will come from organizations’ ability to make their proprietary data usable at scale. That is why MongoDB is focused on bringing retrieval, embeddings, reranking, search, and operational data together on one platform.
For Indian enterprises, the opportunity is particularly significant, given the country’s large developer ecosystem and accelerating adoption of enterprise AI. Preparing the data foundation today will allow organizations to build and scale the next generation of AI-powered applications as these technologies mature.
