At TechSparks 2026, VeloDB's Regional Director for APAC, Nitin Waghulde, delivered a blunt assessment of one of the most hyped categories in enterprise technology: AI agents. In his view, the industry has moved too quickly from impressive demonstrations to expectations of production-grade autonomy, without sufficiently hardening the infrastructure that sits beneath these systems.
Waghulde's argument reflects a growing concern among builders and investors alike. AI agents can appear highly capable in controlled pilots, where the data is curated, the task scope is narrow and human oversight remains close at hand. But once deployed into live environments, they confront inconsistent data, latency constraints, permissioning issues, changing business rules and the need for reliable memory across sessions. That is where many systems begin to break down.
Demo To Deployment Gap
The central problem, Waghulde suggested, is that a successful demo can mask structural weaknesses. A prototype may answer questions well, complete a workflow or generate a compelling output when the underlying dataset is small and the prompts are carefully designed. Production, however, is a different test entirely. Real users ask unpredictable questions, systems must respond quickly, and the agent must operate across multiple tools and data sources without losing context.
That shift exposes the limits of agent architectures that rely too heavily on the model layer alone. If the retrieval system is slow, the database is poorly tuned, or the data foundation is fragmented, the agent's output degrades rapidly. In practice, the user experiences this as hallucination, stale answers, broken workflows or inconsistent responses — symptoms that are often blamed on the model, even when the root cause lies elsewhere.
Waghulde's remarks are especially relevant in India's startup ecosystem, where AI adoption has accelerated across customer support, sales, internal operations and developer tooling. Many young companies are eager to showcase agentic products to customers and investors. Yet the move from pilot to scale requires more than a polished interface or a strong foundation model. It requires systems that can manage state, enforce access controls, retrieve the right information in real time and maintain reliability under load.
The Database Problem
Waghulde's key message was that the database underneath the agent is not a back-end afterthought; it is part of the product itself. For AI agents to function reliably, the storage and retrieval layer must support fast queries, structured and unstructured data, and the ability to keep context fresh as information changes. Without that, even a sophisticated model can produce brittle results.
This is where many enterprise deployments stumble. Companies often invest heavily in model selection and prompt engineering, but underinvest in data architecture, observability and operational controls. The result is an agent that looks intelligent in a demo room but cannot sustain trust in a live business process. In sectors such as finance, logistics, healthcare and commerce, that gap can be costly.
The broader industry lesson is that AI agents are not simply a software feature; they are an end-to-end system. Their performance depends on the quality of the data pipeline, the speed of retrieval, the freshness of records and the ability to trace how an answer was produced. As enterprises move from experimentation to deployment, the database layer becomes a strategic dependency rather than a technical detail.
Production Needs Discipline
Waghulde's comments also underscore a shift in the market narrative. The early agent boom was driven by the promise of autonomy: systems that could act, decide and execute with minimal human intervention. But production environments reward discipline more than ambition. Enterprises need auditability, predictable latency, governance and resilience. They need agents that can fail safely, not just impress in a showcase.
For VeloDB, the message is commercially significant. By framing the challenge as a data infrastructure problem, the company is positioning itself in the stack where enterprise AI becomes operational rather than experimental. That pitch is likely to resonate with technical buyers who have already learned that model quality alone does not guarantee business reliability.
The broader takeaway from TechSparks 2026 is that the AI agent conversation is maturing. The market is moving past the novelty of autonomous demos and toward the harder question of what it takes to run these systems at scale. Waghulde's answer was clear: if the database cannot support the agent, the agent will not survive production.