AI’s next big challenge is making intelligence work at scale: Rangarajan Srirangam

The conversation around artificial intelligence in Indian enterprises is moving beyond experimentation to a more pressing question: how can organisations make AI work reliably at scale? For Rangarajan Srirangam, Senior Regional VP of Solution Engineering – India, Snowflake, the answer begins with getting the data foundation right, while addressing governance, accountability and the economics of increasingly complex AI workloads.

What are the biggest technical and organisational bottlenecks preventing Indian enterprises from moving AI projects from pilot stages to production at scale?

As Indian enterprises move AI from pilots to production, the biggest bottleneck holding Indian enterprises back from moving AI into production is not the model. It is data. A pilot can succeed on a narrow, clean dataset. However, production must work across fragmented data, legacy applications, and multiple clouds, with security, governance, and consistent business context built in from the start. That is where most programmes stall.

The first challenge is data. For example, across legacy estates in banking, manufacturing, and the public sector, data remains siloed and uneven in quality. With thin integration and monitoring, a successful demo struggles to become a reliable process, and the gap widens when DPDP compliance, data residency, and cost controls are added after the pilot rather than designed in.

The second challenge is ownership. When AI is treated as an IT project, business, data, and security teams never share accountability for outcomes, risk, and cost, while skills and change management lag.

The way forward is clear: get the data foundation right, and scale will follow.

As enterprises deploy agentic AI systems that can make decisions and execute tasks autonomously, how must their data architectures evolve to ensure reliability, security and accountability?

Greater agent autonomy demands a stronger data foundation. Enterprise data architectures were built for people running queries and reading dashboards. Agents now read, decide and act independently, so governance must move from the dashboard into every interaction.

Reliability starts with the data itself. Agents need fresh, high-quality data, a shared semantic layer so business terms mean the same thing everywhere, and lineage that traces every decision back to its source.

Security and accountability follow from that foundation. Enterprises must know which agent requested an action, what data and tools it could access, what action it took and what it cost.

We built Cortex AI Gateway for exactly this. It provides a common layer for governing how first-party and third-party agents interact with models, data, MCP servers and enterprise tools, with visibility into agent activity and AI consumption.

With our latest features, Snowflake is extending Snowflake Horizon Catalog, its unified governance foundation, to autonomous agents, putting context, security, and governance at the foundation of every agent instead of adding them after the fact. Every answer and action follows one set of policies, regardless of which cloud, catalog, or engine the data sits in.

As AI spans multiple models, agents and clouds, agents should have the freedom to act within boundaries that are clearly defined and auditable.

With 71% of Indian respondents in Snowflake’s research reporting positive returns from GenAI investments, what metrics should businesses track to determine whether AI is delivering measurable business value?

The 71% figure from the ROI of Gen AI and Agents 2026 report shows that Indian enterprises are turning GenAI investment into returns. The next step is proving those returns with hard metrics, measured against a clear pre-AI baseline.

Businesses should track value at three levels. First are business outcomes, such as revenue uplift, conversion, resolution times or cycle-time reduction. Second is operational performance, including accuracy, reliability and the share of tasks completed without escalation. Third is AI economics: the compute, model consumption and human effort required for each successful outcome.

That last measure matters more with agents, because a single task can trigger multiple model calls and actions. Cortex AI Gateway, a part of Snowflake Horizon Catalog, brings every agent under a unified set of controls, giving enterprises the visibility to secure and govern AI connections, track usage, and keep AI spending in check.

At Snowflake, we call this intelligence efficiency: how effectively an organisation converts compute, models, data and context into measurable business impact.

How can enterprises ensure that their AI systems are built on accurate, governed and interoperable data, particularly when information is spread across multiple cloud platforms and legacy systems?

There is no effective AI strategy without a data strategy. Enterprises need a consistent, governed layer of data and business context across every cloud platform, application and legacy system.

Accuracy starts with strong data quality controls, consistent business definitions and lineage that shows where every number came from. Governance must travel with the data, with the same access policies applying wherever it resides.

Open standards, such as Apache Iceberg™ tables, let organisations work across environments without creating new silos, while legacy estates can be modernised step by step rather than migrated all at once.

Snowflake Business Ontology grounds AI agents in definitions the business has already approved. Teams that own a concept, like net revenue, set its official definition and link it to the data that produces it. Snowflake CoCo and Snowflake CoWork use those approved definitions to ground their answers, and every change is logged, so teams can trace how a number was defined and who signed off.

As AI adoption increases, how can organisations balance the rising costs of model inference and complex workloads with the need for performance, speed and accuracy?

Controlling AI costs starts with being deliberate about AI economics. The most capable model is not automatically the right choice for every task, and organisations should define what each workload truly needs, whether speed, cost, reasoning depth or accuracy, before selecting a model.

Data is an equally important lever. Grounded, governed context means fewer tokens, fewer retries and fewer wrong answers. Running AI where data already resides also avoids the cost and risk of moving it.

Visibility matters too. Teams should track inference cost, latency, token consumption and output quality across workloads and teams, and set clear budgets before usage scales.

Accuracy must be protected as costs fall. Testing models continuously against real business tasks ensures that savings never come at the expense of results.

As agentic workflows grow more complex, managing this balance will determine whether enterprises can scale AI sustainably. Cost discipline and performance are not competing goals; with the right data foundation, they reinforce each other.

How does Snowflake’s Dynamic Model Routing approach address the challenge of selecting the right AI model for different enterprise workloads, and what implications could this have for the economics of agentic AI?

Agents execute multi-step workflows that involve repeated model calls, so choosing the right model at each step shapes both cost and performance.

This is what makes a platform like Snowflake stand out. Dynamic Model Routing through Snowflake's Cortex AI Gateway matches each request to the model best suited to it, using a faster, lower-cost model for classification or summarisation and reserving more capable models for complex reasoning. This gives enterprises more flexibility than standardising every workload on a single model. With visibility into latency, token use and spend, teams can measure every routing decision.

The economic implications are significant. Savings compound across each step of a workflow, budgets become more predictable, and enterprises avoid lock-in to any one model provider. The question shifts from “Which single model should we use?” to “What level of intelligence does each task require, and what is the right cost for that outcome?”

Snowflake is the agentic control plane for enterprises. We help enterprises orchestrate models and agents with trusted data, strong governance and cost discipline making them best positioned to scale agentic AI sustainably.