The AI pilot era is ending. Now enterprises are worrying about the bill, says Snowflake

As AI moves from pilots to production, Indian enterprises are beginning to ask a familiar question: “Kitna deti hai?” Only this time, it is about how much business value they are getting for every rupee spent on AI.

Snowflake is seeing companies pay much closer attention to inference costs, token consumption and returns on AI investments as projects move beyond pilots and are deployed across more employees, customers and business processes, according to Vijayant Rai, Managing Director, Snowflake India.

“The piece which obviously now worries most customers is the cost side of things. How many tokens am I going to utilise for a particular use case? Is it viable? The budgets for it? We are seeing that customers are getting more conscious of that,” Rai told the Economic Times Digital in an interview.


Also read: India a key growth engine for Snowflake: CEO Sridhar Ramaswamy

In the last few years, enterprises have moved from the initial excitement around generative AI to experiments and more recently to larger production deployments.

Rai said Snowflake is now seeing “some really large scale adoption” across its global and Indian customer base. Snowflake has around 13,900 customers globally, with Rai saying about 13,600 were already using AI features on its platform.

However, as adoption widens, AI economics become harder to ignore.

“In most cases, there have been sporadic use cases that have gone into production. As long as they are giving sales some advantage, whether through customer adoption or acquiring new customers, it’s a fairly easy conversation,” Rai said. “But the moment the scale becomes bigger, you’ve got to start thinking about the cost side of things,” he added.

Cost of using the best model for everything

That shift is also behind Snowflake’s latest push around dynamic model routing. Earlier this week, the company announced dynamic model routing within Cortex AI Gateway, which can select an AI model for a particular task based on factors such as cost and performance. The feature is also being integrated into Snowflake CoCo, its coding agent, and CoWork, its agent for business users.

Also read: India’s AI edge will come from talent, not compute, says Snowflake CEO

“You might not want to use a frontier LLM for a use case that is simple, large-scale and repetitive,” Rai said. “You might want to use a lower-cost model that can still do the job,” he added.

Snowflake is also expanding access to open models, including DeepSeek-V4-Flash 0731 and GLM-5.3, alongside models from providers such as OpenAI, Anthropic, Google, Meta and Mistral.

Rangarajan Srirangam, Senior Regional Vice President of Solution Engineering, Snowflake India said model routing is intended to give customers another option rather than take model choice away from them.

“If you strongly feel that one particular model is the best choice, go for it. But if you think the system can make a better decision, you can let it choose,”Srirangam told the Economic Times Digital .

Snowflake said its internal tests showed dynamic model routing improved token efficiency by up to three times in one workload and by 25% in another, though Srirangam said those gains should not be read as equivalent cost savings for every enterprise.

“What you really need to do is take a representative set of use cases, maybe your top 50, run them with and without model routing, compare the token costs, and then make a decision,” he said.

AI's spending problem

As AI moves deeper into day-to-day operations, enterprises are beginning to treat spending controls as part of the deployment itself instead of waiting for the costs to rise.

Rai said companies are taking different approaches to funding AI. Some are carving out money from existing technology budgets, while others are adding fresh spending.

Also read: Enterprises using AI as returns outweigh costs: Snowflake executive

The willingness to spend, however, also depends on whether companies can show a return. Rai said a recent Snowflake survey of around 2,000 customers globally found that 71% of Indian respondents reported positive return on investment from AI, compared with 61% globally.

“As long as there is ROI, I think the cost is justified because it helps an organisation get ahead in different parts of its business,” Rai said.

But the equation becomes more complicated as AI is opened up to more employees, applications and autonomous agents. Enterprises then need to know not just how much they are spending overall, but which users, models and workloads are driving that consumption.

“For us, a lot of this cost conversation is leading to controls on cost,” Srirangam said. “At what granularity can you control costs and what limits can you set and what kind of visibility can you get to spending?”

He said enterprises are increasingly looking at controls such as per-user and service-level limits, budgets, usage tracking and auditing.

Srirangam said that there is also the risk that a simple request can trigger significant computation behind the scenes.

“Then there are other things to worry about, such as runaway spending,” Srirangam added.

Education, silos and context

Cost, however, is only one reason AI projects struggle to scale. Srirangam said the first barrier he sees is often not compute or access to models, but education.

“I think the primary hurdle is education. People need to understand how to use AI and how to get the most value from it,” he said.

Beyond that, Srirangam said enterprises continue to struggle with fragmented data and inconsistent business definitions.

Also read: Timing, not talent, holding back Snowflake’s India R&D expansion: Cofounder Benoît Dageville

“If you get past education, the next two problems are silos and lack of context,” he said.

Rai gave the example of a sales team considering a deal closed when a contract is signed, while finance may consider the same deal closed only once payment or consumption starts.

“If you do not have the semantic sort of clearly defined, the context will go wrong completely,” he said.