Snowflake executives advocate strategic AI use to control costs and maximize results

In a recent discussion, Snowflake executives Baris Gultekin and Bruno Barreto argued that companies should select AI models based on specific tasks and expected outcomes. The goal is to deploy the most advanced technology only where it delivers real value, while simpler solutions can handle other tasks without increasing expenses.

“A Michelin chef isn’t necessarily the person who chops the onions,” said Bruno Barreto, Snowflake’s vice president for Latin America. His analogy highlights the decision companies face when adopting AI: determining when to invest in advanced technology and when a more affordable option suffices.

Barreto explains that the process begins by breaking it down into distinct tasks. One task may require a model capable of analyzing complex data, while another can be handled by a cheaper solution. The key criterion is the level of capability needed to achieve the desired quality of results.

According to the executive, the chef’s work and the ingredient preparer’s role are part of the same final product but require different skills. Applying this logic to AI means identifying where sophistication adds value and where it merely increases costs. Choosing a model becomes a decision about how to allocate resources throughout the process.

“We don’t need to apply the same model to every case, every agent, or every need,” Barreto asserts.

Cost extends beyond a single response

Part of the cost is tied to tokens, the units models use to divide processed and generated content. Many applications accessed via programming interfaces (APIs) base their billing on this volume.

An agent may need to make multiple calls to a model to query information, trigger tools, evaluate received data, and produce a response. Therefore, the cost of a single query doesn’t necessarily reflect the total cost of completing the task.

Gultekin describes the relationship between capability and cost as “intelligence efficiency.” “Are you paying the right amount for the intelligence you’re receiving?” the executive asks.

The choice should consider what the system needs to do and the level of performance required for that application. “If a question can be answered by a cheaper model, the platform can use that cheaper model instead of the more expensive one,” he says.

“We work with our clients to ensure they start with a business goal in mind and then build AI in the most efficient and cost-effective way possible to achieve that result,” Gultekin explains.

Dividing tasks also allows costs to be distributed. Barreto compares this to a kitchen, where professionals with different skills and costs prepare the same dish. In AI, the choice of models should follow a similar logic.

According to the executive, this choice can include a cheaper open-source model for certain applications, rather than always relying on the most advanced systems. The goal is to control costs without sacrificing the necessary capability for each task.

Snowflake’s platform provides tools to set budgets and direct requests to the appropriate model. However, monitoring consumption doesn’t mean stopping it immediately: the company’s documentation states that certain checks are periodic, allowing a limit to be exceeded before a control action is executed.

When the agent needs to call a person

Gultekin distinguishes between activities that can benefit from language models and those that require a strictly repeatable sequence. The choice of tool, he says, must consider this difference.

“If you want something that is always repeatable, at every step, you don’t necessarily want to use a language model for that,” he says.

In sensitive decisions, human supervision can be part of the process. The executive explains that insufficient results in accuracy evaluations are a reason to delegate certain steps to a person.

For Barreto, this ability to supervise is critical to gaining managers’ trust. Concerns include answers fabricated by models, data protection, and compliance with regulated sector requirements.

“How do I ensure this doesn’t leave our environment? Are we acting in compliance with what is required within a regulated industry?” the executive asks, illustrating clients’ concerns.

The result needs to show up in the business

Within Snowflake, Gultekin cites an agent developed to support the sales team. The tool helps salespeople prepare for meetings, track customer product usage, and learn more about the portfolio.

“It’s not just about making them more productive so we need fewer salespeople. It’s the opposite: they’re being much more effective at what they do,” he says.

In Brazil, Barreto points to Azul as an example. According to him, the airline uses data to predict the behavior of loyalty customers during flight delays and guide customer service, with the goal of retaining these passengers.

The executives did not provide percentage increases in revenue from these projects. However, Barreto notes that conversations with clients have shifted from experimentation to charging for results.

The company has also expanded its structure to meet regional demand. According to Barreto, the operation in Latin America has nearly doubled in about a year and a half, with just over 200 employees across Brazil, Mexico, Colombia, Chile, and Costa Rica.

Before the agents, the data

The path to these results begins before selecting a model. For Gultekin, providing ChatGPT or Claude to employees offers a work tool but doesn’t equate to implementing agents that understand how the organization operates.

“The number of companies that have implemented deeper agents, who truly understand the company’s context and data, is still low,” he says.

Preparation requires more than gathering files or granting system access. It’s necessary to explain to AI how the company defines its indicators, what rules it applies to calculations, and what the terms used by teams mean. This database provides the context for interpreting questions and producing useful answers.

“Every company measures revenue and profit differently,” Gultekin says. “All of their metrics need to be described.”

If departments calculate an indicator differently, the agent needs to know which definition to apply. Having access to records alone doesn’t resolve this difference.

Barreto identifies this gap in companies that try to advance with AI without completing the organization of information. “The foundation to start this is missing,” he says.

Next come the selection of applications, team training, and the transition from testing to production. “First, the foundation with a robust data platform. Second, moving from this experimentation to production,” Barreto says.

For Gultekin, this order defines the starting point. “There is no AI strategy without a data strategy,” he asserts.