5 Ways Companies Keep AI Bills in Check

Snowflake CEO Sridhar Ramaswamy on Monday became the latest executive to voice concerns over rising AI costs.

“Are we worried about how much we are spending on AI inference across our internal teams? Absolutely,” he told my colleague Laura during Snowflake’s annual conference in San Francisco. (Snowflake is a major customer of Anthropic, OpenAI and other AI model developers.)

But Snowflake and other companies are also finding ways to save money. As debate rages over whether companies are getting enough of a return on their AI spending, especially as AI providers raise prices, some customers are describing the tricks they use to bring down these costs.

Here are five of the most common cost-saving measures we’ve heard about:

1. Using model routers

Not every AI task requires the most expensive model. Basic chores like summarizing emails or searching through documents can often run on open source models for a fraction of the cost of cutting-edge models, software executives tell us. Firms like Snowflake and Palo Alto Networks have told us they found cost savings by swapping in cheaper models for certain tasks.

Those companies and many others have built their own software known as routers that can choose lower-cost AI models based on the command employees type in internal chatbots or coding apps (which the companies also developed in-house). A founder of a construction AI startup, for instance, told us it was relatively simple to get Anthropic’s Claude models to create code for a bespoke router for the startup. The founder said the Claude-created router does not give special preference to Claude models when deciding which AI should handle which tasks. 

Other customers have used similar routing tools from providers like OpenRouter.

Snowflake’s router in its AI coding tool CoCo, which it also sells to customers, can help lower costs for administrative tasks such as creating reports and slideshow presentations, said Christian Kleinerman, an executive vice president.

2. Prompt engineering

Sometimes, reducing AI costs can be as simple as telling an AI model to do less thinking. Enterprise software firm UiPath has been using prompt engineering to minimize how much models spend “warming up” or “thinking” before doing a task, especially tasks they’ve done before, which led to more than 90% cost savings for some tasks, chief information security officer Scott Roberts said. 

The concept of “prompt engineering”—figuring out the most effective way to give commands to the AI—has been around for years, and public sources of information about that abound, including useful guides that OpenAI, Anthropic, Google, and other model makers have published.

3. Controlling how much AI employees use

As horror stories circulate about runaway AI bills that shock corporate chief information officers, some CIOs are preempting such cases by limiting how much AI employees can use in a given week or month. In some cases, that means setting token ceilings that employees can’t go above without getting approval from their managers. 

Some companies are also controlling which types of employees get access to the most advanced (and expensive) models and which ones don’t. For instance, it may make sense to give software engineers access to OpenAI’s Codex, but not salespeople. That’s precisely what security firm Zscaler does, according to executive vice president Dhawal Sharma.

“Using a very large model to solve a simple problem is a misuse of resources,” Sharma said. “When my legal or marketing team is doing something with AI, they're using lighter models than a software engineer—it’s about governance tied to roles and personas.”

4. Knowing When Not to Use AI 

In a dose of operational reality, some companies are saving money by actively restricting AI usage when traditional, structured software is cheaper. 

For instance, when pharmaceutical giant Novo Nordisk recently analyzed past clinical trial data using Anthropic’s Claude, executives realized the standard version of Microsoft Excel—even without Microsoft’s Copilot AI features—was both cheaper and more reliable. 

“If I can do it better and cheaper and more reliably in Excel, I’m going to tell you to stay in Excel,” chief transformation officer Stephanie Bova said. (The company is still using Claude heavily for other tasks.)

5. Pushing For Discounts Or Escape Hatches

This will seem obvious, but the bigger the customer is, the more leverage they have to get discounts off the sticker price of AI from cloud providers such as Microsoft or Amazon Web Services or from OpenAI.

Snack food giant Mars, for example, selected Google’s Gemini model as its AI chatbot of choice for roughly 62,000 employees because Google agreed to let it pay a flat per-seat fee for most of those employees, rather than usage-based pricing, which can be pricier.

Customers also see an opportunity not to get locked in to any single vendor, which means they can save on cost in the long run. Some companies such as Ralliant and PagerDuty are pushing for shorter-term, one-year deals with Salesforce and other AI providers so they can cut them loose if they need to, depending on how efficient the AI ends up being.

New From Our Reporters

Upcoming Events