Selected companies tighten enterprise AI cost controls
Samsara, Docusign, Yum, Cigna and Compass are using caps, budgets and model choices to manage rising AI use costs.
By Sofia Marchetti · World Affairs Correspondent
3 min read
Selected companies are tightening enterprise AI cost controls after opening AI tools to broad groups of employees. The measures described by Fortune range from daily spending tracking and departmental budgets to limits on access and using less costly models for work that does not need advanced reasoning.
The approach is aimed at managing demand rather than ending AI use. Fortune reported that Samsara has approved Claude, Gemini, ChatGPT and Cursor, while building a system to monitor trackable AI expenses each day. Chief Information Officer Stephen Franchetti said the company capped use for some nontechnical employees, while research-and-development teams have more room to experiment because they need AI for intensive coding and data analysis.
How are companies controlling enterprise AI costs?
Cost forecasting can be difficult because a task’s model choice, token use and chain of model calls may vary, according to McKinsey. Its May survey included 120 enterprise participants, with 75 qualified respondents across five industries; 93% of those respondents said they had exceeded their AI budgets. That finding describes the survey group, not all companies.
Gartner analyst Arunasree Cheparthi told CIO Dive that companies can set hard consumption limits, segment budgets by department or application, and automatically throttle access or require approval once a limit is reached. Gartner also pointed to multivendor routing: assigning routine work to lower-cost or open-source models and reserving premium models for more complex requests.
Docusign made a more targeted change to its coding agents. Chief Technology Officer Sagnik Nandy told Fortune that the agents had been set to draw on the company’s entire codebase before completing a task, raising token use. He said Docusign changed the default so agents retrieve context only for the developer’s narrow task, reducing token consumption by almost half.
At Yum Brands, Chief Digital and Technology Officer Jim Dausch said token use and related costs had been rising, though token use was not yet a material figure. He estimated that perhaps as much as 95% of tasks sent to AI could be handled by more basic, cheaper models. Fortune reported that Yum has expanded training on model selection and urged business leaders to manage digital spending in a manner similar to department headcount budgets.
Why keep access to higher-cost AI models?
The reported controls do not treat all AI tasks alike. Cigna has authorized more than 70 models for internal work and lets employees use smaller language models or earlier, cheaper versions when a task does not require substantial reasoning, Chief Data, Digital and AI Officer Katya Andresen told Fortune. Andresen said Cigna’s computing and token use had risen, but that total spending was growing less quickly; she attributed that result to the company’s multimodel approach.
Compass has set a budget for each engineer after testing AI coding products and selecting Anthropic and Google partnerships with financial limits, CTO Shay Artzi told Fortune. The examples point to a common operating question for companies adopting AI: make higher-cost capacity available for work that needs it, while putting visibility and spending boundaries around wider use.
This story draws on original reporting from Fortune.