Inside Okta’s strategy for managing AI emissions 

Identity management software firm Okta adopted a sustainable strategy in April that guides how the company manages the artificial intelligence software used by its employees.

Work on the strategy started two years ago, initiated as part of a broader evaluation of emerging technologies and spurred by questions from employees. 

The program is led by Okta’s sustainability team, along with representatives from the firm’s engineering and IT teams, said Alison Colwell, senior director of sustainability and responsible technology.

“This partnership ensures that sustainability isn’t just a separate report, but is integrated into how we manage our technology relationships,” she said. For example, Okta now asks strategic AI vendors to disclose their emissions where possible.

The company also sought advice and peer reviews from experts associated with the Business Council on Climate Change, which has an AI and sustainability working group.

“Currently, the program focuses on our internal use of top enterprise AI tools, and these frameworks set the stage for how we think about sustainability across the organization,” Colwell said.

DIY measurement

Okta developed its own methodology based on publicly available data to measure the energy consumption of its most widely used AI software tools, so that it can purchase renewable energy certificates to match that usage.

For example, the company evaluated how many lines of code in a given application or product feature were generated using AI. It then used that information to calculate the number of “tokens,” or blocks of data, that the AI model used to produce that output. The carbon footprint is extrapolated from those calculations. Factors considered include:

• Additional AI model training needed for specific purposes
• Where employees using AI are located
• Data center locations, based on the AI vendor

Okta considers this to be an interim approach that it will use until industry standards emerge, and it’s sharing it to accelerate that process. One lesson was how hard it is to get granular data from AI vendors. Another takeaway: the need to consider AI productivity tools used for day-to-day efficiencies differently than AI services meant for bespoke applications.

“The measurement of these two types of AI deployment have different levels of control, and thus different decarbonization levers that can be used to reduce the environmental impact,” Colwell said.

Employee usage guide   

Okta’s sustainability team collaborated with its engineering and IT teams to compile best practices into a resource its workforce should use when deciding whether to use AI to complete a task. It suggests:

• Limiting AI use only to projects where it can truly have a positive efficiency impact
• Choosing the “lightest” model for the job using energy consumption data

Okta uses a resource called OpenWebUI, which ranks different AI tools by energy efficiency and compute power requirements, and the sustainability team created an agent that suggests ways employees can adjust their choices when it detects there are better options available.

“This occurs right at the point of use when Okta employees are promoting the AI models,” Colwell said.

Renewable energy offsets

Okta plans to estimate energy consumption specific to its AI usage on an annual basis. Based on its initial measurements, Okta expanded its purchases of renewable energy certificates to cover the additional energy needed to run the top two third-party AI tools used by the company.

It does not plan to include those purchases or related emissions reductions in its overall greenhouse gas emissions inventory, because the methodology is still nascent, Colwell said.

“Looking ahead, I hope to move toward more frequent, real-time insights that can inform more agile decision-making and efficiency improvements,” she said.

Okta’s sustainability team will discuss additional details of the company’s strategy as part of the AI x Sustainability program at Trellis Impact 26, from June 23 to 25 in San Francisco.