Army employees who were urged to use generative AI at work are now being told to slow down because the service has burned through a shared token supply, according to WIRED, which reviewed internal emails sent to members of the Army’s Combat Capabilities Development Command, or DEVCOM.
The problem is mundane and revealing: generative AI is metered. Every prompt and response consumes tokens, the billing and capacity unit used by large language model services. The Army’s rollout appears to have treated that meter like a vibes-based accessory until the billable resource became a constraint.
According to WIRED, a DEVCOM email said the Army chief information officer had announced unlimited token access in May 2026, but that by mid-June the CIO’s token pool had been depleted and limits had to be put back in place. The email said the Army had renewed usage at current levels, while leaving unclear whether the pool would be renewed after Oct. 1.
The Army and the Department of Defense did not respond to WIRED’s requests for comment. Ask Sage, the vendor whose platform the Army uses, also did not respond, according to WIRED.
What the Army is using
The Army’s generative AI workspace runs on Ask Sage, a multimodal platform that lets users access different large language models, including Alphabet’s Gemini, Meta’s Llama, and OpenAI’s ChatGPT, according to WIRED. The Army has said on its website that Ask Sage supports an enterprise LLM workspace and can be used for tasks such as reclassifying personnel descriptions.
Ask Sage says its defense product is accredited for Controlled Unclassified Information. WIRED also reported that the Defense Department’s Chief Digital and AI Office uses Ask Sage for acquisitions.
An Army employee who spoke with WIRED anonymously because they were not authorized to talk to the press said the service had encouraged workers to use generative AI. Emails viewed by WIRED showed employees received at least 200,000 tokens a month, with more automatically added if they used up the first allotment. Workers who had signed up for Ask Sage but were not using it much received messages nudging them to consume more of their allocation, WIRED reported.
The Army’s subscription included access to 100,000,000 tokens through an annual enterprise pack, according to WIRED. For Ask Sage, one token equals about 3.7 characters, based on documents WIRED reviewed. That makes tokens less like a magic AI ration and more like cloud compute with a smaller costume.
The Pentagon’s AI appetite is broader than office work
The token crunch comes after Business Insider reported that nearly half of the Defense Department’s 3.5 million employees were using AI at work. It also comes as the Pentagon keeps expanding AI use beyond administrative chores.
Breaking Defense reported that the Defense Department used about 20 billion tokens per day during the 38-day Operation Epic Fury in Iran. WIRED said it is unclear whether ordinary Defense Department employees and users working with classified or secret material draw from the same token pool.
The Intercept reported Monday that the Pentagon has reduced staffing at the Civilian Protection Center of Excellence, whose work includes preventing civilian casualties in conflict zones, while developing an AI tool intended to speed up assessments previously handled by that staff.
The Army is not alone in discovering that “use more AI” can become “please stop using so much AI” when token budgets meet employee enthusiasm. The New York Times reported that Meta removed an internal token-use leaderboard after encouraging employees to maximize usage. TechCrunch reported that Instagram head Adam Mosseri discussed possible token caps for Meta engineers. Fortune reported that Uber engineers used a year’s worth of generative AI tokens in four months.
The anonymous Army employee told WIRED the tools had not been especially useful for their work and described them as unreliable, including one model that claimed to have completed a task it had not done. The employee said some federal bureaucracy could benefit from the technology, but that careless deployment would not produce a reliable or efficient rollout.
This story draws on original reporting from WIRED.
