The Army was sold broad access to Ask Sage and then hit the oldest software problem in a new costume: usage is real, and usage costs money.
The Army told some users to slow down on Ask Sage after they burned through artificial intelligence tokens far faster than planned. According to Wired, members of the Army's Combat Capabilities Development Command received an email saying the Army CIO had announced unlimited tokens in May 2026, but by mid-June the Army CIO pool had been exhausted and limits had to come back.
That's the story. Not a theory. Not a gloomy forecast about enterprise AI economics. A real customer, a real deployment, and a real email telling people that the supposedly open tap had a meter on it after all.
The scale makes it more awkward. Wired reported that the Army had access to 100,000,000 tokens through an annual Ask Sage enterprise pack. Breaking Defense reported in June 2025 that the Army Enterprise Large Language Model Workspace had officially launched in May with a five-year, $49 million Ask Sage contract, and that then-Army CIO Leonel Garciga said the rollout had gone from about zero users to roughly 19,000 in less than 45 days.
The math was never friendly.
Ask Sage's own June 18, 2025 press release said its $10 million first-year partnership with the Defense Department's Chief Digital and Artificial Intelligence Office and the Army would provide unlimited access for Combatant Commands, the Joint Staff, and the Office of the Secretary of Defense. That word now does a lot of work. In software, unlimited often means unlimited until a finance model notices what customers are actually doing.
The Army noticed quickly. Wired said employees had been given at least 200,000 tokens a month, with more allocated automatically when they used the first batch. Users who weren't active even received reminders nudging them to use more, according to emails Wired reviewed. You can't encourage people to consume a metered resource and then act surprised when the meter spins.
The promise broke before the product did #
This isn't really an argument about whether soldiers and civilian staff should use generative AI. The Army's own public CIO page describes Ask Sage as the engine behind its Enterprise LLM Workspace and says it supports Controlled Unclassified Information work at Impact Level 5. Wired reported that the platform can route users to models including OpenAI's ChatGPT, Alphabet's Gemini, and Meta's Llama. The Army has paperwork, acquisition tasks, personnel descriptions, and dense internal processes. A language model is useful there.
The problem is the contract story around it. If you tell an organization of that size to use AI broadly, you need to price the thing as if people may believe you. The old SaaS habit doesn't fit cleanly here. A dormant seat in a dashboard app costs very little. A heavy user in an AI workspace can burn real compute every time they paste in documents, ask for revisions, or let a model generate long answers.
Frankly, this is where founders should pay attention. If you're selling enterprise AI on a flat or seat-based plan, the Army just showed you the bad version of success. Adoption is what you wanted. Adoption is also what can wreck your margin when each extra action carries an inference cost.
Garciga had already warned about this before he left the CIO job on May 1, 2026. In January 2025, Breaking Defense quoted him saying that unexpected generative AI bills could force hard conversations when a department has already spent half its cloud budget in the first quarter. That wasn't abstract caution. It was a preview.
Your AI budget needs a hard stop #
The Army isn't alone. TechCrunch reported in June 2026 that Uber capped employee spending on agentic coding tools after burning through its AI budget in four months, with Bloomberg reporting a $1,500 monthly cap per employee and per tool for products such as Claude Code and Cursor. Wired also pointed to Uber and Meta as examples of companies pulling back after pushing employees to use more tokens.
The FinOps Foundation's 2026 State of FinOps report gives you the broader picture. 98% of respondents now manage AI spend - up from 63% in 2025 and just 31% in 2024. That's a shift fast enough to make most finance teams nervous. And the top requested tooling capability, according to the same report, was granular monitoring of AI spend: tokens, LLM requests, GPU use, all of it.
That tells you where the market is going. The first phase was access. The next phase is control.
Ask Sage still has a useful model-routing argument. A platform that can send simpler work to cheaper models has a real advantage over a single-model rollout where every request hits the expensive option. But routing isn't magic. If the pool is finite and the incentives point toward more usage, users will empty it faster than a procurement deck expects.
Don't miss the simple lesson. If you buy enterprise AI, insist on usage dashboards, team-level budgets, alerts, and hard caps before launch, not after the first bad email. If you sell it, stop hiding behind the word unlimited unless you're prepared for customers to behave as if it means what it says.
The Army's public embarrassment is useful because it makes the invisible bill visible. The product may still be worth using. The rollout still needed a better meter.
Also read: Glow raises $180 million at a $1.2 billion valuation to secure the AI agents that CrowdStrike wasn't built to see • Alphabet reports Q2 2026 earnings today with its $190 billion AI bet under the microscope • Computer science enrollment falls for the first time in 20 years as AI reshapes who gets hired