- AI agents consume approximately 50 times the computing power of simple chatbot prompts (per Goldman Sachs), and Goldman predicts AI agents will increase total token consumption 24 times over the next four years — with business AI agents increasing consumption 55 times by 2040; even as model prices fell roughly 50% from December 2024 to December 2025, tokens consumed grew 4.5 times in the same window (per Bain).
- Bristol-Myers Squibb’s chief digital and technology officer Greg Meyers told the WSJ the company is preparing its CFO and board for “pretty high token consumption” that will be “orders of magnitude higher than what we spend today” — a sign enterprise AI cost awareness is moving from IT departments to the C-suite and boardroom.
- The dominant cost-control playbook borrows from cloud-era FinOps: real-time usage dashboards by department and manager (Smartsheet), monthly token reports to the CFO and CTO (Priceline), “chargeback” models that attribute AI spend to business units (OpenText reports this alone reduces costs 20-30%), and automated alerts before employees hit token limits.
- Companies are also substituting smaller, open-source, and on-device models for routine tasks instead of defaulting to expensive frontier models — Qualcomm runs AI on its own hardware to save costs, Lowe’s uses smaller models to prevent “token wastage,” and Principal Financial designs for model flexibility so higher usage doesn’t automatically mean higher spend.
What Happened?
Corporate CIOs are confronting a new cost problem: AI is billed by usage, all employees have access, and AI agents — which complete multi-step tasks autonomously — consume vastly more computing tokens than simple chatbots. Priceline’s senior IT finance director described the dynamic bluntly: “With AI, you’re putting the credit card in the hands of the end user.” Companies are responding by transplanting the FinOps discipline developed during the cloud computing era — real-time spend visibility, departmental accountability, and model tiering — into their AI governance frameworks. CIOs from Smartsheet, Qualcomm, Principal Financial, Bristol-Myers Squibb, Lowe’s, and OpenText all described variations of the same playbook: visibility, accountability, and model selection optimization.
Why It Matters?
The AI token cost problem has a structural twist that makes it harder to manage than cloud costs: the shift from prompt-based AI to always-on autonomous agents means consumption is no longer discrete and predictable — agents run continuously, interact with other agents, and rack up tokens over extended periods on tasks employees may not even be monitoring. Goldman Sachs estimates agent-driven token consumption will increase 55 times by 2040. Meanwhile, the counterintuitive lesson from the past year is that cheaper tokens don’t mean lower bills: as prices fell 50%, consumption grew 4.5x. The cost problem is adoption-driven, not price-driven — and adoption is accelerating, not slowing. This means the governance and FinOps infrastructure companies are building now will determine whether AI delivers the ROI executives are promising boards, or whether AI costs become a drag on operating margins.
What’s Next?
The critical near-term development is whether AI agent costs become a material line item in corporate earnings discussions — the point at which AI spend transitions from a growth investment to a cost management challenge. Companies like CVS Health are already hiring dedicated AI Ops and GPU cost governance executives, suggesting the function is institutionalizing. The “chargeback” model — where business units bear the cost of their AI consumption and are accountable for the business outcomes — is emerging as the most rigorous governance approach, and its adoption across large enterprises will likely be a leading indicator of which companies are building AI operations that can scale sustainably versus those running up tabs they’ll eventually have to explain to shareholders.
Source: The Wall Street Journal
