Agents Don't Need Seats. But They Still Cost Money.

OpenAI moved workspace agents to token-based credit pricing on July 6. Cost now scales with what the agent actually does, not whether it was invoked. That sounds like progress. It is also a new category of financial risk that most enterprise procurement teams have not priced in.

The $234B SaaS-at-risk number gets quoted as a vendor problem, and it is one. But the other side of that equation doesn't get discussed: when your agents replace seat-licensed users, the seat cost goes away. The agent's token spend is now variable, usage-driven, and almost completely invisible to the budget owners who approved the AI initiative in the first place.

I've had this conversation with finance leaders at our clients. They approved a headcount-equivalent budget for AI tooling. Clean line item. No one told them the agent would run autonomously at 3 am processing a batch job, and that the token bill from that one run would exceed the monthly seat cost it replaced. That gap is not theoretical anymore. Our teams are in these environments. We are watching it happen... in real time.

FinOps teams are already feeling this: 98% now manage AI spend as a priority. But most enterprise procurement and legal functions haven't caught up to what token-based pricing actually means at agent scale. You've traded a predictable cost model (seats) for an unpredictable one (token consumption that scales with agent autonomy and usage). Governance needs to cover the spend model, not just the behavior.

Has your organization built a financial governance model for agent token spend? Not a budget line. A model: who owns the forecast, who gets alerted when consumption spikes, who can throttle a runaway agent at any time? If that accountability doesn't exist, you don't have an AI governance program. You have an AI experiment with an open tab.