It is the end of the quarter, beginning of a new quarter, and in the past 7 days (keeping in mind the July 4th holiday), I have had 3 completely separate and independent conversations about the budgetary impacts of AI spend. To that end, I wanted to get some thoughts out about my perspective on AI spending and token sprawl, and why this issue deserves greater attention.
If your last budget review included a moment of stunned silence over the AI line item, you're not alone. Across engineering, security, and finance organizations, the same story is playing out: teams adopted generative and agentic AI in good faith, and the invoice came back looking nothing like the forecast.
I (and others) choose to call this dynamic Token Sprawl — the uncontrolled growth of AI token consumption that happens when organizations deploy AI systems without a matching discipline around how those tokens get used. It's easy to mistake this for a pricing problem. It isn't. What's rising is volume, and more specifically, unmanaged volume — the kind that compounds quietly in the background until someone finally asks why the bill doesn't look anything like the plan.
Part of what makes Token Sprawl so disruptive is that it doesn't respect org charts. It shows up differently depending on where you sit, but it's really the same underlying issue wearing three different hats.
Engineering and DevOps leaders see it in agentic coding and automation tools, where a single workflow can balloon in cost simply by running longer than anyone anticipated. Security leaders see it in AI-powered detection and investigation tools, where a single deep investigation can consume a startling amount of budget with little warning. And finance leaders see it last, and worst — as a line item that no longer behaves like the predictable, seat-based software spend they're used to forecasting.
Because each group only sees its own slice of the problem, the instinct is to solve it locally: cap a budget here, throttle a tool there. But treating Token Sprawl as three separate problems is exactly why it keeps recurring. It's one problem, viewed from three desks.
Out-of-control AI spending and Token Sprawl isn't really anyone's mistake. It's a structural byproduct of how modern AI systems work. Agentic workflows re-process context at every step. Systems retry and self-correct without anyone watching the meter. Background agents run continuously, whether or not a human ever asked them to. And it's often simpler, at least at first, to default to the most capable model for every task, regardless of whether the task needs it. None of these behaviors are wrong on their own. Together, unmanaged, they add up fast.
The organizations that get this under control aren't the ones with the tightest forecasts — they're the ones that treat token usage as something to actively govern, not just track. That means knowing where consumption is coming from, matching the right tool to the right task instead of defaulting to the most powerful option, and making sure engineering, security, and finance are having this conversation together rather than separately.
Token Sprawl isn't going away on its own. But it is manageable, for organizations willing to treat it as an operational discipline rather than an end-of-quarter surprise.
______________________
This is just an overview. EMA will be releasing a whitepaper that dives deeper in to this subject in the very near future. Stay tuned!