Your AI Budget Is Now A Meter, Not A Licence, And Finance Will Find Out

Daniel Fallmann, founder and chief executive of enterprise search vendor Mindbreeze, argues in Forbes that the urgent problem for companies that have moved past AI pilots is no longer model performance, accuracy or security but the economics of consumption itself ^[1][2]. His claim, and the one number in the piece worth staring at, is that token usage is becoming one of the fastest-growing lines in enterprise technology budgets while most organisations lack the means to manage it ^[3].

The structural change is simple and unforgiving. Traditional enterprise licensing was largely predictable, and per-user costs often fell as adoption spread ^[4]. Generative AI inverts that: every prompt, retrieval, reasoning step and agent action consumes tokens that convert directly into cost, individually trivial and collectively volatile ^[5]. Fallmann describes "AI sprawl", where staff adopt overlapping tools with no central coordination, producing duplicated work, fragmented workflows and rising token consumption without matching productivity ^[6].

Agents change the slope of the curve. An autonomous agent rarely makes a single model call; it retrieves, invokes tools, reasons through intermediate steps, validates its own output and often repeats part of that loop ^[7]. Across thousands of concurrent workflows, the spend starts behaving like cloud infrastructure rather than software: elastic, usage-driven and easy to lose control of ^[8].

There is evidence this is not hypothetical. According to the piece, Axios reported in June 2026 that Databricks shipped enterprise controls specifically to cap AI spending and monitor usage across providers, after some organisations found AI bills reaching tens of millions of dollars a month ^[9]. Taken at the bottom of that range, ten million dollars a month annualises to roughly 120 million dollars a year, which is a capital-allocation decision arriving through an invoice rather than a business case ^[10]. The pattern of uncoordinated adoption driving consumption with little measurable value has acquired a label, "token maxxing" ^[11].

Operators have seen the shape of this before. Cloud made infrastructure trivially easy to provision, usage outran budgets and governance, and FinOps emerged as a discipline because oversight lagged consumption ^[12]. Fallmann's point is that AI runs the same cycle faster, because token consumption scales invisibly once AI is embedded in daily workflows instead of being requested through a visible provisioning step ^[13]. He also reports that many European enterprises expanding deployments are diversifying providers and hardening cost management because autonomous systems consumed far more tokens than projected ^[14].

His remedy is architectural: much unnecessary consumption traces to AI systems without governed access to the right information, which compensate with repeated retrieval, wider context windows and redundant reasoning, so a unified knowledge foundation should cut tokens for the same or better outcomes ^[15][16]. Note the interest. Fallmann sells enterprise search and knowledge management, and the prescribed fix is the product category he runs ^[17]. That does not make the mechanism wrong, but it should be tested against your own logs rather than accepted as diagnosis.

What to watch, in your own house rather than in the commentary. Whether token spend is attributable to a team, a workflow and a business outcome, or arrives as one aggregated vendor line. Whether anyone owns a hard cap and an alerting threshold before the bill, given that vendors are now selling those controls as features ^[9]. Whether duplicate tools are being consolidated on cost grounds ^[6]. And whether the measurement is usage or value, because a falling cost per token is not the same as a falling cost per completed task.