Summary
- Agentic AI can incur costs through multiple autonomous model calls, tool invocations, retries, and escalations.
- Grossman argues that cost per successful business outcome is a better measure than cost per token.
- He advocates named workflow owners, real-time consumption telemetry, spending limits, and approval controls.
New Gartner research puts a number on something finance leaders have been feeling for months: as much as $234 billion of enterprise software spend is now at risk as agentic AI reshapes how technology gets bought and consumed. The evidence is already public. Uber burned through its entire 2026 AI budget in four months. Accenture has told staff to stop using AI for unnecessary tasks as costs surged. These are perfect examples of what happens when a new spending model outruns the finance function’s ability to see it.
The shift is structural. Traditional software cost is initiated by a person: someone logs in, runs a report, sends a query. Agentic AI breaks that link. A single request to an autonomous agent can trigger dozens of model calls, pull data from multiple systems, invoke external tools and validate its own output before a human ever sees the result. Every one of those steps has a cost, and none of them requires a person to press a button. Enterprises have become internal token economies, where machines are now the primary drivers of consumption. Most finance functions are still budgeting as if the old rules apply.
That’s why cost per token is becoming a misleading metric. It measures the efficiency of individual model interactions, not the value of the work being completed. An agent might use relatively few tokens on each model call, yet still be an expensive way of solving a problem if it has to retry multiple times or escalate before producing a usable outcome. For some low-risk tasks, several iterations may be perfectly acceptable. For others, such as customer service or operational decision-making, repeated retries quickly erode both cost and productivity. The number that actually matters is cost per successful business outcome: what it took, end to end, to get a correct, usable result. Without that lens, finance teams end up optimising the wrong variable.
A new era for finance leaders
None of this means finance leaders need to become AI engineers. They don’t need to understand model architecture or write a line of code. What they do need is enough AI literacy to ask the right questions of the teams deploying these systems: how many calls does a typical agent workflow generate, what triggers retries and escalations, and who owns the decision when an agent’s cost profile changes overnight because a workflow got more complex. That’s a governance skill, and it’s no different in kind from finance’s traditional role in any other cost centre. It just requires learning a new vocabulary.
The playbook here isn’t new either. Enterprises went through a version of this with early cloud adoption, when workloads scaled faster than anyone had budgeted for and bills arrived as a surprise rather than a forecast. The lesson from that cycle was that visibility and ownership have to be built in before the spend accelerates, not retrofitted afterwards. The same applies now. Every agentic workflow needs a named owner, a budget attached to that owner, and telemetry that shows consumption in near real time rather than at the end of the billing cycle. If finance can’t answer “which agent, which team, which workflow” for a spike in spend, the governance model isn’t ready yet.
In practice, that means embedding budget controls beyond dashboards and directly into the agents themselves. This could look like hard tokens or spend ceilings per agent and per workflow, automatic throttling or human approval gates when a task is about to exceed its budget, and kill switches that stop runaway processes before they compound. It also means treating budget as a design constraint from the start of a project, the same way security and compliance already are, rather than a control that gets added once costs have already become a problem.
Agentic AI changes who, or rather what, is consuming enterprise technology. That’s a fundamental shift for finance. The organisations that treat AI as another software licence will struggle to explain why spend keeps accelerating. The ones that recognise they’re now managing an economy of machine-initiated transactions will build the controls, reporting and accountability needed to keep that growth commercially sustainable.
| About the author | |
|---|---|
Scott Grossman is the Chief Financial Officer at Ensono |  |
