The paradox every AI budget is hitting in 2026
AI has never been cheaper per unit, and enterprise AI bills have never been harder to control. That contradiction is the defining problem of agentic AI costs this year, and it now sits on both the CFO’s and the CIO’s desk.
The unit price is collapsing. Ramp’s enterprise spending data shows the average price of a million tokens dropped from roughly $10 to $2.50 in a single year, according to CRN Asia. Yet McKinsey research from July 2026, summarized on HackerNoon, found that 93% of enterprises running agentic AI are over their AI budgets, and one in five has already restricted usage because of cost.
Both facts are true at once. They measure different things. The price sheet measures the cost of a token, while the invoice measures how many tokens your organization burned to get work done. AI agents have changed the second number far more than vendors have changed the first.
The story is already in the headlines. Uber reportedly exhausted its entire 2026 AI budget by April, as CloudZero recounts, and Microsoft reportedly pulled most internal licenses for one agentic coding assistant over runaway token bills. Neither case was a technology failure. Both were governance failures.
Why do AI agents make AI costs rise?
Agents consume tokens per task, not per question, and a single task can trigger dozens of model calls. A chatbot answers once. An agent plans, calls tools, checks its own work, and retries when a step fails.
Gartner’s March 2026 forecast puts agentic workloads at 5 to 30 times the tokens per task of a standard chatbot exchange (HackerNoon). Three mechanics drive that multiplier.
- Chained reasoning. Agents can launch sub-agents, recursive calls, and branching logic, so cost grows geometrically rather than in a straight line with users (elvex).
- Quality loops. McKinsey estimates about 60% of an agentic task’s cost goes to refinement, meaning retries, re-prompts, and validation passes. That work is what makes the output reliable enough to use, so it cannot simply be switched off.
- Model defaulting. Teams tend to send every task to the most capable and most expensive model, a habit elvex calls token maxing. A model that was right-sized at launch is often oversized six months later, and few organizations re-benchmark.
For finance leaders the implication is simple. Enterprise AI has moved from a fixed cost you license to a variable cost you meter. EY describes it as a shift from comparing AI against fixed software and labor costs to managing variable compute consumption (EY). Falling unit prices will not save a budget when volume grows faster than price falls, which is exactly what Gartner expects through 2030 (CGI).
What is the true cost of an AI agent?
The token invoice is only the visible part of agentic AI costs, and most business cases stop there. EY illustrates the shift with a customer service assistant whose cost per interaction grew from about $0.04 for a simple chat to roughly $1.20 once tools, planning, and sub-agents were added, around a 30x jump (EY).
EY breaks the total cost of an agent into seven categories and notes that most companies budget only the first three. The rest surface later, in budgets that belong to other executives.
| Cost category | Where it usually shows up | Who typically owns it |
|---|---|---|
| Tokens and API calls | Model vendor invoice, after the work is done | CIO, CTO, business unit |
| Subscriptions and licenses | Several vendor contracts | CIO, procurement |
| Platform infrastructure | Cloud bill, filed as infrastructure | CTO, CIO |
| Governance and compliance | Risk budget, often as headcount | CRO, audit |
| Organizational change | HR, training, consulting budgets | COO, CHRO |
| Failure and recovery | Nowhere, until an incident | CFO |
| Possible future AI taxes | Not yet a cost, only a regulatory signal | CFO, legal |
Adapted from EY’s agentic FinOps framework.
This is why neither executive can solve the problem alone. The CIO sees the token and infrastructure lines. The CFO carries the risk, the forecast, and the board conversation. Without a shared view, every agent looks cheaper on paper than it is in practice.
Why traditional IT budgeting fails for agentic AI
Annual budgets assume costs you can predict, and agent consumption is neither predictable nor visible until the invoice arrives. Four structural gaps explain why.
The bill arrives after the damage. Token spend is usually known at month end, once the work is done. Quali puts it well, comparing billing dashboards to reading a factory’s electricity bill to understand its production costs (Quali). Monitoring and alerts worked for cloud. Agents need limits enforced where the spend is created.
Subsidized prices are ending. EY notes that some compute costs have been absorbed or strategically priced by upstream providers, and that many vendors are moving from subscriptions to consumption pricing. Token prices also depend on a constrained supply chain of chips, power, and data centers, which can show up as usage caps or sudden repricing.
The cost center has moved from building to running. Gartner’s August 2026 forecast found inference spending passed training spending for the first time this year, at 55% of AI-optimized infrastructure spend (HackerNoon). Run costs recur every month and scale with adoption, the opposite of a one-off project budget.
Nobody can link spend to value. A June 2026 CloudZero survey of 260 finance leaders found only 22% can tie AI spend to business outcomes, while 87% say they need to within a year (CloudZero). Without that link, the only lever left is rationing, which is how Uber ended up out of budget by April.
The good news is that FinOps teams are mobilizing. The FinOps Foundation’s State of FinOps 2026 report shows 98% of practitioners now manage AI costs, up from 63% a year earlier, and AI cost management is the top skill teams want to add (CIO Dive).
A CFO and CIO playbook for governing agentic AI costs
The goal is not to spend less on AI but to make every token accountable to an owner and an outcome. These seven moves can be started within one planning cycle.
- Name a single owner for agent economics. EY recommends one executive accountable for model usage, cost leakage, and value realization across all seven cost lines. In practice this works best as a joint CFO and CIO mandate, with FinOps doing the day-to-day work.
- Put fully loaded cost in every business case. Include infrastructure, governance, change, and human review alongside tokens. An agent that only looks profitable on its token line is not profitable.
- Install circuit breakers before you scale. Set spend ceilings, call-volume caps, and automatic shutoffs at the agent, workflow, and business unit level. Enforce them in an AI gateway so limits apply at runtime rather than in next month’s report (MuleSoft).
- Route every task to the cheapest model that does the job. Reserve premium reasoning models for complex work, use caching, and re-benchmark model assignments every quarter. Unit prices fall fast, so last quarter’s right-sized model is often this quarter’s overspend.
- Allocate spend to teams, products, and agents. Showback first, then chargeback. The FinOps Open Cost and Usage Specification (FOCUS) is adding AI-specific columns that identify the model billed and the person, service, or agent behind each charge (FOCUS).
- Require a value metric on day one. No agent gets approved, scaled, or renewed without a defined output per dollar, whether revenue, hours saved, cycle time, or risk reduced.
- Renegotiate how you buy AI. Vendors are layering consumption meters onto seat licenses. Push for usage caps, price protection, and visibility into credit burn before renewal, not after (Flexera).
The split of responsibilities is straightforward. The CIO owns the controls: gateways, routing, telemetry, and allocation data. The CFO owns the rules: budgets, ROI thresholds, and the decision on which agents earn the right to scale.
Which AI cost metrics should CFOs and CIOs track?
Stop reporting cost per token and start reporting cost per successful task, broken out by use case. A blended average across all AI activity hides exactly the workloads that are running away.
| Metric | What it reveals | Primary owner |
|---|---|---|
| Cost per successful task | The real unit economics of each agent | CFO and CIO jointly |
| Tokens per task, by step | Where planning, tool calls, and retries inflate usage | CIO |
| Retry and refinement rate | How much spend goes to quality loops | CIO |
| Days since last model re-benchmark | Whether workloads run on oversized models | CIO |
| Share of AI spend allocated to an owner | How much spend is still unaccountable | FinOps |
| Spend versus budget, by agent and business unit | Early warning before a budget runs out | CFO |
| Value per dollar of AI spend | Whether an agent should scale, change, or stop | CFO |
The first four metrics draw on the measurement framework proposed on HackerNoon. None of them come from a vendor price sheet. All of them come from instrumenting your own production traffic.
Frequently asked questions about agentic AI costs
Why are AI costs rising if token prices are falling? Because agents use far more tokens per task. Gartner estimates agentic workloads consume 5 to 30 times the tokens of a standard chatbot exchange, so volume growth outpaces price declines.
What are agentic AI costs? Agentic AI costs are the total cost of running AI agents, including tokens, licenses, infrastructure, governance, organizational change, and the risk of failures. Tokens are the most visible part but rarely the largest.
How should companies budget for AI agents? Budget per task and per use case rather than per year, include fully loaded costs in the business case, and set hard spend ceilings at the agent and business unit level.
Who should own AI cost governance, the CFO or the CIO? Both. The CIO owns the technical controls and cost data, while the CFO owns budgets, ROI thresholds, and scaling decisions. FinOps teams typically connect the two.
What is FinOps for AI? FinOps for AI applies cloud financial management practices to AI spend, covering tracking, allocation, forecasting, optimization, and linking spend to business value.
How can enterprises reduce AI token costs without hurting quality? Route simple tasks to smaller models, cache repeated prompts, limit retries, and re-benchmark model choices regularly. Measure cost per successful task so savings do not come at the expense of rework.
The bottom line for finance and technology leaders
Cheaper tokens will not fix agentic AI costs, because the problem is volume and visibility, not price. The organizations that pull ahead in 2027 will not be the ones that spend the least on AI. They will be the ones that can show, agent by agent, what each dollar produced.
That requires the CFO and the CIO to govern AI together, treating agent capacity like capital that must earn its place. Start with one owner, one metric, and one set of circuit breakers, then scale only what proves its value.