The bill has arrived. After years of AI experimentation, enterprises are now staring down real invoices – and for many, the numbers are deeply uncomfortable.
Gartner’s latest forecast puts worldwide AI spending at $2.67 trillion in 2026, up 49.5% from 2025. Yet McKinsey finds that only 37% of survey respondents attribute any EBIT impact to AI. That gap – between what the market is pouring in and what organizations can connect to earnings – is the defining business challenge of this moment.
For every organization, the question is no longer “should we invest in AI?” It is: “Can we prove this investment is working?”
The Token Economy Is Harder to Manage Than Anyone Predicted
Tokens (and credits) are a core unit of AI consumption and an important driver of cost. Every prompt sent, every document processed and every agent action taken can consume tokens. But tokens are not the whole bill: licences, infrastructure, retrieval, storage, tool calls, monitoring and human review can also contribute. Unlike traditional fixed-price software, much of this spending varies with use, making it harder to forecast.
The volatility is catching enterprises off guard. In an independent February 2026 survey conducted by Sapio Research for DoiT, covering 500 finance leaders at large US and UK organizations, 79% reported AI-related cost overruns in the previous 12 months. A separate July 2026 WitnessAI survey of 300 US business executives found that 68% said at least some AI initiatives had run over budget.
One widely reported example involved Uber’s engineering organization and Claude Code. Adoption expanded rapidly across roughly 5,000 engineers, with agentic-tool usage reaching 84% by March 2026. By April, Uber had reportedly exhausted its full-year AI tooling budget. Reported monthly costs ranged from about $150-$250 for average users to $500-$2,000 for power users. One rollout, four months, a full-year budget.
This is not an outlier story. It is a warning signal.
Agentic AI Is Amplifying the Cost Problem
The shift to agentic AI – systems that can plan, reason and execute multi-step tasks – is making cost management significantly harder. Gartner analysis cited across industry reporting estimates that agentic workflows can consume 5 to 30 times more tokens per task than a standard chatbot query. Separate reporting indicates inference now exceeds training within AI-optimised cloud infrastructure spending; however, the exact share of an individual enterprise’s total AI budget will vary.
Gartner’s June 2026 research predicted that AI coding costs could surpass the average developer’s salary by 2028 as token consumption rises. It also warned “token discipline will not emerge through developer choice alone, as developers tend to optimize for speed and convenience over cost efficiency.”
Meanwhile, Gartner’s July 2026 analysis projected that up to $234 billion in enterprise application software spending could be exposed to “agentic arbitrage” by 2030 – roughly 20% of enterprise software-as-a-service spending – as agents complete tasks across systems and reduce reliance on traditional interfaces. This is a forecast, not an observed loss, but it signals that the economics of enterprise software are being rewritten.
The Visibility Gap: You Cannot Manage What You Cannot See
Here is the core problem that underpins nearly every AI budget blowout: organizations are scaling spend faster than they are building the visibility to understand it.
Only 26% of organizations report full, real-time visibility into what their AI systems cost to operate, according to KPMG’s US Q2 2026 AI Quarterly Pulse survey of 204 senior leaders. KPMG’s separate Global AI Pulse survey of 2,145 senior leaders across 20 countries and territories found that organizations with strong cost visibility were five times more likely to report established ROI than those without it (15% vs. 3%).
The Tokenomics Foundation’s 2026 State of Tokenomics survey, based on 472 organizations across 11 industries representing $4.6 trillion in combined revenue, makes the gap even starker: 96% use frontier model providers, yet roughly three in four cannot confidently connect AI spend to a measurable business outcome their CFO would accept.
If you cannot show your CFO where the money is going, you are not managing an AI strategy – you are managing an AI expense.
Rethinking the Metric: Cost Per Task, Not Cost Per Token
One of the most important shifts in AI financial thinking in 2026 is a change in the unit of measurement. Tracking cost per token is like measuring a factory’s efficiency by counting how many bolts it uses rather than how many finished products it ships.
The more meaningful metric is cost per completed task – the total attributable cost of useful work that meets defined acceptance criteria, including model usage, retries, failed attempts, tool calls and human intervention. The FinOps Foundation’s unit-economics framework supports connecting technology cost to business units such as a transaction, case resolved or workload, rather than treating a raw consumption metric as the outcome.
As a September 2026 AIwire analysis argued, optimising cost per token alone can encourage organizations simply to use less AI, rather than use it more effectively. More durable measures include cost per completed task, workflow or resolved unit of work, assessed alongside quality and compared with the equivalent human or legacy-software cost.
What Organizations Getting ROI Are Doing Differently
Research from the Tokenomics Foundation and KPMG points to several operational disciplines associated with stronger AI value and ROI:
1. They Have Defined Ownership
The State of Tokenomics survey found that companies with a clearly defined owner for AI economics were 3.7 times more likely to show their CFO a measurable outcome from AI spend. Accountability is a structural advantage.
2. They Are Routing Intelligently
Not every task needs a frontier model. Routine, high-volume work such as classification, extraction and summarisation may be suitable for smaller or open-weight models, while complex reasoning and high-stakes work may justify frontier models. The Tokenomics Foundation found that 86% of respondents were evaluating or using model routing, and router users were four times more likely to be able to show CFO value. Routing is a major cost lever, but its benefits depend on workload quality, risk and performance requirements.
3. They Measure Outcomes, Not Activity
The organizations seeing returns have moved from tracking “AI usage” to tracking business outcomes: tickets resolved, documents processed, hours saved, revenue influenced. The metric that matters is not how many tokens were consumed – it is what changed in the business as a result.
4. They Have Embedded Cost Reviews Into Governance
KPMG’s Q2 2026 US survey found that 61% of organizations had approval processes in place, but only 36% had direct token or usage controls. The practical lesson is to attach a cost model, an outcome target and a review cadence before an AI workload moves into production.
The Bottom Line
AI investment is not slowing down. The question is whether your organization is positioned to turn that spend into demonstrable value – or whether you will find yourself explaining to the board why the AI budget tripled and the results are still intangible.
Token spending without governance is not a technology problem. It is a business risk. And the organizations that are pulling ahead are not necessarily spending more – they are spending more deliberately.
At Insentra, we work with organizations to move from AI experimentation to AI accountability. Meaning we help organizations build the visibility, governance frameworks, and outcome-measurement practices to turn AI spend into a line item you can defend, optimize and scale with confidence.
If your AI investment is not showing up in your outcomes, it is time to change the conversation.
Insentra’s AI Momentum programme trains and coaches your people to use AI where it delivers real value, with the governance to keep spend under control. Capability your team owns, and outcomes you can defend. Explore AI Momentum.
