FinOps for AI: Five Reasons AI Spend Spirals [Part II of IV]

FinOps for AI: The new billing model every technology leader needs to understand

This is the second in a four-part series on FinOps for AI, written for organizations actively scaling AI workloads. 

This article names five major structural pain points in AI spend. The first article defines what FinOps for AI is and why it differs from Cloud FinOps. The third lays out a four-phase playbook, Understand, Attribute, Optimize, Operate, for acting on those pain points. The fourth shows how WWT ran that playbook on itself, and it closes with recommendations for where to start.

Executive summary

AI spend spirals for five primary reasons, and enterprises facing unexpected AI costs are usually dealing with some combination of them. The five issues compound one another, so a fix aimed at just one leaves an unrealized built-in gap, such as a dashboard nobody acts on or a policy nobody can enforce.

Token-maxxing, as defined in the first article in this series, is the symptom that brings executives to the table. The causes underneath it are structural. No unified visibility means AI spend is scattered across cloud providers, SaaS vendors and internal API calls, with no complete picture available anywhere.

No cost attribution means a shared endpoint's cost lands in a central budget that no business unit believes it owns, leading to significant experimentation with inefficient choices of tools, models and data sources. No agentic cost control means workflows that loop and re-prompt on their own. No ROI framework means every AI investment runs as a faith-based budget line. And a lack of AI-specific governance leaves model choice, integration approval and shutdown authority unguided, which is where shadow AI takes root.

Each of these structural issues feeds into the next, snowballing into a budget crisis. This article gives each of the five causes a diagnostic for an executive team to apply in one sitting.

Where the problem starts

The spend goes live before anyone owns it

Take the moment a new AI feature ships. The product team added a new feature to the support workflow mid-sprint, and the demo left everyone impressed. No contract crosses the procurement's desk, because the feature was added by an account that already existed, and the spend grows like small charges on a credit card that no one notices until the statement arrives.

AI spend starts in most enterprises as a small, reasonable decision made at the team level, invisible to every function that would normally govern it. High token consumption without a defined business outcome behind it generates AI spend rather than AI adoption, and by the time the number is large enough to notice, usage inefficiencies have compounded.

- Chuck Balog, WWT Cloud Strategy & FinOps Practice Director

Why the five issues compound

These five issues do not operate independently. Missing visibility makes attribution impossible, since you can't assign a cost to what you can't see. Missing attribution dissolves accountability, and unaccountable spend lets agentic workflows run unchecked. Unchecked workflows make ROI impossible to measure, and without ROI measurement, governance has no foundation to enforce.

The compounding is why partial fixes are not enough. An organization that buys a visibility tool while leaving attribution and governance untouched gets a better view of the soaring spend, but still cannot govern it. 

Here we are naming the five common AI spend pain points we see enterprises facing today, along with a simple self-diagnostic to see where you are.

The five structural pain points of AI spend

Pain point 1: AI spend scatters across sources no one team can see end-to-end

A finance analyst gets asked for the company's total AI spend for a technology finance review. What comes back is a spreadsheet of information fragments: AI services pulled from the cloud bill, subscriptions found in expense reports and an API invoice paid on a corporate card, with no way to know what the list is missing. In some periods, the licensing structures have changed in ways that hinder month-on-month comparisons.

AI spend is distributed across cloud providers, SaaS vendors and internal API calls, and no single team has a complete picture because existing FinOps tools were never built to aggregate across these sources. Across organizations, we have seen Cloud FinOps frameworks, software lifecycle management and product portfolio management. Yet as is, none of these tools really captures AI well.

The gap is industry-wide: The FinOps Foundation's State of FinOps 2026 report notes strong demand for new FinOps tooling capabilities as organizations seek greater visibility into AI-specific cost drivers, such as tokens and inference costs. Finance sees line items it cannot decode, while engineering sees throughput metrics it cannot translate to dollars. We hit the same wall ourselves when our cloud billing dashboards could not break down AI spend by use case or team. We had to build a custom attribution layer for AI.

Key question: If leadership requested a complete picture of AI spend today, is there a single team that could provide it end-to-end?

Pain point 2: The unattributed bill lands centrally, and accountability collapses

When the assembled number finally reaches a leadership meeting, the head of each business unit scans the bill only to find nothing regarding their own share. Why? Because the bill was never broken down to show attribution. The line stays parked in a central budget, and most teams leave the room confident that the problem belongs to someone else.

Attribution fails for a mechanical reason. A single shared model endpoint can serve many teams at once, and the API call that drives it does not carry department labels. User IDs on the vendor side are not always synced with the org structure, or the data is hard to link back. Without attribution, no one takes ownership of the numbers or the initiative to solve overspend or token-maxxing issues. 

The FinOps Foundation recognizes that shared AI services often need supplemental tagging, governance and cost-tracking mechanisms to allocate spend accurately. Attribution is therefore an active design decision, not an out-of-the-box capability.

Key question: Can every AI-related expense from last month be traced to a team that reviews, understands and supports the associated costs?

Pain point 3: Without owners or limits, agentic workflows will compound spend

Workloads keep moving while the ownership question stalls. An engineer checks the overnight queue and finds an agent workflow still running. It was designed to complete its task in a few steps, but it has spent the night looping through retries and tool calls, and no circuit breaker stopped it.

Single-step AI tasks are predictable, but agentic workflows are a different economics. Every loop, tool call and re-prompt consumes more tokens, and the workflow decides on its own how many of those it needs. AI agent costs at scale are highly unpredictable, and budget overruns compound quickly in architectures where agents interact heavily with one another.

Key question: Within the next hour, could your organization name its three largest AI cost drivers by team, tool and use case?

Did you know: Gartner explains that "Quadratic complexity describes a situation where, as the number of AI agents or services in an AI system increases, the potential for cost overruns grows rapidly — specifically, the cost rises with the square of the number of agents." [1]

Pain point 4: Spend that can't prove value runs as a faith-based budget line

At the quarterly budget review, the AI program lead presents adoption charts that climb along even faster cost charts. When asked what the program returned, the lead offers usage numbers, and nobody in the room can map the cost to business returns. Anecdotes of good value abound. In early adopter organizations, so do anecdotes of highly inefficient and sometimes unnecessary use.

Without a cost-per-completed-task measure, every AI investment is a faith-based budget line. AI programs that cannot prove value can lose executive sponsorship before they can demonstrate results. The unit economics work, called for in our first article, converts that faith into a defensible number.

Key question: For your three most expensive AI use cases, what is the cost of one completed task, and can business value be calculated for it?

Pain point 5: Governance built for procurement cannot answer AI's questions, and shadow AI fills the gap

When the runaway workflow from the overnight queue finally surfaces in an escalation channel, the thread fills with people asking who has authority to shut it down. Security, the platform team and a business unit each carefully consider the risk of stopping work someone needs, while no one raises a hand. The workflow keeps running while they decide. Once shut down, the specific user learns to never run such workflows again. Everyone else with the same access? They never get the memo, and the gap persists.

Governance structures built for software procurement do not extend to AI consumption. Nobody regulates the new types of decisions AI raises: which model to use for a task, which new API integrations to approve, and when to shut down a runaway workflow. Gartner describes "shadow agents" as "AI agents created by developers or business units without governance oversight, which often go unnoticed until they produce cost leakage or compliance breaches." [2] The result is spend that compounds before anyone notices.

Key question: Who holds the standing authority to pause a runaway AI workflow without waiting for a meeting? If the honest answer is no one, this fifth gap is open, whatever the other four look like.

How the five compound into a budget crisis

Follow the issues to where they lead, and the budget crisis stops being mysterious. Organizations that cannot see spend cannot assign it. Nothing assigned means nothing is defended. Workflows run without limits, and value goes unmeasured. Governance arrives last and holds a policy it can't enforce. The crisis is the compounded interest on five separate debts.

Most technology leaders reading this will immediately recognize several of the five issues in their own organization. Importantly, this recognition is the first step the diagnostic is designed to produce. The honest answers are uncomfortable, but the discomfort is cheaper than the invoice that otherwise delivers the same news.

The next article in this series lays out a four-phase playbook that addresses the issues in order, but it only helps organizations that recognize their gaps.
 

Essential terms

• Agentic workflow: An AI workflow in which autonomous agents plan, loop, call tools and trigger further model calls on their own.
• Attribution convention: The agreed tagging and labeling scheme that maps every AI cost to a team, workload, and environment.
• Chargeback trigger: The point at which a team's attributed AI costs begin billing to its own budget rather than a central one.
• Circuit breaker: An automated control that halts an AI workload when its spend or behavior crosses a set limit.
• Shadow AI: AI tools or services adopted by teams outside any IT, security or procurement review.

Additional references

• [1] Gartner, "Don't Let AI Agents Burn Your Budget," Yogesh Bhatt, Deven Tasgaonkar, Ben Yan, Mike Fang, 1 March 2026. GARTNER is a trademark of Gartner, Inc. and/or its affiliates. 
• [2] Gartner, "FinOps Is Critical to Maximizing ROI of AI Agents," Deacon D.K Wan, Tigran Egiazarov, Aaron Harrison, Bill Blosen, 9 February 2026.