The short answer: Cloud spend scales with infrastructure. SaaS spend is fixed by contract. AI spend is driven by content — what users type, how agents reason, how features are designed — and that makes it fundamentally unpredictable in a way that breaks standard governance frameworks. Most IT leaders reported unexpected AI-related charges in 2025, and the number is growing year over year.¹ The problem isn’t a lack of effort. It’s that the tools, processes, and mental models built for cloud and SaaS procurement simply don’t translate.
It Doesn’t Behave Like Any Cost You’ve Managed Before
Every governance framework your team has applies a version of the same logic: usage is somewhat predictable, contracts set ceilings, and monthly reconciliation catches drift. AI breaks all three assumptions simultaneously.
With cloud infrastructure, a spike in compute correlates with a spike in product traffic. The relationship is legible and forecastable. With AI, a single change to a system prompt, a new agent tool call, or a feature redesign that adds one reasoning step can multiply token consumption by 10x or 100x — with no corresponding change to traffic, infrastructure, or contract terms. Finance doesn’t see it coming because there’s no infrastructure signal to catch.
Cost is no longer a function of servers or seats — it’s a function of tokens, and tokens are driven by content. A poorly crafted prompt can cost ten times as much as an optimized one doing equivalent work, and neither generates a ticket or a purchase order.
Six Reasons AI Spend Resists Governance
The governance gap isn’t a single problem — it’s a compounding set of structural mismatches between how AI costs work and how finance and operations teams are set up to manage them.
- Consumption-based pricing with no natural ceiling. Traditional SaaS costs are fixed at contract. AI API costs are metered per token, per call, or per task. There’s no cap unless you explicitly set one — and most teams don’t until after the first surprise invoice.
- Cost is driven by content, not infrastructure. A change to a feature prompt or agent instruction set can produce a step-change in spend overnight without touching any server, storage, or network resource. Traditional FinOps tooling has no way to detect or forecast this.
- No procurement gate. Engineers add model API calls directly in code. Product managers enable AI features without a purchase order. The spend is live before finance knows it exists.
- Attribution is structurally broken. Most companies use a single shared API key for all model calls. A shared key collapses all inference spend into a single line — no customer attribution, no feature attribution, no team attribution. You get a total. Not a breakdown.
- Shadow AI is already inside the organization. More than half of employees are using unsanctioned AI tools. This spend doesn’t appear in any procurement record — it shows up as expensed subscriptions, corporate card charges, or not at all.
- Forecasting models don’t work. AI workloads are highly variable and can spike unpredictably — unlike infrastructure costs that scale linearly with load. Traditional forecasting assumes seasonal patterns and stable growth rates. AI spend has neither.²
The Shadow AI Problem
Shadow AI is the fastest-growing governance blind spot in software companies. Unlike shadow IT — which typically involved a team subscribing to an unauthorized SaaS tool — shadow AI is harder to detect because it often lives inside authorized tools. A sales rep using a CRM’s embedded AI summarization at scale. A developer querying a model API directly from a personal account. A customer success manager running weekly reports through a consumer AI tool with no data controls.
Gartner projects that a significant share of enterprise applications will feature embedded AI agents within the next few years — up sharply from where things stood in 2025.¹ Many of those embedded agents will generate inference spend that sits entirely outside any FinOps or procurement framework — because the application itself is approved even if the AI usage within it isn’t governed.
The spend consequences are real, but the data risk is arguably larger. Employees feeding sensitive customer data or internal financials into unsanctioned tools creates a liability that governance frameworks weren’t built to catch — because the tool is used inside an approved workflow.
The Agentic Wildcard
Agentic AI systems — those that can plan, use tools, take multi-step actions, and self-direct toward a goal — introduce a category of governance problem that has no precedent in cloud or SaaS management. AI agents burn 50x more tokens than standard chat interactions because they send the full conversation history with every tool call, accumulating context across multiple reasoning steps.
The cost runaway risk is concrete. One team’s pipeline ran for eleven days and generated a $47,000 bill before anyone intervened. The dashboards, alerts, and provider-level spending caps in place at the time were observability tools — not enforcement mechanisms. Observability tells you what happened. It doesn’t stop the spend.
Agents can enter recursive loops, over-query external systems, or expand tasks beyond their original scope with no human checkpoint. The spend compounds at machine speed, not human speed. What often looks like an engineering optimization problem surfaces later as a profitability problem — especially when unmanaged AI feature costs quietly compress customer margins long before finance notices.
What Governance Actually Requires for AI
Standard FinOps practice — tag resources, forecast from history, reconcile monthly — doesn’t translate. Governing AI spend requires a different operating model built around four capabilities:
- Real-time cost attribution. Every model call must be tagged with a customer ID, feature name, model version, and team at the application layer before it reaches the API. A shared key with no tagging makes governance structurally impossible.
- Enforcement, not just visibility. Dashboards and alerts are retrospective. Governance requires hard limits: per-run caps, per-feature budgets, per-customer cost ceilings, and automated throttling or termination when thresholds are breached.
- Prompt and model governance. Since a single prompt change can 10x spend without any infrastructure change, the prompt and model selection layer must be treated as a cost management surface — reviewed before deployment, tracked after.
- Sanction and visibility for AI tool adoption. A process for approving new AI tools and tracking usage — especially embedded AI within approved SaaS products — prevents shadow AI from accumulating into unmanaged spend.
The companies adapting fastest are moving toward AI-specific spend metrics that connect cost to customer, product, and business outcomes — rather than relying on infrastructure-era reporting that was never designed to capture this kind of variability.
The companies that govern AI spend well treat it as a product operations problem, not a finance reconciliation problem. The spend is generated in real time, at the feature and customer level, by decisions made in code. That’s where the governance needs to live.
The Bottom Line
AI spend is uniquely hard to govern because it violates every assumption that cloud and SaaS governance frameworks are built on. The cost isn’t fixed by contract, it doesn’t scale with infrastructure, it can’t be forecasted from historical patterns, and it’s generated at points — in code, in agent workflows, in unsanctioned tools — that sit entirely outside traditional procurement.
That doesn’t make it ungovernable. It means governance has to move closer to where the cost is generated — into the application layer, the model routing layer, and the agent runtime. Most companies have built the observability layer but not the enforcement layer. That gap is where AI spend gets away from you.
Brad Perry is the CEO of Cogs’z, a profitability management platform built for B2B SaaS companies. Brad co-founded DealerSocket, an end-to-end platform in the automotive industry, where he experienced firsthand the margin challenges that Cogs’z is designed to solve. Cogs’z automates customer-level cost allocation so finance, CS, and sales teams share a single, accurate, view of who’s profitable and why. Learn more or request a demo at cogsz.com.
References
- Gartner — AI Value Realization & Enterprise AI Adoption Research
- FinOps Foundation — State of FinOps: AI Workloads