98% Track AI Costs. 11% Can Forecast Them

By Amanda Wagner | Trace3 Senior Cloud FinOps Consultant

An AI Cost Governance Use Case · Partner Deep Dive: Mavvrik

In our first article, we argued tracking AI spend is no longer the hard part and governance is where organizations are falling behind. Mavvrik's 2026 State of AI Cost Governance research, conducted with Benchmarkit across 396 enterprises, puts numbers on that claim. We read it as practitioners; here is where it matches the environments we work in, where we would add nuance, and what to do about it.

The Gap

Nearly every organization surveyed tracks AI infrastructure cost and assigns a formal AI budget. Almost none can forecast what that budget will consume. Tracking became universal; the discipline it exists to support did not follow.

Read together, those numbers describe an organization that can produce a report but cannot make a decision from it. That is the governance gap, stated in data. 

What Matches the Field

The finding closest to our client experience is the coverage cliff. Reporting is strong where an environment produces structured billing data and falls off in proportion to the distance from a billing API: public cloud sits at 67% coverage, while agentic workflows and GPU infrastructure sit at 36%. That is an inheritance problem more than a tooling failure. The frameworks doing the governing were built to reconcile cloud invoices, and AI runs in places that never issue one.

The second is the attribution gap. Team and department attribution lead the survey, and both answer the same question: who spent the money. Attribution by agent, model, customer, or user, the kind that explains why it cost what it did, shows up far less. We see this constantly. A client can name the cost center that owns a line item and has no idea which model choice or retry loop produced it.

The third should give every FinOps practice pause. Forty-four percent of organizations rate themselves Advanced, drawn from the same population where 89% cannot forecast within ±10%. Self-assessment measures inputs: budgets assigned, tools installed, teams named. The blind spots that should lower the score are the same ones hiding from the scorer.

Where We Would Add Nuance

A ±10% forecast is a demanding bar for consumption spend that rides on runtime behavior, and we would not treat a miss as proof of a broken practice. What matters more is whether the variance is explainable. A team that misses by 20% and can decompose it into model mix, retry volume, and adoption growth is in better shape than one that lands inside 10% by accident.

Worth stating plainly for anyone building an AI cost program: the research finds data-platform overages, not tokens, are the top source of unexpected AI cost. A framework that instruments tokens beautifully and ignores Snowflake, Databricks, or BigQuery query volume will miss the largest surprise category in the survey.

The Distinction Most Teams Miss

The most useful idea in the report is not a statistic. What drives AI cost depends on how AI is deployed, and each pattern breaks differently and needs a different lever. Product AI and workflow AI look nearly identical on an invoice while representing entirely different problems: one scales with customers and hits gross margin, the other scales with employees and accumulates as unvalidated operating expense. Reported together, they produce a number that is accurate in aggregate and risky to act on.

This is the part we would hand a client first because it reframes one AI cost problem as three, with three owners.

Governing the Spend the Billing API Never Sees

The pattern in the field looks like this. An organization has a competent cloud FinOps practice and a reporting cadence finance trusts. Then AI arrives, and the same practice starts producing numbers nobody can act on, because the reliability of the cost data is eroding. Agents run everywhere and report almost nowhere. Hybrid and on-premises AI issues no invoice to reconcile. Developer tooling routes around the AI budget entirely, at 98% adoption and 42% reporting. And 35% find an overrun only when the invoice lands.

That is not a reporting inconvenience. Sixty-two percent say an unexpected AI cost materially changed a business decision last year. Forty-nine percent were forced to reprice an AI product, usually mid-cycle in front of customers, 40% escalated to the board, and one in four delayed or cancelled an initiative.

Mavvrik's argument is AI cost visibility must be built rather than inherited. In the cloud era, visibility was a property of the platform. In the AI era, the signals sit in environments never designed to report cost in normalized format, and collecting them is an engineering commitment. In practice, that means extending reporting past cloud billing into hosted models, data platforms, developer tooling, agent runtimes, and private infrastructure, then attributing spend along the dimensions that explain it rather than the ones that happen to exist in the org chart.

The honest framing for anyone evaluating this category is no single vendor covers the full AI cloud spend today. The pattern we see working is layered: an aggregation platform as the system of record, paired with in-path enforcement that constrains spend before it is incurred. Judge Mavvrik, or any alternative, against the dimensions your own allocation model needs to resolve.

Where to Start

Four interventions we would prioritize. Finance should make pre-deployment cost review a standard control and measure AI against COGS rather than revenue, since 81% report margin impact but 70% have never required COGS-based tracking. Infrastructure leaders should treat unified signal collection as an architecture requirement and build agent-level tracking in before scale. FinOps should extend the reporting perimeter past cloud billing and cut detection lag toward real-time. Product should price the feature rather than the platform, separating product AI from workflow AI in the unit economics.

One sequencing note, because it is the most common mistake: these are dependent, not parallel. Forecasting, attribution, and ROI measurement all rest on complete cost signal coverage. Coverage comes first.

Start With Coverage

The organizations that forecast accurately are not smarter or better funded. They made one decision differently: they pushed coverage past cloud billing, into agents, GPUs, developer tools, and private infrastructure, before spend in those categories became material. That choice is far cheaper now than after a repricing event.

If your AI cost reporting stops at the cloud bill, you already know where the next surprise is coming from. Reach out to FinOpsTeam@Trace3.com to schedule a 30-minute AI spend review.

Go Deeper

If this is a live problem in your environment, three reads worth your time:

• Tooling up for the Tokenomics Era (Trace3 Innovation) — how enterprises got here: adopt first, ask questions later, and a bill nobody forecast.

• 2026 State of AI Cost Governance (Mavvrik and Benchmarkit) — the full dataset behind the numbers above, with coverage, attribution, and forecasting benchmarked across 396 enterprises.

• AI FinOps: The Invisible Budget Line (Trace3) — our practitioner take on building coverage before spend becomes material.
  
  

References

• Mavvrik and Benchmarkit. "2026 State of AI Cost Governance." Survey of 396 organizations, April to May 2026. mavvrik.ai

• FinOps Foundation. "State of FinOps 2026." finops.org

• Trace3 Cloud FinOps. "Your CFO Started Asking About Tokens: How FinOps Brings Order to AI Spend." 2026.

If you’re interested in learning more, please don’t hesitate to contact us at finopsteam@trace3.com, or visit our website.