Ninety-two percent of CFOs and top finance executives feel pressure to prove that their AI investments are delivering a return. Half of those surveyed say AI agents have yielded only limited measurable ROI. And only 7% say their organization prioritizes AI governance over speed of adoption. Those three numbers, from Avalara's July 21 survey of 1,505 finance leaders, describe the exact position most enterprise AI programs are in right now: spending at scale, governing at minimum, proving almost nothing.
McKinsey's response to this moment is a framework called AI FinOps — applying the financial operations discipline that tamed cloud spending to a new class of unpredictable, consumption-driven AI costs. For CFOs, CIOs, and anyone accountable for an AI budget that's growing faster than its verified business impact, this is the operating model that changes how the math works.
The Governance Gap Is Getting Expensive
The Avalara data, published July 21, 2026, covers 1,505 CFOs and senior finance executives. The findings are a systematic description of an enterprise AI program that has outrun its oversight infrastructure.
44% of respondents are only "somewhat confident" they could explain the actions of an AI agent to an auditor or regulator. In regulated industries — financial services, healthcare, insurance — that confidence level is a compliance liability, not a management gap.
76% say their organizations lack the in-house expertise to understand how their AI operates. This is the accountability void McKinsey's AI FinOps framework is designed to close: when no one in the organization fully understands the system generating costs and decisions, financial controls become impossible to enforce.
30% haven't updated their internal controls within the last year — a period that saw the most significant shift in enterprise AI deployment in history, including the transition from single-model deployments to multi-agent systems with compounding costs.
23% say responsibility for an AI error would fall to no one or would be unclear. For any CFO signing off on an AI budget, that's not a technical detail — it's a fiduciary issue.
46% have incident response plans that are untested or still in development. Only 28% require documented audit logs showing how their AI agent reaches decisions.
"Nearly every leader we surveyed feels personal pressure to prove these investments are delivering," said Jayme Fishman, EVP at Avalara. "Meanwhile, the controls, expertise and accountability meant to sit underneath them are still being built."
That sentence describes the gap AI FinOps is designed to close.
The Scale of What's at Stake
Gartner forecasts worldwide AI spending will reach $2.6 trillion in 2026 — a 47% increase from $1.76 trillion in 2025. Enterprise AI spend specifically hit $407 billion in 2026, up 34.8% year-over-year. By 2030, total AI investment is projected at $5.62 trillion.
Against that backdrop, the Avalara finding that 50% of AI agents have yielded only "limited measurable ROI" represents an enormous capital efficiency problem. At $407 billion in enterprise AI spend, even a 50% limited-ROI rate implies hundreds of billions of dollars deployed without verifiable business return.
AI governance is already emerging as the fastest-growing budget line in enterprise AI: it now claims 8–12% of total AI spend according to enterprise spending analysis, up from negligible levels in 2024. The market is responding to the governance gap even as most organizations haven't implemented the frameworks to make governance meaningful.
This is exactly the dynamic McKinsey's AI FinOps framework is designed to address before spending scales further.
What Is AI FinOps?
FinOps — financial operations — emerged as a discipline in 2019 to manage the unpredictable, usage-based cost structure of cloud infrastructure. Before FinOps, cloud spending was opaque, attribution was inconsistent, and engineering teams had no real-time feedback on the cost implications of their decisions.
AI spending has the same structural problem, with additional complexity. Cloud costs are driven by compute and storage consumption. AI costs are driven by token consumption — and tokens per task vary by model choice, prompt design, context window usage, reasoning effort level, and output length. A team that switches from GPT-5.6 Luna to GPT-5.6 Sol without changing their prompts can see a 5x cost increase overnight. A poorly designed agentic loop can generate 10x more output tokens than a well-designed one on the same task.
McKinsey's AI FinOps framework applies four disciplines from cloud FinOps to this environment:
1. Cost Visibility
The first discipline is tagging and attribution — understanding which teams, products, and workflows are generating which AI costs. Most enterprise AI deployments today route through a central API key, which makes cost attribution nearly impossible. AI FinOps requires per-team, per-product, per-workflow cost tracking at the token level.
This isn't a finance problem. It's an infrastructure problem: the tooling to capture this data at the model API level needs to be in place before finance can do anything with it.
2. Demand Governance
Demand governance means establishing who can use which models for which workloads — and requiring business justification for premium model usage. In practice, this looks like a model tiering policy: default to mid-tier models (GPT-5.6 Luna, Gemini 3.6 Flash) for standard workflows, require approval for flagship model usage (GPT-5.6 Sol, Claude Fable 5), and reserve the most expensive reasoning tiers for tasks where the quality differential has been demonstrated.
Most organizations don't have this policy. Engineers make model selection decisions based on ease of access and perceived capability, without cost accountability. AI FinOps makes model selection a governed decision with documented business rationale.
3. Usage Accountability
Usage accountability means holding teams accountable for AI spend the same way they're accountable for headcount and software licenses. This requires budget allocation — AI token costs as a line item in team budgets — and regular review of spend-per-output-unit rather than just spend-per-period.
The unit economics question is: what does it cost to complete one task of this type? For a customer service agent: cost per resolved ticket. For a code review agent: cost per reviewed PR. For a document extraction workflow: cost per processed document. When teams see cost-per-unit rather than aggregate spend, optimization behavior follows naturally.
4. Unit Economics Measurement
The final discipline is the one most directly connected to CFO pressure: establishing the ROI numerator. AI spend is the denominator. The ROI numerator — what business value did this spend generate — requires measurement infrastructure that most organizations haven't built.
This means instrumenting AI workflows to capture outcome data: did the agent-completed task produce a usable output? How did it compare to human-completed work on quality, time, and cost? What was the downstream business impact?
The Avalara finding that only 28% of organizations require documented audit logs for AI agent decisions is a direct indicator of how few organizations have this instrumentation in place. Without it, ROI claims are qualitative at best.
Why This Is a CFO Problem, Not Just a CTO Problem
The temptation is to treat AI FinOps as a technical discipline — something for the engineering team to implement. That framing is wrong, and it's the reason most AI governance initiatives fail to change spending behavior.
AI FinOps works when it's a joint CFO-CTO function. The CFO provides the financial governance structure: budget allocation, cost attribution requirements, ROI measurement standards. The CTO provides the technical infrastructure: model tiering policies, API cost tracking, workflow instrumentation.
Without CFO ownership, AI FinOps becomes an engineering cost-optimization exercise that doesn't change the business case for AI investment. Without CTO ownership, it becomes a finance reporting exercise that doesn't change how engineers make model selection decisions.
The 92% CFO pressure figure from Avalara reflects what happens when this joint ownership doesn't exist: finance leaders feel the pressure to justify AI spend but don't have the instrumentation or governance infrastructure to actually measure it.
A Practical Starting Framework
For enterprise teams at the beginning of an AI FinOps implementation:
Week 1–2: Cost visibility baseline
- Implement per-team API key tagging or cost center routing at the model API level
- Run a 30-day spend audit: which teams, which models, which workflows
- Identify your top 5 cost centers by token spend
Week 3–4: Model tiering policy
- Define three tiers: standard (mid-tier models), enhanced (mid-flagship), premium (flagship)
- Set default tier as standard for new workflows
- Require documented business justification for enhanced or premium usage
- Review existing workflows — most can likely drop to a lower tier without quality impact
Month 2: Unit economics baseline
- For your top 3 AI workflows by cost, define the unit of output (ticket, document, PR, report)
- Instrument each workflow to capture output count
- Calculate current cost-per-unit
- Set a 90-day improvement target (typically 20–40% reduction is achievable through model tiering and prompt optimization alone)
Month 3: ROI instrumentation
- For each instrumented workflow, capture the business outcome: time saved, error rate, output quality score, downstream conversion
- Build the ROI numerator
- Connect cost-per-unit to value-per-unit
This 90-day implementation path doesn't require new tooling in most cases. It requires governance decisions — who approves premium model usage, how are costs attributed, what counts as a successful output — that are leadership decisions, not engineering ones.
The Compliance Dimension
For regulated industries, AI FinOps has a parallel compliance benefit that goes beyond ROI measurement.
The Avalara finding that 44% of CFOs couldn't confidently explain AI agent actions to an auditor describes a regulatory exposure that's becoming more acute. The EU AI Act's high-risk AI system requirements, which took effect in August 2026, require documented decision audit trails for AI systems in financial services, healthcare, and certain HR applications.
Organizations with mature AI FinOps implementations — particularly those with documented audit logs (which only 28% currently have) — will find regulatory compliance significantly easier. The instrumentation required for cost accountability and ROI measurement is largely the same instrumentation required for regulatory audit trails.
Building AI FinOps now serves two purposes: it answers the CFO's ROI question and it prepares the organization for regulatory requirements that are already in force in the EU and approaching in the U.S.
The Bottom Line
The Avalara data from July 21 is not a warning about AI's future. It's a description of an organizational maturity gap in the present. Spending is scaling. Governance is not. ROI measurement is largely absent. And 92% of the people responsible for the budgets feel the pressure of that gap personally.
McKinsey's AI FinOps framework is the operating model that closes it. Not by slowing AI adoption — by making the cost of AI visible, attributable, and accountable in the same way that cloud costs became visible, attributable, and accountable a decade ago.
The organizations that implement AI FinOps in the next 90 days will have a clear answer to the board's ROI question. The ones that don't will still be at the 92% pressure figure when the next survey runs.
Sources: Avalara "Agents of Change" survey (1,505 CFOs and senior finance executives, July 21, 2026), CFO Dive reporting, McKinsey AI FinOps framework (July 2026), Gartner AI spending forecast (May 2026), enterprise AI spending analysis via AboutChromebooks/ValueAddVC (July 2026).
