The FinOps maturity paradox: why your most disciplined teams post the biggest AI overruns

If cloud FinOps discipline transferred cleanly to AI spend, the most mature organizations should have the AI budgets furthest under control. New 2026 survey data says the opposite: the more mature a company's FinOps practice, the more likely — and the more severely — it overran its AI budget.

The number that shouldn't be possible

A February 2026 survey of 500 finance leaders across the US and UK — companies with 1,000+ employees, fielded by Sapio Research on behalf of DoiT — split respondents by self-assessed FinOps maturity and asked a simple question: did your AI initiatives run over budget in the past 12 months?

Among organizations that rated themselves very mature or leading-edge on FinOps, 89% had experienced AI cost overruns, with a mean overspend of 30.9%. Among early-stage FinOps organizations, only 69% overran, with a mean overspend of 16.1%. The teams with the most cost discipline on paper posted overruns 20 points more often and nearly twice as large.

That is not a rounding error. It is a direct contradiction of what every cloud FinOps program has trained finance and engineering leaders to expect: that maturity buys control. For AI spend specifically, in 2026, it does not — at least not on its own.

What FinOps maturity was built to solve — and what it wasn't

Cloud FinOps matured around a specific cost shape: relatively stable unit prices, resource-level granularity (an instance, a bucket, a function), and spend that scales roughly with provisioned capacity. Tagging, rightsizing, reserved commitments, and showback dashboards were built for that shape, and they work well against it.

AI spend has a fundamentally different shape. Cost varies 100x or more between models on the same task. A single feature's cost profile can shift overnight when someone edits a system prompt, swaps a model version, or ships an agent that loops five extra times before it succeeds. The unit of spend is a request, not a resource — and the price of that request depends on a chain of runtime decisions no capacity plan captures.

The DoiT survey data points at exactly this gap. Mature organizations aren't failing because they're careless; the report's own framing is that "mature organizations run larger, more complex AI programs" — they have more surface area, more models in production, more teams building agentic features, and better instrumentation to actually catch the overrun when it happens. They see the problem clearly. Seeing it is not the same as having a lever built to stop it.

Where the overruns are actually coming from

A separate July 2026 report from WitnessAI, surveying 300 business executives, found 68% of US companies experienced AI budget overruns in the past year, with 33% saying it happened "mostly or always." Asked why, 30% pointed directly to unmanaged or poorly governed AI usage, and 27% reported delayed or canceled initiatives as a direct consequence of governance gaps — not the other way around. Governance failure isn't a side effect of AI cost overruns in this data; for nearly a third of respondents, it's the cause.

The same report breaks down where ungoverned usage concentrates by department: IT and infrastructure teams lead at 47%, followed by sales and business development at 34% and marketing at 33%. These are exactly the teams closest to shipping AI features fastest — and furthest from the finance function that owns the budget line.

Compounding this, a Q2 2026 KPMG report found the share of organizations running multiple concurrent AI agents doubled from 9% to 18% in a single quarter. Multi-agent architectures compound cost in ways single-model chat features never did — each additional agent in a pipeline resends context, retries independently, and can trigger downstream agents of its own. Scaling agent count is scaling a cost multiplier, not just a feature count, and most budget models built for the chat era don't have a term for it.

The governance line item nobody budgeted for

There's a second, quieter number worth sitting with: according to Gartner, AI governance now claims 8-12% of the average enterprise AI budget in 2026, up from just 3-5% in 2024 — making it the fastest-growing line in the AI budget, ahead of model spend growth itself. Enterprise AI spending overall reached an estimated $407 billion in 2026, up 34.8% from $302 billion in 2025, and governance is eating a larger share of that growth every quarter.

This is the tell. Mature organizations aren't overspending because they ignore governance — they're overspending in part because they're now paying for governance as a distinct, growing cost category that a cloud-era FinOps model never had a bucket for. Treating AI governance as a compliance afterthought rather than a budgeted, forecasted line item is itself a source of the overrun, not just a response to it.

Why detection isn't the same as control

There's real pressure behind all of this. A CloudZero survey found 87% of finance leaders feel pressure to demonstrate AI ROI within one year of investment — but only 22% have actually achieved that. And per the DoiT data, even among organizations investing seriously in cost visibility, only 15% can calculate AI ROI without significant bottlenecks. The top three barriers cited: the pace of technological change (40%), finance and engineering defining "success" differently (37%), and a plain lack of clear financial attribution (36%).

Put together, this describes organizations that can increasingly see their AI spend — dashboards, monthly reports, model-level breakdowns — but still can't act on it fast enough to stay inside budget. Visibility without a control loop that operates at the speed AI cost actually moves is exactly what produces a 30.9% average overspend in your most instrumented teams.

What AI-native cost control actually requires

The organizations closing this gap share a few traits that don't come from the cloud FinOps playbook:

• Attribution at the request and agent-run level, not the monthly invoice level. A dashboard that updates after the billing cycle closes is a post-mortem, not a control.
• A shared cost definition between finance and engineering, set before the quarter starts. The 37% gap on what "success" means is a governance failure that shows up as a budget failure three months later.
• Governance budgeted as its own forecasted line, not folded into "AI spend" as a single number. If it's 8-12% of your AI budget and growing faster than the rest, it needs its own forecast and its own owner.
• Hard per-agent and per-feature ceilings enforced in real time, not just alerts reviewed the next morning — the gap between "multi-agent orgs doubled" and "cost stayed under control" is entirely a function of whether a ceiling exists before the agent runs, not after.
• Ownership assigned at the team that ships the feature, matching the WitnessAI finding that ungoverned usage concentrates exactly where features ship fastest: IT, sales, and marketing.

The bottom line

FinOps maturity was never the wrong instinct — it's just calibrated for a different cost shape than the one AI spend actually has. The 2026 data is blunt about the consequence: without AI-native attribution and real-time control, being more disciplined about cloud costs correlates with running bigger AI overruns, not smaller ones, because maturity buys you bigger programs and better detection before it buys you control. The fix isn't more dashboards. It's closing the loop between seeing the spend and stopping it, at the speed AI spend actually moves.