62% of companies had an AI cost surprise reach the board this year

A cost surprise used to mean an awkward finance meeting. In 2026, a new industry survey finds it's more likely to mean a board memo. 62% of organizations say an unexpected AI cost materially altered a business decision in the past year — and for 4 in 10 of them, that decision required board-level escalation.

From line item to board agenda

The Mavvrik/Benchmarkit 2026 State of AI Cost Governance Report, published August 11, 2026, found 62% of surveyed organizations experienced an unexpected AI cost that materially altered a business decision over the past year. Of those, 40% required board-level escalation, 33% triggered emergency spending freezes, and 25% delayed or canceled an AI initiative outright.

These aren't marginal numbers describing an edge case. They describe a majority outcome: a cost overrun in 2026 has real odds of becoming a strategic event that reaches the board, not a line item that gets quietly reconciled at month-end.

The forecast is getting worse, not better

The more striking number in the report isn't the overrun rate — it's the trend on forecasting. Only 11% of organizations can forecast AI expenditure within ±10% accuracy, down from 15% in 2025. Despite a full additional year of deployment experience, forecast accuracy declined.

The likely cause isn't a lack of effort. It's that the thing being forecast keeps getting more variable faster than forecasting practice can adapt: growing model diversity, reasoning and agentic workloads with highly non-linear token consumption, and multi-provider stacks all add variance to the spend curve every quarter. A forecasting method built for last year's simpler usage pattern doesn't automatically scale to this year's.

Where the surprise actually comes from

43% of organizations cite token costs specifically as the primary source of unexpected AI spending — not GPU or infrastructure procurement, not staffing, the per-request line itself. That tracks with what's driving usage in 2026: reasoning models bill invisible "thinking" tokens as output, and agentic workflows can burn 5-30x more tokens per completed task than a single chatbot turn. Both mechanics are inherently spiky and resist the kind of flat, per-seat or per-project budgeting assumption that still underlies most AI cost planning.

Why finance discipline alone doesn't fix this

An emergency spending freeze is a lagging control. By the time a project gets frozen or canceled after board escalation, it has usually already generated most of the cost that triggered the escalation in the first place — the intervention arrives too late to change the outcome, only to stop it from repeating. What actually prevents the escalation is catching the anomaly at the point of spend: a threshold that fires when a project's burn rate crosses a line, not a forecast that gets revisited once a quarter.

Three signals worth watching before they reach the board

• Forecast variance over time, not just the current quarter's forecast number — a widening gap between forecast and actual is the leading indicator, not the overrun itself.
• Spend concentration in agentic and reasoning workloads, since that's where the 43% of token-driven surprises cluster.
• Whether alerts fire before a threshold is breached, or only after the invoice reconciles — the difference between a control and a postmortem.

The bottom line

A board-level escalation is what happens when every earlier control has already failed silently. The 62% figure isn't a story about companies being careless with AI spend — it's a story about visibility arriving too late to matter, in an environment where the underlying cost is getting harder to forecast every year, not easier. Fixing the escalation rate starts well before the board agenda: at the request, the feature, the agent run where the spend actually happens.