67% of enterprises overran their AI agent budget this year

IDC's Future Enterprise Resiliency and Spending Survey, Wave 4, fielded in July 2026, found 67% of enterprises overran their AI agent spend budget by more than 10% over the past 12 months. The average enterprise with visibility into its own agent costs now reports monthly agent spend of $117,558 — over $1.4 million a year, annualized.

The overrun isn't a rounding error

IDC's breakdown shows the overrun isn't clustered at the edges: 43.1% of enterprises ran moderately over budget, by 10% to 25%; 18.5% ran significantly over, by 26% to 50%; and 5.4% blew past forecast by more than 50%. Put together, two-thirds of enterprises running agents in production missed their own budget by a double-digit margin — and for nearly a quarter of them, that margin was severe enough to call the original forecast meaningless.

Why agent spend forecasts keep missing

Agentic workloads don't scale the way the budgeting process assumes they will. A single agent run can spawn sub-tasks, retry failed steps, call external tools, and re-inject context on every loop — each of those is a separate, metered cost event that doesn't show up until the bill does. Teams that shipped multi-step agents into production have described burning through an entire year's AI budget in weeks, not the quarter it was scoped for, because the forecasting model was built around a per-request cost assumption that agentic loops don't respect.

The average hides a wide spread

$117,558 a month is the average among enterprises that can actually measure their own agent spend — which is itself a filtered group. Organizations without that visibility aren't in this average at all; they're the ones most likely to discover their real number on an invoice, after the fact. The gap between "average reported spend" and "actual total spend across the portfolio" is exactly where budget overruns hide until someone reconciles the books.

What separates the 33% that stayed on budget

The enterprises that didn't overrun weren't running smaller agent programs — they were the ones treating agent cost as a metered, per-task number to track continuously, not a fixed line to check quarterly. That distinction matters because agentic cost variance compounds fast: a workflow that costs 20% more per run than modeled doesn't show up as a 20% miss at year-end, it shows up as a 20% miss multiplied by every run since the model was set, which is how a "moderate" 10-25% overrun becomes a "severe" 50%+ one within a single fiscal year.

What to check before your next agent rollout

• Whether agent cost is tracked per completed task, including retries and sub-agent calls, or only as a monthly total against the provider invoice.
• How far into the fiscal year your current forecast is already off — a small early miss compounds faster in agentic workloads than in flat-rate usage.
• Whether you'd know about a 25%+ overrun this month, or only when the quarterly budget review catches it.

The bottom line

A budget overrun rate of 67% isn't a story about enterprises underestimating AI — it's a story about a forecasting method built for stable, per-seat costs being applied to a spend pattern that compounds per task, per retry, per loop. The 33% that stayed on budget didn't get lucky; they built the tracking that catches the drift before it becomes a severe overrun instead of a moderate one.