Summary: TechOrange’s October 17, 2026 report on the State of AI Cost Governance 2026 study by Mavvrik and Benchmarkit lays out an uncomfortable reality for builders: nearly everyone tracks AI costs, almost nobody forecasts them well, and the biggest leaks — agentic workflows and AI coding tools — sit exactly where finance can’t see them. If you build agents or ship AI features, the data here explains why your budget misses and what to fix first.
The gap between tracking and controlling
The study, covering 396 enterprises, found that 98% of companies now track AI infrastructure costs and 95% have formal budgets. Yet only 11% keep forecast error within ±10%. That mismatch produces real damage: 62% of companies say unexpected AI costs materially changed business strategy, a quarter of projects were paused or cancelled over them, and 49% had to reprice customer contracts mid-term, according to TechOrange’s report.
Why agents blow up budgets
Agentic AI has non-deterministic costs. Unlike traditional software, an agent chooses its own execution path per task; a parsing failure or system error can trigger retry loops and heavy orchestration, so the same workflow can cost 2x or even 10x more on a second run. Adoption is near-universal — 98% of firms run agentic AI in production — but only 36% report those costs properly in financial statements.
The surprise charges don’t come mainly from model tokens. Data platform fees top the list at 47% (think agents hammering Snowflake or Databricks), ahead of LLM token costs at 43% and retry-orchestration costs at 29%.
The second leak: coding tools bought in fragments
98% of engineering teams use AI coding assistants like Claude Code, GitHub Copilot, and Amazon Q Developer, averaging 2.4 tools per company. Most are purchased as individual licenses or departmental expenses; only 42% land in a formal AI budget. 39% of companies admit actual spend far exceeded procurement expectations.
This resonates with a pattern we covered when Uber capped employee AI tool spend at $1,500 a month — distributed tool purchases are hard to see until someone finally totals them up.
Misattribution eats margin
The downstream damage is concrete. 81% of companies report moderate-to-severe gross margin impact from AI costs; 63% saw margins fall 6–15 percentage points, and 18% fell more than 16 points. A root cause: 70% never put product-facing AI compute and API costs into cost of goods sold. When those costs hide in R&D or ops overhead, leadership can’t compute true per-customer cost — so pricing underestimates, and the more the product sells, the more it loses.
Timing is the other failure. 62% of companies apply no financial controls at the design stage: 15% do no cost review before invoices arrive, and 47% only start tracking after deployment. 46% lack real-time overspend alerts, and 63% still rely on monthly manual reviews.
Hybrid infrastructure widens the hole. 68% of enterprises run hybrid cloud across AWS (73%), Azure (65%), and GCP (58%). Traditional FinOps tools are built on single-provider billing APIs, which private GPU clusters, self-hosted models, and agent workflows don’t expose. Visibility drops from 67% on public cloud to 36–44% on private infrastructure.
What to actually do
The report’s playbook for a “proactive financial line of defense” breaks down by role:
- Finance: require pre-deployment cost reviews at architecture design stage; shift from fixed-cost extrapolation to driver-based forecasting (cost per workflow run or interaction); push product AI spend into COGS.
- CIO/CTO: build unified cost data collection across hybrid and private GPU environments, and bake agent- and model-level cost tracking into architecture now, not post-launch.
- FinOps: extend monitoring beyond cloud billing APIs to agents and third-party hosted models; treat overspend detection latency as the top metric; automate real-time alerts.
- Product owners: separate Product AI (margin and customer experience) from Workflow AI (employee efficiency) as distinct financial problems; price on measured per-interaction cost before committing.
- Engineering leads: pull the average 2.4 coding tools into one formal budget, and integrate tool billing and contract quota monitoring before agentic SDLC scales.
Visibility before ROI
The report’s closing message is blunt: “Visibility first. Everything follows.” Total cost is the denominator of any ROI calculation — an incomplete denominator makes every ROI number meaningless. That echoes what we found in where enterprise AI ROI forecasts break down: the forecast fails long before the value question gets answered.
For builders, the actionable takeaway is narrow and doable: instrument per-workflow cost at design time, split product-facing costs into COGS, and get your coding tools onto one budget line. None of that requires new slogans — just earlier, better-placed measurement.