Summary
The price of a given level of AI capability has fallen steeply since 2021. Spending has not. Usage has grown faster than prices have fallen, and much of that usage now runs through coding agents, personal accounts and API keys on individual laptops, outside the systems finance uses to allocate cost.
Independent and vendor surveys from 2026 describe the same position: surprise bills, no clear owner, few people who can trace a spike. The paper sets out what the evidence shows, why the fix starts with attribution by device and team, and the unit economics finance can use.
The numbers
- 72%had an unexpected AI cost spike or bill in the past yearHarness · 2026 · N=700
- 79%of finance leaders saw AI-related cost overruns in 12 monthsDoiT · 2026 · N=500
- 52%say there is no clear owner for AI costHarness · 2026 · N=700
- 26%report full, real-time visibility into what AI systems cost to runKPMG · Q2 2026 · N=204
Key findings
- Price and spend move in opposite directions. a16z measured a 10x annual fall in the cost of equivalent LLM performance, yet Gartner expects overall inference costs to rise because token use grows faster.
- AI cost is now standard FinOps work. 98% of State of FinOps 2026 respondents manage AI spend, up from 31% two years earlier.
- Personal accounts are falling, not gone. 47% of genAI users still used personal apps in Netskope's 2026 data, down from 78%.
- Policies exist; visibility does not. 73% of Harness respondents have AI cost policies, but only 13% have basic visibility into AI spend.
- Few can explain a spike. If AI spend doubled, 20% could find the cause within hours.
What is inside
- 01Cheaper tokens, bigger bills
- 02Where AI spend hides
- 03Personal keys, company work
- 04The survey evidence
- 05When the coding budget ran out
- 06Who owns the bill
- 07Unit economics finance can use
- 08What your own data should show
- 09From crawl to run