Even an AI cost-management vendor can lose control of its agent spending
Unmonitored AI agents can destroy systems and rack up huge costs, as even a cost-management vendor learned.
The study, which analyzed 109,000+ incidents, found that in at least nine documented cases over the past year, AI agents used valid credentials to wipe data and delete databases on their own. Standard monitoring did not detect the activity until damage appeared. It also found that median resolution times for failures have been roughly flat since 2023, and the single most common fix is waiting for another company's engineers.
Revenium's engineers described how the four-day session began on May 13, when a developer opened an AI coding session on his laptop. The session stayed open for four days, ran 4,819 calls and cost $3,762, with no budget allocated and no alert fired, they said. They added that there was nothing unusual about the session, and that the budgeting model for AI sessions, which treats them like SaaS subscriptions with flat per-seat or per-token costs, breaks down when looking at actual usage.
In another case cited by Revenium, two AI agents entered an infinite conversation loop that ran undetected for 11 days and consumed $47,000. A mid-sized e-commerce customer saw AI infrastructure costs jump from $5,000 per month during prototyping to $50,000 per month in staging, a tenfold increase driven by unoptimized RAG queries and recursive agent loops during high-volume periods.
The company's internal audit of its own agentic AI practices found that averages hide the true cost. Over 90 days, the median cost for agent-based work was $2.24 across 557 code-implementation tasks, while the most expensive task reached $300.97. If you manage an AI bill against an average, you have no visibility into what could happen tomorrow morning, the engineers cautioned.
The cost distribution is highly skewed. Of 14,680 AI runs tracked over 90 days, the top 1% of runs represented 46% of total spend, the top 5% accounted for 77%, and the bottom 90% accounted for only 12%. The engineers stated that AI spending lives in the tail of the distribution and that SaaS cost controls aim at the wrong part of the curve.
Interactive use, not automated pipelines, drives most of the expense. Among 10,005 interactive agentic sessions studied, the bill came to $109,118, while 4,171 automated software development lifecycle tasks cost only $6,723. The automated pipeline that implements and reviews pull requests costs under 6% of the bill; the remainder is engineers using AI through the day.
Pull-request costs also vary widely. Over 30 days, 12 engineers each merged at least 10 pull requests, for a total of 1,721. Cost per merged pull request ranged from $4.05 to $103.66, with a median of $16.59. The spread reflects different work and patterns of AI use, the team noted, and this variance sits in a budget dimension almost nobody is measuring.
The report warns that published pricing guidelines do not capture such runaway costs. BakedWith estimates a basic chatbot at $20 to $50 per month, a mid-level agentic assistant at $100 to $500 per month, and a custom enterprise agentic solution at more than $10,000 upfront. But these numbers exclude under-supervised agents, which can erase return on investment. The report said that each action was rational in isolation but the cumulative cost was not, underscoring the need for organizations to manage deployed AI agents and monitor their proliferation.