Follow ZDNET: Add us as a preferred source on Google.
ZDNET’s key takeaways
- AI agents risk spinning out of control and running up charges.
- One provider recounted its issues with uncontrolled agents.
- Look for high-end AI use case costs, not averages.
AI agents are proliferating, often beyond IT managers’ control, and the costs of unseen agentic activity can add up quickly, a recent analysis shows.
StackGen’s latest State of Reliability Report found that unregulated or semi-supervised agents are proving to be problematic on many levels.
Also: 4 AI development skills you need, according to Andrew Ng – and what experts say he’s missing
In at least nine documented cases over the past year, AI agents destroyed live company systems by wiping data and deleting databases on their own, using valid credentials. Their efforts were invisible to standard monitoring until the damage appeared.
The study, which analyzed 109,000+ incidents, also found that as failure modes multiply, recovery isn’t speeding up: median resolution times have been roughly flat since 2023, and the single most common fix is waiting for another company’s engineers.
Cumulative costs
From a financial perspective, AI agent costs can rack up. Recently, engineers at Revenium, an AI spending solutions provider, reported an agentic AI session where a developer left one AI coding assistant running on his laptop for four days straight. By the time anyone noticed, the agent had made 4,819 calls at a total cost of $3,762, without anyone’s knowledge.
While expenses vary considerably with use case, basic guidelines pulled together by BakedWith put the cost of a basic chatbot at $20 to $50 per month, a mid-level agentic assistant at $100 to $500 a month, and a custom enterprise agentic solution exceeding $10,000 upfront.
Still, these are upfront costs, and pricing does not account for runaway costs with under-supervised agents, which, of course, can quash ROI. This potential issue outlines how organizations must manage the AI agents they deploy and underscores the risks of their unchecked proliferation.
Also: ‘Specialists aren’t required’ anymore: How to stay valuable in an AI agent workplace today
Revenium cited an example from one of its customers, a mid-sized e-commerce company, which “saw AI agent infrastructure costs jump from $5,000 per month during prototyping to $50,000 per month in staging — a 10x increase driven by unoptimized RAG queries and recursive agent loops during high-volume periods.”
Every individual agent action looked like good engineering: “Each action was rational in isolation, but the cumulative cost was not.”
In another case, a team’s AI agents “entered an infinite conversation loop that ran undetected for 11 days, burning through $47,000 before anyone noticed,” the report added.
Also: Why replacing staff with AI backfires – and 5 ways smart leaders generate real value instead
“Two agents got stuck talking to each other while the team slept, while they worked, while they believed the system was just running smoothly.”
Lessons learned
In a recent internal study, Revenium’s engineering team turned attention to its own agentic AI practices and also found costly activity.
Here’s what the team learned from a particular unmonitored AI agent incident and related audit:
- A four-day session involving a stray AI agent came to $3,762: “On May 13, one of our developers opened an AI coding session on his laptop,” members of the company’s engineering team reported. “It stayed open for four days. By the time it closed, it had run 4,819 calls and cost $3,762. Nobody had budgeted for it. No alert fired.” There’s actually nothing unusual about this kind of session, they added. The issues arise because AI sessions are budgeted similarly to SaaS sessions, with a flat per-seat or per-token cost, tracked as an average. “That model breaks down once you look at what people are actually running,” the engineers added.
- Averages hide what AI actually costs: The median cost for agent-based work was $2.24 across 557 code-implementation tasks over 90 days, which looks obviously very sustainable. The most expensive task came in at $300.97. However, tracking averages hid the potential for runaway AI agent costs, as shown with the $3,762 session cost mentioned above. “If you’re managing an AI bill against an average, you have no visibility into what could happen tomorrow morning,” the engineers cautioned.
- Watch out for that top 1% of agentic runs: The top 1% of runs drove nearly half the spend, the engineers observed. Of 14,680 AI runs tracked over 90 days, the top 1% represented 46% of total spend. The top 5% of runs came to 77% of spend. The bottom 90% of runs were just 12% of AI spending. “AI spending lives in the tail of the distribution. SaaS cost controls aim at the wrong part of the curve,” the engineers stated.
- The costs come out of interactive AI, not automated AI: Among 10,005 interactive agentic sessions studied, the bill came to $109,118. For the 4,171 automated software development lifecycle tasks studied, the cost was $6,723. “The automated pipeline that implements and reviews pull requests costs under 6% of the bill,” the engineers reported. “The other 94% is engineers using AI through the day. That category rarely gets broken out as its own line item.”
- The cost of AI pull requests varies depending on unmeasured circumstances: Over 30 days, 12 engineers on the team each merged at least 10 pull requests, for a total of 1,721. Cost per merged pull request ranged from $4.05 at the lowest to $103.66 for the highest. The median was $16.59. “The spread reflects different work and different patterns of AI use,” the team members related. “We’re not arguing some engineers are ‘good’ and others ‘bad.’ We’re pointing out that this variance sits in a budget dimension almost nobody is measuring, and it’s bigger than any per-seat negotiation an enterprise will ever have with a vendor.”
- Cost scales with depth of use, not headcount: The team using AI grew fourfold from January to May, from seven to 28 engineers. Consumed AI value grew 420 times, while per-engineer consumption grew roughly 100×. “Per-seat budgets don’t see that acceleration.” The initial team of seven engineers using AI saw an API-equivalent value, tokens consumed via subscription-priced tools and priced at public API rates, of $109. By May, with 28 engineers using AI, they saw an API-equivalent value of $45,728.
While these numbers represent just one engineering organization, the Revenium team felt “the shape of the distribution holds broadly, and it tracks with patterns we see in early customer deployments.”
Also: Companies embracing AI the most are hiring more people – including entry-level
At the same time, “lower cost per pull request does not mean better engineering. We’re not suggesting anyone replace their higher-cost engineers. The point is about measurement: the variance is currently invisible to the budget owner, and the invisibility is the problem.”
