AI Cost Overruns Hit 78% of Companies, Survey Finds

The invisible expense of AI deployment

Most organizations know whether their AI systems are operational. Far fewer can explain what those systems cost in the past hour—or predict what they'll spend tomorrow. A survey of 218 technology leaders reveals the scale of the problem: 78% reported unexpected charges tied to AI or consumption-based pricing in the last year, according to findings first reported by Fast Company. The consequence was severe: 61% were forced to cancel projects to cover the overruns.

The root cause is a fundamental shift in how software is priced. Traditional enterprise software operated on fixed contracts with predictable annual costs. AI systems run on usage-based pricing, charging for every token—the small units of text a model reads and writes. As customer interactions scale and AI agents perform more complex tasks, costs rise automatically. The billing meter runs continuously, but most finance teams only see the total when the monthly invoice arrives.

Why it matters

AI is becoming cheaper per unit while total bills climb—a paradox that's forcing companies to cancel promising initiatives mid-flight. Organizations that treat AI cost as a post-hoc reconciliation exercise rather than a real-time operational metric are discovering too late that their profitable-looking products have become money losers. The gap between technical success and financial viability is widening, and it's catching leadership teams off guard.

The agent amplification effect

AI agents compound the cost challenge in ways simple chatbots never did. A chatbot answers a question with a single model call. An agent pursues a goal through multiple steps—retrieving documents, invoking tools, and retrying failed attempts. Gartner forecasts that agentic tasks will consume five to 30 times more tokens than equivalent chatbot exchanges.

Several factors multiply the expense invisibly. Input tokens—the context and instructions fed to the model—often exceed the cost of generating responses. Long context windows and retrieval-augmented generation workflows increase token consumption dramatically. An agent stuck in a retry loop can execute dozens of paid API calls attempting to recover from a single error, all without triggering traditional operational alerts.

Gartner projects that by 2028, AI coding costs will surpass the average developer's salary, driven by rising token consumption and the shift to consumption-based pricing. The firm also forecasts that running inference on a one-trillion-parameter model will cost providers more than 90% less in 2030 than it did in 2025—yet enterprise AI spending continues rising because usage growth outpaces price declines.

Building cost observability into operations

Engineering teams already practice observability—monitoring metrics, logs, and traces to understand system behavior. AI operations must extend this discipline to include consumption as a first-class signal. Tracking latency and uptime while ignoring token usage is equivalent to monitoring a vehicle's speed but never checking the fuel gauge.

The practical implementation requires capturing usage data in near real-time and attributing it to specific workflows, features, or customers. Distributed tracing becomes critical when an agent fans out into a chain of model calls—teams need visibility into which step consumed the most budget. The objective is transforming an opaque monthly total into actionable signals available while spending is still occurring.

Translating tokens into business metrics

Raw token counts mean little outside engineering teams. The metrics that matter to executives tie cost to outcomes: cost per customer interaction, cost per resolved support ticket, cost per completed transaction. These figures reveal whether AI features generate positive returns.

Sasi Kiran Malladi, a principal technical account manager at Amazon writing for Fast Company, recommends four practices for leaders deploying AI features: implement usage dashboards the same week features launch, express costs as per-business-outcome rather than raw tokens, set spending spike alerts similar to downtime alerts, and assign clear ownership of AI unit economics.

Cost architecture becomes a design decision rather than an afterthought—routing simple tasks to smaller models, caching repeated requests, and choosing traditional code over model calls where appropriate. The organizations that succeed will be those that see what AI costs them in time to act, while features remain successful rather than becoming expensive surprises.

These findings and recommendations were first reported by Fast Company.