AWS is urging teams to measure AI-agent cost per successful task rather than comparing models only by token price. A cheaper model can still cost more if it tokenizes inputs poorly, produces excessive output, retries tool calls or uses long reasoning traces.
The guide’s illustrative example compares a lower-priced model requiring three turns with a higher-priced model completing the job in one. Repeated context and extra output make the nominally cheaper option cost $0.0042 per task, versus $0.00145 for the alternative. These are teaching figures rather than results from a named production deployment.
For real measurement, AWS recommends recording every workflow under one trace ID, including model tokens, retries, tool calls, human-review minutes, escalation and whether the result was accepted or reopened. Cost per call can flatter systems that loop or fail; dividing total spend by successful outcomes exposes that waste. Human time must remain in the calculation even when an agent performs the first draft.
Return on investment also needs a baseline collected before launch and a holdout or matched control group. Teams can then compare completion time, rework, total cost and business outcomes rather than attributing every improvement to the agent. The framework recommends scaling only when quality and value meet thresholds, tuning ambiguous results, and retiring agents that remain expensive without measurable benefit.