Measure Where the Waiting Time Actually Goes
The model may not be the slowest part of your AI workflow. Retrieval, tool calls and repeated validation attempts can consume most of the waiting time.
For a synthetic support task, record the duration of input preparation, the model request, each tool call and final validation. Include the time until a usable answer, not just the first visible token. Keep the same acceptance conditions when comparing configurations.
Inspect the slowest representative cases as well as the average. A workflow that feels quick on easy questions can stall on exactly the exceptions users care about. Improve the measured bottleneck first. Shorter output, fewer unnecessary serial calls or better retrieval may help more than switching to a different model without understanding the delay.
