A county deploying a custom document-classification model to a SageMaker endpoint needs to pick an instance family that balances the permitting chatbot's latency needs against monthly compute cost. What should guide this decision?

Benchmarking a handful of candidate instance families under realistic request patterns measures how each actually performs for this specific document-classification model, and comparing that measured latency and throughput against each family's price lets the county pick the option that meets the permitting chatbot's response-time requirement at the lowest sustainable cost. Defaulting to the largest GPU-backed instance ignores that document classification, especially a lighter traditional model, may not need or benefit from that much compute, wasting budget without necessarily improving latency. Accepting whatever the console lists first treats a UI default as a recommendation, when defaults are typically generic starting points, not a workload-specific analysis. Choosing the cheapest, smallest instance assumes complexity correlates neatly with resource needs, but a classification model handling high request volume or larger documents can still need meaningful compute to hit a latency target, regardless of task type. Scope caveat: instance-family selection should be revisited if traffic patterns or the model itself changes materially, since a benchmark done once can go stale. A concrete operational check: load-test the top two candidate instance families with production-like request volume and confirm the latency numbers hold before committing to one for production.