An insurance company's nightly job extracts fields from 200,000 scanned claim letters for next-morning reporting. It currently calls the synchronous API one document at a time using the largest model, and the run regularly overruns into business hours, competing with the interactive claims assistant for capacity. Which TWO changes are best? (Select two.)

Answer: B and D — Move the job to asynchronous batch processing, since nobody is waiting on any individual document's response / Evaluate a smaller model on a labelled sample of letters and switch if extraction accuracy holds on the hard cases

This is a throughput workload: batch processing (such as the Message Batches API) fits work nobody is waiting on and removes contention with interactive traffic (b). Model tier should be an evidence-based choice; if a smaller model holds accuracy on representative and hard cases, it cuts time and cost (d). (a) improves perceived latency for a human reader, who doesn't exist here, and doesn't shorten the run. (c) adds a full extra pass to a job that already overruns, without evidence it's needed.