A company produced two models: Model A detects fraudulent transactions and will be invoked on every transaction; Model B forecasts next-month sales and will be called once per month. Both must be deployed to production using SageMaker AI. Which hosting approach should the company use?

Model A requires immediate predictions for every transaction, making a real-time endpoint the appropriate choice for low-latency inference. Model B forecasts sales once per month, meaning it does not require real-time inference. Batch transform is ideal for infrequent, large-volume inference tasks like this, as it processes data asynchronously and efficiently without maintaining a persistent endpoint. Hosting both models in a single container behind one real-time endpoint would be inefficient for Model B and potentially overload the endpoint. Hosting Model A on an asynchronous endpoint would introduce unacceptable latency for fraud detection. Using batch transform for Model A would also introduce significant delays, compromising its effectiveness.