A financial services company is building a serverless order-processing workflow on Azure. Each order must run through a sequence of steps: credit check, inventory reservation, payment capture, and shipping notification. Some steps call external APIs that can take several minutes, and one step requires waiting for a human approval that may arrive hours later. The team wants to keep code in a serverless model, maintain state across each step automatically, and support compensation logic if a later step fails. Which compute approach best meets these requirements?