Turning voicemail orders into structured data

A Retail business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had four focus areas, each central to a proof of concept the client could trust with real orders.

An user requests a pre-signed URL through Amazon API Gateway and a Lambda function, then uploads the voicemail recording directly to an Amazon S3 bucket. The upload triggers an AWS Step Functions workflow: a Lambda function places the event on an Amazon SQS queue for controlled processing, retrieves the audio and sends it to Amazon Transcribe, respecting service limits by returning the message to the queue if too many files are already in flight.

The proof of concept met its success criteria: a pipeline that accurately extracts ordered items from real voicemail recordings, assesses their quality, validates them against historical data and returns a structured JSON output, deployed into the client's own AWS environment.

• Orders extracted automatically — Amazon Transcribe converts each voicemail to text and a Bedrock model extracts the items into a structured list ready for confirmation.
• Quality flagged — The system assesses each recording and distinguishes clear orders from those affected by noise or other issues.
• Grounded in history — A Bedrock Knowledge Base backed by OpenSearch compares each order with the customer's past orders and recommends whether it should be checked.
• Resilient and cost-aware — Cross-region replication protects the data, and lifecycle policies move inactive files to Glacier to keep storage costs down.
• Fully traceable — Every stage of the pipeline is logged in DynamoDB, with CloudWatch monitoring across the workflow.
• Handover — Live demonstrations, recorded walkthroughs and all code in source control left the client's team able to operate and extend the service.

With a proven, serverless pipeline in place, the client has a foundation it can extend across more of its order intake, moving from a proof of concept towards automated, AI-assisted order capture at scale.