Capability C6
AI extraction and summarization
The AI reads the records and returns structured fields and summaries. Every output is linked to its source and carries a confidence value. A person approves each output before it is used.
Reference builtReference implementation existsExtraction, cited summaries, and streaming answers have reference implementations.
What happens
- Records are read and fields are extracted, each with its source.
- Summaries are drafted, with a citation for each statement.
- Outputs below the confidence threshold are sent to a person.
How the AI platform addresses it
- Extraction runs on the smallest model that can handle the document type. A language model is used only for summaries and drafts.
- A citation check removes any statement without a source.
- Guardrails check the purpose of the request, the access rights of the caller, and the output format before an output is shown.
AI role
Extraction by document and neural models, and summaries by language models, with citation checks and a confidence threshold.
Approval
A person approves each output. The approval is recorded with the output version.
Systems involved
- Document Intelligence
- Azure AI Search
- Azure AI Foundry
- Audit ledger
Reference components
- Medical records extraction flow
- Clinical visit scribe
- Claims answer streamer with citations