AI-103 Practice Question 30: Unstructured Document Extraction
Microsoft Azure AI Apps and Agents Developer Associate (AI-103) · Test 3 · question 10 of 10
A health insurer receives referral letters from several hundred clinics. They are free-text, multi-page and narrative; no two practices lay one out the same way. For each letter the insurer needs the referring specialty, the urgency, and a prior-authorization flag that is almost never stated outright and has to be inferred from what the letter describes. There is no labelled corpus and no budget to create one, and low-confidence extractions must be routed to a human.
Which approach matches Microsoft's guidance?
Show answer and explanation
Answer:
Microsoft's decision guide names this scenario almost word for word: "Custom extraction without labels, or for unstructured documents (for example, policies, referral letters, doctor notes)" → Content Understanding custom analyzer (zero-shot or with knowledge source). Two details close the case. Content Understanding can infer fields that are not explicitly present; the guide's own examples are deriving a contract end date from a start date and duration, or a jurisdiction from the parties' addresses, which is what the prior-authorization flag requires. And its confidence and grounding output (estimateFieldSourceAndConfidence) is designed to route low-confidence extractions to human review while high-confidence ones flow straight through.
- A is wrong because Document Intelligence custom models are the recommendation for *highly structured* documents with labelled samples. Labelling five samples for each of several hundred free-form layouts is the cost the insurer explicitly cannot pay. - C is wrong because layout analysis returns structure (paragraphs, tables, reading order) not semantics. Regular expressions over narrative prose from hundreds of authors are brittle, and no regular expression can *infer* an unstated flag. - D is wrong because it is the "build your own" column of the same comparison: no confidence scores, no grounding, and full ownership of prompt tuning, orchestration and scaling, so the insurer must either accept every result or review every result.
Reference: Choose the right Azure AI tool for document processing