AIF-C01 Practice Question 29: Human Review Low Confidence
AWS Certified AI Practitioner (AIF-C01) · Test 3 · question 9 of 10
A claims team extracts fields from scanned forms with a document analysis model. Most scans are clean, but poor handwriting and faint copies produce extractions the model itself returns with a low confidence score, and those are currently written straight into the claims system, where they cause costly corrections later. What is the most appropriate first step?
Show answer and explanation
Answer:
This is the human-in-the-loop pattern. Amazon Augmented AI (Amazon A2I) lets you send low-confidence predictions, or a random sample for ongoing auditing, to human reviewers before the results are used, and it supplies the review workflow, task templates and reviewer management so you do not build that yourself. Poor scans and bad handwriting are the textbook example: the model handles the clean majority, humans handle the uncertain remainder (option C).
- A is wrong because retraining takes time and data, and no amount of it eliminates genuinely illegible handwriting. It is a longer-term improvement, not the first step, and it leaves bad values flowing into the claims system meanwhile. - B is wrong because it moves in the wrong direction. Accepting more low-confidence extractions increases exactly the errors the team is paying to correct. - D is wrong because it pushes the model's uncertainty onto the customer. A reviewer can usually read the field in seconds, so rejecting the whole form adds delay and frustration for no accuracy gain.
Reference: Using Amazon Augmented AI for Human Review