Amazon Augmented AI (A2I)
Human review loops for low-confidence ML predictions.
❓ What is it?
A service for inserting human review into ML workflows: define confidence conditions, and predictions that fail them route to human reviewers (your team, vendors, or Mechanical Turk) whose verdicts return to your application.
💡 Why does it exist?
No model is right 100% of the time, and some decisions (content takedowns, ID verification, loan documents) are too costly to get wrong silently. Human-in-the-loop buys accuracy where it matters and produces fresh labels for retraining.
⏱️ When should you use it?
Use it when predictions below a confidence threshold need human eyes — moderation appeals, document field verification, random sampling audits of model output.
🗺️ Where does it fit?
Immediately after inference: built-in integrations trigger from Textract and Rekognition results, or wrap ANY model (including SageMaker or third-party) with a custom human-loop start call.
🔌 How do you integrate it?
Define a flow definition (worker task template + workforce + trigger conditions); reviews land in S3 as structured JSON your pipeline consumes; reviewed items can feed back into training data.
🧩 Commonly integrated with
🎯 Exam angle (AIF-C01)
- A2I reviews predictions AFTER inference; Ground Truth creates labels BEFORE training — the exam contrasts them.
- Trigger phrase: "route low-confidence predictions for human review" → A2I.