Amazon Augmented AI (Amazon A2I) is a managed human-review service for machine-learning predictions. It helps teams route selected results to people when confidence is low, a required field is missing, or a review policy calls for a human check. For the AWS Certified AI Practitioner (AIF-C01) exam, recognize A2I as a human-in-the-loop workflow—not a model-training service or a general-purpose chatbot.
What problem does Amazon A2I solve?
Machine-learning systems can produce predictions that are uncertain, incomplete, or sensitive enough to require human verification. A fully automated process may pass an incorrect result downstream; sending every result to a person, however, can be slow and expensive. A2I provides configurable human-review workflows so an application can send selected predictions to a workforce and collect the reviewed output.
Think about a loan-processing application that extracts fields from scanned forms. Most fields may be clear, but a poorly scanned date or handwritten amount may have a low confidence score. A review rule can send that item to a human reviewer instead of treating the machine’s prediction as unquestionably correct.
How an A2I human-review workflow works
- Choose the input and review task. A workflow can use a supported built-in task type or a custom task for an application.
- Define when review is required. For example, trigger review when a form key is missing, a confidence score falls below a threshold, or a business rule selects a sample for audit.
- Configure the flow definition. Specify the workforce, worker instructions and task UI, IAM permissions, and the S3 location used for input and output data.
- Start a human loop. A built-in integration can initiate review when its configured conditions are met. A custom application can call the A2I Runtime API to start a human loop.
- Use the reviewed result. The application retrieves the human-review output and applies its own validation and downstream business logic.
A2I does not magically guarantee that the final answer is correct. Review quality still depends on clear instructions, qualified reviewers, appropriate sampling or confidence thresholds, access controls, and validation of the returned result.
Built-in task types and custom workflows
AWS documents two built-in A2I task types:
- Amazon Textract – key-value pair extraction: human review can be triggered for specified form keys when confidence is low or a key is missing.
- Amazon Rekognition – image moderation: human review can be triggered based on moderation-label confidence.
A2I also supports custom human-review workflows for an application’s own machine-learning predictions. In a custom workflow, the application defines the review conditions and uses the A2I Runtime API to start the human loop. Do not confuse the two built-in task types with the broader set of services or models that can be connected through a custom implementation.
Core components to recognize
- Flow definition: the configuration for the human-review workflow, including the task type, workforce, worker task template, review settings, and output location.
- Human loop: an individual review process that sends one or more tasks to human workers.
- Workforce: the people who perform the review. AWS documentation describes private workforces and vendor workforces; check current availability and eligibility before planning a new deployment.
- Worker task template: the instructions and interface presented to reviewers so they can inspect the input and submit a result.
- Amazon S3: stores workflow input and review output. The bucket and workflow resources must be configured in compatible AWS Regions.
- IAM permissions: control who can create workflows, invoke the relevant service APIs, access S3 objects, and start or manage human loops.
Amazon A2I vs. SageMaker Ground Truth
| Service | Primary purpose | Exam clue |
|---|---|---|
| Amazon A2I | Human review of selected machine-learning predictions or outputs | Send low-confidence or selected results to people for verification |
| Amazon SageMaker Ground Truth | Label datasets used to prepare training data | Workers annotate images, text, or other data so a model can be trained |
| Amazon SageMaker Clarify | Help detect potential bias and explain model behavior | Analyze bias or explainability—not route predictions to reviewers |
| Amazon Textract | Extract text and structured information from documents | Read forms, tables, and document fields; A2I can add human review to supported Textract workflows |
Memory hook: Ground Truth labels data for model development; A2I sends selected predictions for human review; Clarify analyzes bias and explainability; Textract extracts document content.
Exam traps and design considerations
- Do not select Ground Truth for every human task. If the requirement is to verify a prediction produced during an operational workflow, A2I is the historical service concept being tested. If the requirement is to label a dataset for training, Ground Truth is the closer fit.
- A2I is not a model. It coordinates review; it does not train a model or replace the service that produced the prediction.
- Review can be conditional. A2I is useful when only selected predictions need attention, such as low-confidence fields or samples chosen for quality checks.
- Permissions matter. The workflow needs suitable IAM permissions and access to the S3 data. Permissions for A2I do not automatically grant permission to invoke every integrated service API.
- Check Region alignment. AWS notes that Region availability can differ across A2I and integrated services. Keep related resources in supported Regions.
- Know the current service status. A2I is no longer open to new customers and AWS does not plan new features. Avoid presenting it as an unrestricted greenfield choice.
Test yourself
Practice question 1
A company uses Amazon Textract to extract key-value pairs from loan applications. It wants a person to inspect a specific field when its confidence score is below a defined threshold. Which service is designed to support this human-review workflow?
- Amazon SageMaker Clarify
- Amazon Augmented AI (A2I)
- Amazon SageMaker Ground Truth
- Amazon Comprehend
Click to reveal the answer & explanation
Correct answer: B. A2I supports configured human-review workflows for supported Textract tasks. Clarify focuses on bias and explainability, Ground Truth labels datasets, and Comprehend analyzes text.
Practice question 2
A team needs workers to label thousands of images so the labeled dataset can be used to train a computer-vision model. Which AWS service is the best match?
- Amazon Augmented AI (A2I)
- Amazon Rekognition
- Amazon SageMaker Ground Truth
- Amazon SageMaker Clarify
Click to reveal the answer & explanation
Correct answer: C. Ground Truth is used to label training datasets. A2I is for human review of selected predictions in an inference workflow.
Key takeaway
For AIF-C01, remember the distinction between preparing labeled training data and reviewing live or operational predictions. Amazon A2I coordinates human review for supported workflows and custom applications, while Ground Truth supports data labeling. Also remember the current AWS availability notice: existing A2I customers can continue using the service, but it is no longer open to new customers and no new features are planned.
Related AWS topics
Amazon Textract · Amazon SageMaker Clarify · Amazon SageMaker · Amazon Rekognition · SageMaker Ground Truth
✦ CloudExamPro Premium
Ready for the AWS Certified AI Practitioner (AIF-C01)?
Move beyond memorizing terms. Practice with realistic, scenario-based questions and detailed explanations that help you understand the AWS AI services and concepts behind each answer.
Original practice questions. No dumps. No shortcuts.