Amazon SageMaker Clarify and Amazon Augmented AI (A2I) are both no longer open to new customers, and the AWS AI Practitioner (AIF-C01) exam guide does not depend on either. Its responsible AI domain tests practices (subgroup analysis, label quality, human audits) and the tools that remain: Amazon Bedrock Guardrails, Amazon Bedrock Model Evaluations and Amazon SageMaker Model Cards.
Last updated October 2026.
The SageMaker developer guide now carries a note on each service: "Amazon SageMaker Clarify is no longer open to new customers" and "Amazon SageMaker A2I is no longer open to new customers." Both notes are about new customers: an account that never used the service cannot start now.
The exam was already moving the same way. In the AIF-C01 version 1.1 guide, objective 4.1.7 asks you to "describe tools to detect and monitor bias, trustworthiness, and truthfulness" and gives analyzing label quality, human audits and subgroup analysis as the examples. Those are practices, not products. Objective 4.2.2 names Amazon SageMaker Model Cards and Amazon Bedrock Model Evaluations as transparency tools, and neither Clarify nor A2I appears in the guide's in-scope services list.
Service status from the Amazon SageMaker AI developer guide, checked 23 October 2026. Objectives from the official AIF-C01 exam guide, version 1.1.
What this means for study: learn what Clarify and A2I did, because the ideas behind them are tested. Do not spend time memorising their features as the answer to a question.
Domain 4 is 14% of scored content, about 7 of the 50 questions. Most of it starts from this list, and many questions describe a failure and ask which feature is missing:
| Feature | What it asks | Failure that signals it |
|---|---|---|
| Low bias | No systematic tilt for or against a group | Errors pile up on one group |
| Fairness | Similar people, similar outcomes | Same income and credit, different loan decisions |
| Inclusivity | Works for everyone it serves | Voice assistant fails on regional accents |
| Robustness | Stable on noisy or adversarial input | A typo flips the answer |
| Safety | No harmful output, resists misuse | Gives dangerous instructions |
| Veracity | Outputs are true, not just fluent | Confident hallucinated facts |
The trap pair is fairness and veracity. A biased loan decision is "wrong", but the failing feature is fairness, because similar applicants are treated differently. Veracity is about whether a statement is factually true.
Clarify computed standard bias metrics on training data and predictions. The exam guide describes the method instead, and the method works with any tool:
Repeat the subgroup analysis after launch on fresh predictions. A model that was fair at launch can drift as new customer groups arrive.
A2I sent low-confidence predictions to people for review. The principle is still on the exam, under human-centered design for explainable AI (objective 4.2.4): route predictions below a confidence threshold to a reviewer, require human review for medical, legal, financial and employment decisions, and review a random sample of confident decisions too.
For foundation models, Amazon Bedrock Model Evaluations includes a human evaluation workflow that uses your own team or an AWS-managed workforce. It suits subjective criteria such as tone or helpfulness, where automatic metrics fall short.
Checks input and output against content filters, denied topics, word filters, sensitive information filters and grounding checks. The managed answer for safety and veracity.
Automatic, LLM-as-a-judge and human evaluation of foundation models, including responsible AI metrics such as toxicity and stereotyping.
One versioned record per model: intended use, risk rating, training details and evaluation results, exportable for auditors.
AWS's own documentation of intended use and limitations for its AI services. Read before choosing a managed service.
Not as a named tool in the current objectives. Clarify is closed to new customers, and the version 1.1 exam guide tests bias detection as practices (label quality, human audits, subgroup analysis) and names Model Cards and Bedrock Model Evaluations as transparency tools. Know what Clarify did, but do not rely on it as an answer.
There is no single replacement service: the practice of routing low-confidence or high-stakes predictions to a person is built into the application workflow. For foundation models, Amazon Bedrock Model Evaluations offers human evaluation with your own team or an AWS-managed workforce.
Use subgroup analysis, label quality analysis and human audits. Breaking results down by group, rather than reporting one overall accuracy, is what reveals that errors fall unevenly on some people.
About 7 of the 50 scored questions. Guidelines for Responsible AI is 14% of the exam, the same weight as Security, Compliance, and Governance for AI Solutions.
The Certified AIF-C01 course covers Domain 4 in eight short chapters, from the six features to Guardrails, bias detection and Model Cards, each with exam-style questions.
Open the AIF-C01 course →