Copyright (c) 2026 MindMesh Academy. All rights reserved. This content is proprietary and may not be reproduced or distributed without permission.

3.6.6. Human Oversight Workflows for AI Operations

💡 First Principle: When an AI-proposed action is consequential or irreversible, make the human decision a required state transition in a durable workflow, with a timeout and an audit record, rather than an instruction in a prompt.

Scenario: An agent drafts vendor payments and infrastructure changes. Payments over $10,000 and any production change must be approved by a person, approvals can take up to three days, and auditors need to see who approved what.

The Step Functions callback pattern.
  1. A Standard workflow invokes the model or agent (Step Functions has an optimized "Amazon Bedrock" integration and an "AgentCore" integration) to produce a proposed action and a risk score.
  2. A Choice state auto-approves low-risk proposals and routes the rest to an approval task.
  3. The approval task uses .waitForTaskToken: it sends the task token with the proposal to the reviewer (through "SNS" email, an "SQS" queue read by an approval app, or a "Lambda" that posts to a ticketing tool) and pauses.
  4. The reviewer's choice calls SendTaskSuccess (approve) or SendTaskFailure (reject); the workflow then executes or discards the action.
  5. TimeoutSeconds or HeartbeatSeconds fail the task if nobody responds, so a Catch can escalate instead of waiting forever.

Why this design. A Standard workflow can pause for up to a year without running compute, and its execution history, plus the CloudTrail record of the SendTaskSuccess caller, is the approval evidence. Express workflows only support the Request Response pattern, so they cannot wait for a callback. Task tokens must be returned by principals in the same account as the workflow, so a cross-account approval application needs a role in that account.

Related controls and what they are not:
  • Policy in AgentCore gives an automatic, deterministic deny or allow on a tool call; it is a guard, not a human review.
  • Bedrock Evaluations human jobs rate model outputs offline for quality; they do not approve live actions.
  • Amazon Augmented AI (A2I) and Bedrock Agents Classic user confirmation are closed to new customers, so they are not the design for a new build.

⚠️ Exam Trap: Distractors often have a Lambda function poll a DynamoDB "approved" flag in a loop, or use an Express workflow for the approval. Polling wastes compute and loses the built-in timeout and audit history, and Express workflows cannot use .waitForTaskToken at all.

See how it connects
Alvin Varughese
Written byAlvin Varughese
Founder•20 professional certifications