1.4. Reflection Checkpoint

Key Takeaways

• Treat AI as a probabilistic forecaster, not a calculator: design for a known error rate, decide where automation is safe and where human review is required, and expect performance to need monitoring.
• Trace value through the whole chain (data, model, prediction, decision, action, outcome, feedback); most failures are business links, not model links.
• Work backwards from a measurable outcome; choose the technology last, and be willing to conclude that AI is not needed.
• Keep decisions with the right owners: the strategist sets problems, value measures, risk tolerance, and adoption plans; technical teams choose algorithms, pipelines, and configuration.
• Place AWS services by layer: Amazon Quick to use AI now (seat-based), Amazon Bedrock to build with existing foundation models (consumption-based), Amazon SageMaker AI to build custom models (instance-based), AWS Marketplace to buy.
• Under shared responsibility, AWS secures the cloud; you remain accountable for data, access, configuration, appropriate use, and outputs. CAF reminds you that readiness spans six perspectives, three of them business-facing.

Connecting Forward

Phase 2 builds the vocabulary that the rest of the exam assumes: what training and inference are, how AI, ML, and generative AI relate, why data quality decides outcomes, when rules beat AI, what makes an agent different from a chatbot, why models drift, and how prompting, RAG, and fine-tuning adapt generative AI to business needs. Each concept connects back to the value chain you just learned.

Self-Check Questions

• A vendor claims its AI "never makes mistakes." Which first principle does that claim violate, and what three questions would you ask before believing it?
• Your team has an accurate model, but the business result has not moved. Which links in the AI value chain would you inspect, and in what order?
• Why might a company choose Amazon Quick for one need and Amazon SageMaker AI for another in the same quarter, and how would the pricing models differ?
• A department head argues that AWS's compliance certifications mean the department's AI use case is compliant. What is missing from that argument?