Why this guide exists and why you should care

If you are cramming for the AWS Machine Learning AI Specialty exam you will face questions that look like a multiple choice ransom note. The two usual suspects are Bedrock and SageMaker. One is great for managed foundation model inference with minimal babysitting, the other is built for custom training full MLOps and fine control. This guide walks you through a repeatable checklist so you can answer fast and sound smart in the exam room.

Step 1 read the requirement carefully

Stop skimming like the question is a terms and conditions page. Note whether the focus is training, inference latency, data privacy, cost or operational maintenance. Keywords to watch for include managed foundation model, prompt based inference, batch processing, hyperparameter tuning and MLOps pipeline. Those keywords are the exam whisperers.

Step 2 map requirements to service capabilities

Here is the blunt truth. Choose Bedrock when the question asks for managed foundation models, prompt based inference or low ops effort. Choose SageMaker when custom training, hyperparameter tuning, full lifecycle MLOps or model registry features are required. SageMaker is the heavy duty toolbox. Bedrock is the hands free option for generative AI workloads.

Quick mapping list

• Use Bedrock for pay per inference managed foundation models and quick prompt based prototypes.
• Use SageMaker for custom model training, distributed training, hyperparameter tuning and CI CD model deployment pipelines.
• Prefer SageMaker when you need specific instance types or fine grained control over training jobs.
• Prefer Bedrock when you want managed inference with less operational overhead and you can accept black box models.

Step 3 check security and data flow

Exam writers love security caveats. Always verify encryption at rest and in transit, use of VPC endpoints, and least privilege IAM roles for data access. For generative AI questions confirm data residency, content filtering and logging requirements. If the prompt mentions sensitive data or regulatory compliance you are probably pointing at a VPC enabled SageMaker setup or a specific data residency requirement with Bedrock configuration.

Step 4 consider cost and scalability

Cost logic will save your exam life more often than you think. If the workload is bursty and inference only a pay per call model using Bedrock might be cheaper. If heavy training or continuous batch jobs are required then provisioned instances in SageMaker with managed spot or distributed training will usually be more cost efficient. Also think autoscaling and cold start impacts on latency.

Step 5 answer with a concise rationale

Write one strong sentence that ties the core requirement to the service capability. Then add one sentence about a trade off. Example answer style for the exam would be something like this

• Because the question asks for prompt based inference with minimal ops overhead choose Bedrock. Trade off is less control over model internals and potential cost per inference differences compared to a provisioned SageMaker deployment.

Operational checklist to quote in the exam

• Confirm training versus inference focus
• Validate data residency and content logging needs
• Call out encryption in transit and at rest
• Note autoscaling or burst traffic requirements
• Estimate cost model pay per call versus provisioned instances

Exam time tips and final mantra

If you get stuck pick the option that best matches the primary requirement and then write one clear sentence about security or cost to support it. That single sentence often turns a guess into a defensible answer in the certification review. Remember to mention MLOps needs if the question implies lifecycle management or continuous delivery pipelines.

Use this framework in practice questions to build speed. You will sound decisive and technically correct which is basically what the exam wants from you. Also breathe. You can still pass even if you spill coffee on your notes.