Bedrock vs SageMaker, and Why the AI Practitioner Exam Keeps Blurring the Two

The AWS AI Practitioner exam makes you choose Bedrock or SageMaker, and the honest answer is that the line between them is thinner than the study guides admit. I made two videos here, one mapping the 15 AWS AI services and one breaking down the exam after I passed it, and this post is a standalone study aid so you get the map and my notes without needing to watch.

The name game

AWS named most of its AI services after what they do, so the names are half the study. Textract is text plus extract, pulling text and tables out of scanned documents. Rekognition looks at images and video. Comprehend reads text for sentiment and entities, with a Comprehend Medical variant for clinical notes. Transcribe turns audio into text and Polly does the reverse, text into speech, which is the trap the exam sets: Polly speaks, Transcribe listens, do not swap them. Translate handles languages and Lex runs the chatbot logic behind Alexa.

The two big ones

Bedrock is the vending machine: pick a foundation model like Claude, Llama, Mistral, or Titan, send prompts, pay per token, run no servers. SageMaker is the full kitchen: you pick instance types, control the training pipeline, and manage the endpoints. The cheat sheet I give in the video is simple. If the question says serverless or foundation model, that is Bedrock. If it says custom training or instance types, that is SageMaker.

The combinations

Services get more useful chained together. A voice chatbot is Transcribe as the ears, Lex as the brain, Polly as the mouth. RAG is Bedrock plus Knowledge Bases, where the Knowledge Base retrieves and Bedrock generates. A Bedrock Agent adds instructions, action groups on Lambda, and knowledge bases so a model can act, not just answer. Guardrails and temperature are concepts to know rather than services, with temperature as the creativity dial from zero, always the same answer, up to one, different every time.

Then the actual exam

I passed AIF-C01 with 13 AWS certs already done, and it is harder than Cloud Practitioner. Cloud Practitioner felt like a vocabulary test that asks you to name things. This one asks you to choose, often with several plausible answers where you pick the cheapest or fastest or most controlled option for a business outcome. Responsible AI runs all the way through it: fairness, bias, hallucinations, explainability, and human oversight. ML fundamentals are real too, including model scoring metrics and NLP scores like BLEU for translation and BERT for semantic similarity, plus security you already know from other AWS exams like IAM, KMS, and VPC endpoints. Model customization trade-offs come up as RAG versus fine-tuning versus prompt engineering. If you already hold AWS certs your foundation carries a lot of the weight and two to four weeks of focused study is enough. If it is your first AWS exam, do Cloud Practitioner first. My honest read: the real payoff is the knowledge, not the badge.