I’ll disagree with @kakeru on one thing: you can safely ignore most trad ML math, but if you skip all numbers/diagrams you’ll get tripped up on a few “which phase is this?” style questions. You don’t need gradients, but you should recognize a very simple workflow picture and basic eval metrics.
To untangle the noise, think of prep in 3 buckets instead of a long step list:
1. “Boardroom” skills the exam secretly loves
This cert is more “AI product manager 101” than “ML engineer lite”. A lot of questions are basically:
A business stakeholder wants X, with constraint Y. Which approach / AWS service is most appropriate?
So practice reading a question and instantly labeling it as:
- Business framing
- What is the actual goal?
- Is this classification, prediction, summarization, recommendation, translation, etc.?
- Risk / compliance
- Are they worried about PII, region, regulation, or brand safety?
- That’s your cue for responsible AI / data residency / guardrails.
- Cost / complexity
- Are they a small team, non-technical, or just experimenting?
- That nudges toward managed services and no‑code tools, not custom training.
A good exercise:
Take 5–10 random AI use cases from real life (YouTube comments moderation, HR resume screening, marketing copy, customer email triage, etc.) and force yourself to answer 3 things for each:
- What kind of ML problem is this?
- Would I use a classic AI API (Comprehend, Rekognition, Translate…) or Bedrock / Amazon Q?
- Any obvious responsible AI concerns?
This mirrors how the exam questions are structured way more than rote memorizing service names.
2. Minimal AWS surface area you actually can’t skip
@kakeru gave a solid service list, so I won’t rehash it all. Instead, here’s where people over or under study:
Understudied but important:
-
Amazon Q vs Bedrock
Too many people treat them as the same “chatbot thing.”- Q is “AI assistant for your org / your AWS stuff / your business apps”
- Bedrock is “platform to build your own AI solutions with FMs”
If the question smells like “help staff work with AWS better” or “internal productivity,” that’s usually Q, not Bedrock.
-
Where data lives + how AI touches it
You just need to know:- S3 commonly holds training data, documents, KB content
- RDS / DynamoDB as typical app data sources
- Vector store concept with Bedrock KBs (not implementation details)
Questions may say “large collection of PDFs” or “transactional app data” and expect you to know which storage choice is even plausible.
-
Evaluation / limitations
Especially for gen AI questions:- “Hallucinations” vs “incomplete context” vs “wrong prompt”
- When to use RAG / knowledge bases instead of hoping the base model “knows everything”
Overstudied and mostly useless:
- In‑depth IAM policy JSON
- Specific instance types for training
- Precise pricing math
- Glue / EMR / Redshift details for data engineering
If your notes include three types of VPC endpoints, you’re studying for a different exam.
3. Smarter practice instead of more content
Instead of binging 3 full video courses (which all repeat each other), do this:
A. “Reverse engineer” questions
Grab any AI Practitioner‑style practice questions you find (official set, or ones explicitly updated for this cert). For each question, try to label:
- What concept did it test?
- Was it service‑selection, risk, cost, or conceptual AI/ML?
- Why were the wrong options bad choices?
You’ll start seeing patterns like:
- Two options technically work, one is “overkill,” one is “wrong service,” one is “not responsible / risky,” and one is the “just right” AWS‑friendly answer.
B. Hands‑on, but small and surgical
You do not need to build a full project. Instead, spend one short session per area:
- A quick Bedrock playground session: try different prompts and temperatures, observe output change.
- Run a sentiment or entity detection job in Comprehend.
- Use Textract or Rekognition on a sample file.
The point is not depth. The point is that “service mental images” stick better than reading bullet points. It also makes the “which service?” questions feel obvious.
C. One‑page “brain sheet”
Rather than 20 pages of notes, compress everything into a single doc/table with:
- Row: use case
- Columns:
- Problem type (summarization, forecasting, translation, etc.)
- Recommended AWS service(s)
- Gen‑AI vs traditional AI
- Any risk / privacy gotcha
If you can fill that out without looking things up, you’re very close to ready.
Last thing: don’t underestimate responsible AI. On this exam it’s not fluff; it’s a scoring category. Think like a cautious product owner:
- “Could this harm users or leak data?”
- “Do I need guardrails, PII controls, region selection?”
- “Should I keep humans in the loop for sensitive decisions?”
If you can talk through those tradeoffs out loud, pick the right AWS service on a multiple‑choice list, and distinguish classic AI APIs from Bedrock / Q, you’re aligned with how the exam is actually written, not how random YouTube thumbnails make it sound.