The AIF-C01 exam and the AWS AI stack: what is tested and where each service sits

It has 65 questions of which 50 are scored and 15 are unscored trial items, runs 90 minutes, and reports a scaled score from 100 to 1000 with 700 as the pass mark.

• Question types: multiple choice, multiple response, ordering, matching and case study.
• No penalty for wrong answers, so nothing should be left blank; multiple response has no partial credit.
• A 30-minute extension is available for non-native English speakers, requested before the sitting.

The scaled score is not a percentage, so 700 does not mean 70 percent of the questions right.

Fundamentals of AI and ML at 20 percent, Fundamentals of generative AI at 24, Applications of foundation models at 28, Guidelines for responsible AI at 14, and Security, compliance and governance for AI solutions at 14.

• Domains 2 and 3 together are 52 percent and are dominated by Amazon Bedrock and its features.
• Domain 1 is definitional; domains 4 and 5 are lists of named tools and standards.
• Each domain breaks into numbered task statements that list the in-scope terms.

The heaviest single domain, applications of foundation models, is also the most scenario-driven: RAG versus fine-tuning, prompts, agents and evaluation metrics.

Infrastructure at the bottom (EC2 GPU instances, Trainium, Inferentia, S3, HyperPod), the ML platform above it (Amazon SageMaker), AI services above that (Bedrock and the task services such as Rekognition, Comprehend and Transcribe), and applications on top (Amazon Q, PartyRock).

• Each layer trades control for convenience: SageMaker gives full control over a custom model; Bedrock gives an API; Q gives a finished assistant.
• Scenario keywords map to layers: "custom model" is SageMaker, "least operational overhead" is an AI service, "employees ask about internal documents" is Q Business.
• Trainium and Inferentia are AWS's own chips for cost-efficient training and inference.

Amazon Q Developer is itself built on Bedrock, an application-layer product sitting on a service-layer one.

Retrieval augmented generation with Amazon Bedrock Knowledge Bases, which indexes the documents into a vector store and grounds the model's answers in retrieved passages with citations.

• The phrases "own documents", "citations" and "without training" all point at retrieval rather than fine-tuning.
• Amazon Q Business is the finished-application alternative if no custom app is being built.
• Fine-tuning would teach style, not facts, and would need retraining as documents change.

Bedrock does not use customer data to train its base models, which is why the "no training" constraint is satisfied by design.

Bedrock is a serverless API where you pay per token and never see infrastructure; JumpStart deploys a model onto SageMaker endpoints that you pay for by the hour and can fine-tune with full control over the weights.

• "Serverless", "pay per use", "quickest to start" means Bedrock.
• "Open-source model", "deploy to an endpoint", "full control" means JumpStart.
• Both can fine-tune, but Bedrock's custom models then need provisioned throughput.

JumpStart lives inside SageMaker Studio and offers one-click deployment plus notebooks, which is the giveaway phrase in a question.

"Least operational overhead", "fully managed", "no ML expertise", "quickly" and "without managing infrastructure" all point up the stack to an AI service or Bedrock.

• "Custom model", "own training data", "data scientists" and "full control" point down to SageMaker.
• "No code" points at SageMaker Canvas or PartyRock.
• The modality then picks the service: images to Rekognition, documents to Textract, text to Comprehend, speech to Transcribe.

When both a managed service and a SageMaker build would work, the exam wants the managed one unless the scenario needs something the managed service cannot do.

SageMaker Model Monitor detects data and model quality drift on a deployed endpoint; SageMaker Clarify explains predictions through feature attributions and also measures bias.

• Model Monitor compares live traffic with a baseline and raises alerts on drift.
• Clarify runs before training (data bias), after training (model bias) and for explainability.
• Amazon Augmented AI adds human review for low-confidence predictions.

Model Monitor can also watch bias drift and feature attribution drift, using Clarify under the hood.

Artifact hands out AWS's own compliance reports; Audit Manager collects evidence from your account against a framework; Config records and evaluates resource configurations against rules; Inspector scans workloads for software vulnerabilities; CloudTrail logs every API call for audit; Trusted Advisor checks the account against best practices for cost, security and performance.

• "Reports for our auditors" is Artifact; "continuous evidence" is Audit Manager.
• "Was this bucket ever public" is Config; "who deleted the model" is CloudTrail.
• "Unpatched CVE on the instance" is Inspector.

These six appear together as options in domain 5 questions, so the one-line distinctions are worth memorising.

They raise the temperature, and possibly top-p or top-k, which widens the set of tokens the model samples from and makes output more varied.

• Lower temperature makes output more deterministic, which suits extraction and classification.
• These are inference parameters, changed per request, not training settings.
• Maximum tokens and stop sequences control length, not creativity.

Temperature of zero does not guarantee identical output on every call, because the exam's own term for the model's behaviour is nondeterminism.

A real AWS service placed in the wrong role: Macie where Comprehend is wanted, Inspector where Audit Manager is wanted, fine-tuning where retrieval is wanted, Bedrock where a classical ML service such as Forecast fits.

• Every option is a real service, so "does this exist" never eliminates anything; "does it fit this constraint" does.
• Two options often both work, and the constraint word (cheapest, least effort, no internet) picks one.
• Reading "AI" as "generative AI" sends classical problems to Bedrock wrongly.

Reading the question sentence before the scenario and underlining the constraint word is the practical defence.

It means the customer's prompts, documents and outputs must stay private and never improve the provider's base model; Amazon Bedrock satisfies this by design, and adding knowledge through a Knowledge Base or prompts trains nothing at all.

• Bedrock keeps data encrypted in the customer's account and region and does not share it with model providers.
• Fine-tuning on Bedrock produces a private copy of the model, still not shared.
• PrivateLink endpoints keep the traffic off the public internet if the scenario also asks for that.

The same sentence often appears together with "citations", which confirms retrieval as the intended answer.

They are AWS-designed chips on the infrastructure layer, Trainium for training and Inferentia for inference, offered as a lower-cost, energy-efficient alternative to GPUs.

• They appear in questions about cost-efficient or sustainable model training and serving.
• SageMaker and Bedrock can run on them without the customer choosing hardware.
• Instance sizing and cluster design are out of scope.

Responsible-AI questions about environmental impact accept "use purpose-built AI chips and smaller models" as a sustainability measure.

Roughly in proportion, but with extra time on domains 4 and 5, whose 28 percent combined is made of named tools and standards that must be memorised rather than reasoned out.

• Domain 3's 28 percent is mostly Bedrock, RAG, prompting and metrics, which are intuitive once seen.
• Domain 1 is definitions and a service catalogue; flashcards work.
• The Bedrock article pays off across all five domains.

Working each article's question set aloud is closer to the exam than re-reading, because the exam is recognition under a time limit.

Define the business goal, collect data, explore and pre-process it, engineer features, train the model, tune hyperparameters, evaluate, deploy, then monitor and re-train.

• The generative-AI equivalent is the foundation model lifecycle: data selection, model selection, pre-training, fine-tuning, evaluation, deployment, feedback.
• The RAG query flow is another ordering favourite: embed the question, search the vector store, augment the prompt, generate.
• Ordering questions give no partial credit, so the whole sequence must be right.

Article 4 maps a SageMaker feature onto each stage, which is how the exam usually asks about it.