AWS Certified AI Practitioner (AIF-C01)

AIF-C01 Exam Prep Notes

Based on the official exam guide scope as of October 2026About 30 min read

A complete summary of the AIF-C01 exam scope, short enough to reread in the last 30 minutes before the exam. Use it as a final check after working through all five domains in the learning course and finding your weak spots with the practice questions. Evaluation metrics, customization methods, and telling SageMaker AI tools apart are where points are most often lost, so focus on the tables and traps.

Exam facts and approach

The exam has 65 questions in 90 minutes, of which 50 are scored, and you cannot tell which 15 questions are unscored.
Results are reported as a scaled score from 100 to 1,000, and the passing score is 700.
Scoring is compensatory, so there is no per-domain passing line and you only need 700 overall.
There are four question types: multiple choice, multiple response, ordering (3 to 5 items), and matching (3 to 7 pairs), and every type except multiple choice scores only when fully correct.
Unanswered questions are scored as incorrect and there is no penalty for guessing, so answer every question.
The domain weights are Applications of Foundation Models 28%, Fundamentals of Generative AI 24%, Fundamentals of AI and ML 20%, Guidelines for Responsible AI 14%, and Security, Compliance, and Governance for AI Solutions 14%.
The target candidate uses AI rather than builds it, and coding models, tuning hyperparameters, building pipelines, and statistical calculations are out of scope.
Non-AI basics such as EC2, S3, Lambda, IAM, the shared responsibility model, and pricing models are also tested as prerequisites.
Decide between the last two options by requirement words such as lowest cost, least operational overhead, or explainability required.

Question cues and the right answer

Cue in the question	Direction of the answer
Predict from data with correct labels	Supervised learning
Group unlabeled data into similar clusters	Unsupervised learning (clustering)
Predict a numeric value such as sales or price	Regression
Predict a category such as fraud or not	Classification
A fixed, deterministic result is required and can be written as rules	Do not use AI or ML
The basis of decisions must be explained to regulators	An explainable traditional ML model
Run inference on large data in bulk overnight	Batch inference
Respond to each request immediately	Real-time inference
Large inputs that take time, and waiting is acceptable	Asynchronous inference
Rare calls, and no paying for idle time	Serverless inference
Use multiple foundation models through an API without managing infrastructure	Amazon Bedrock
Train and deploy your own model	Amazon SageMaker AI
Answer from internal documents without retraining	RAG (Amazon Bedrock Knowledge Bases)
Store embeddings and run similarity search	Vector database (such as OpenSearch Service)
Connect agents to external tools and data in a standard way	MCP (Model Context Protocol)
Make answers consistent with less variation	Lower the temperature
Show the format with a few examples in the prompt	Few-shot
Have the model reason step by step	Chain-of-Thought
Version and reuse prompts	Amazon Bedrock Prompt Management
Match the wording or format of a specific domain	Fine-tuning
Add knowledge from large amounts of unlabeled domain data	Continued pre-training
Transfer a large model's capability to a smaller, cheaper model	Model distillation
Measure summary quality automatically	ROUGE
Measure translation quality automatically	BLEU
Have another LLM grade the answers	LLM-as-a-judge
Reduce missed cases in a classification model	Prioritize recall
Reduce false positives in a classification model	Prioritize precision
Block harmful content, denied topics, and personal data	Amazon Bedrock Guardrails
Detect and explain bias in data or models	SageMaker Clarify
Monitor production model degradation and data drift	SageMaker Model Monitor
Have people review low-confidence predictions	A human-in-the-loop review
Document a model's intended use, evaluation, and risks	SageMaker Model Cards
Keep traffic to Bedrock off the internet	AWS PrivateLink (VPC endpoint)
Cut cost and latency of sending the same long preamble every time	Prompt caching
Secure throughput for steady, high-volume use	Provisioned Throughput

Wrong-answer patterns

Rule out fine-tuning when the model must answer from internal information that changes often.
When asked for the lowest-cost method, rule out pre-training a model from scratch.
Rule out any option that says raising the temperature improves accuracy or consistency.
Rule out any option that says hallucinations can be eliminated completely.
Rule out choosing generative AI for processing that must always give the same rule-based result.
Rule out using a large, hard-to-explain model as is for decisions that require an explanation.
Rule out the reversed pairing of BLEU for summaries and ROUGE for translation.
Rule out evaluating a classification model on highly imbalanced data by accuracy alone.
Rule out building or tuning a model yourself when a managed service would do.
Rule out the mix-up of Model Monitor for bias detection and Clarify for production monitoring.
Rule out handling personal data only by asking for it in the prompt.
Rule out any option that says IAM policies alone are enough against jailbreaks.
Rule out any option that says prompts sent to Bedrock are used to train the foundation models.
Rule out any option that makes managing customer data and access permissions the responsibility of AWS.
Rule out choosing Config to find out who invoked Bedrock.
Rule out any option that describes AI Service Cards as documents customers write about their own models.

Key points by domain

Domain 1 Fundamentals of AI and ML (20%)

ML sits inside AI and deep learning inside ML, generative AI creates content with foundation models, and agentic AI autonomously plans and runs tools toward a goal.
The three types of learning are supervised (classification and regression), unsupervised (clustering and anomaly detection), and reinforcement learning (maximizing reward).
Data types are distinguished as labeled or unlabeled, tabular, time series, image, text, and structured or unstructured.
The four inference methods are real-time, batch, asynchronous, and serverless.
AI is a poor fit when the cost outweighs the benefit and when a fixed outcome rather than a prediction is required.
Choose between traditional ML and foundation models by regulation, the need for explanation, and operational constraints.
The ML cycle repeats data collection, preprocessing and feature engineering, training, evaluation, deployment, and monitoring and retraining.
MLOps covers experiment management, repeatable processes, scalability, managing technical debt, and production monitoring and retraining.
Model metrics are accuracy, precision, recall, and F1 score, and business metrics are cost per user, development cost, customer feedback, and ROI.
Transcribe, Translate, Comprehend, Lex, Polly, Rekognition, and Personalize are managed AI services you call through APIs without training.

Domain 2 Fundamentals of Generative AI (24%)

A token is the unit of text a model handles, chunking splits documents for retrieval, and an embedding represents meaning as a numeric vector.
Generative AI models include LLMs built on transformers, multimodal models that handle text, images, and more, and diffusion models that create images.
The foundation model lifecycle is data selection, model selection, pre-training, fine-tuning, evaluation, deployment, and feedback.
Context engineering is choosing and assembling what to pass to the model from instructions, retrieval results, conversation history, and tool results.
The basics of agents are tool use, memory management, workflow control, multi-agent collaboration, and MCP for connecting to external systems.
Generative AI's strengths are adaptability, responsiveness, conversation, and content creation, and its weaknesses are hallucinations, low interpretability, inaccuracy, and nondeterminism that changes results for the same input.
Choose a model by type, performance requirements, capabilities, constraints, compliance, cost, latency, and complexity.
Measure business value by ROI, efficiency, conversion rate, average revenue per user, and customer lifetime value.
The AWS tools for building generative AI are Bedrock, SageMaker AI, JumpStart, Amazon Quick, Kiro, Strands Agents, and Bedrock AgentCore.
With token-based pricing, longer inputs and outputs raise cost and response time, and cost also depends on whether you use Provisioned Throughput or custom models.

Domain 3 Applications of Foundation Models (28%)

Choose a foundation model by cost, modality, latency, multilingual support, size, customizability, input and output length, and prompt caching support.
Lowering the temperature makes answers consistent and raising it makes them varied, and the maximum output length limits cost and answer length.
RAG passes retrieved internal information to the model as the basis for its answer, and you can build it with Bedrock Knowledge Bases.
Vector databases for storing embeddings include OpenSearch Service, Aurora, Neptune, and RDS for PostgreSQL.
Prompt engineering is the cheapest customization, followed by RAG, fine-tuning, and pre-training in rising cost, and distillation lowers running costs with a smaller model.
An AI agent autonomously works through multiple steps and tool calls toward a goal and is used for customer support and business automation.
Build prompts from context, instructions, input, output format, and a negative prompt stating what to avoid, and techniques include zero-shot, single-shot, few-shot, Chain-of-Thought, and templates.
Prompt risks are exposure, poisoning, hijacking, and jailbreaking, and Bedrock Prompt Management handles prompt versioning.
Training elements are pre-training, fine-tuning (instruction tuning, domain adaptation, transfer learning), continued pre-training, and distillation, data needs quality, governance, size, labels, and representativeness, and RLHF aligns the model with human preferences.
Evaluate with human evaluation, benchmarks, Bedrock model evaluation, ROUGE, BLEU, BERTScore, and LLM-as-a-judge, and in business look at task completion rate, satisfaction, and cost per interaction.

Domain 4 Guidelines for Responsible AI (14%)

The features of responsible AI include bias, fairness, inclusivity, robustness, safety, and veracity.
Bedrock Guardrails blocks harmful content, denied topics, and personal data in both inputs and outputs.
When choosing a model, consider environmental impact and sustainability as well as performance and cost.
The legal risks of generative AI are intellectual property infringement claims, biased outputs, loss of customer trust, risk to end users, and hallucinations.
Datasets should be inclusive, diverse, from curated sources, and balanced.
High bias leads to underfitting and high variance to overfitting, and both cause inaccuracy for particular groups.
Find bias through label quality analysis, human audits, subgroup analysis, and SageMaker Clarify.
A transparent model shows how it works, an explainable model can give the reason for each prediction, and larger deep learning models trade explainability for performance.
Ways to show transparency are SageMaker Model Cards, SageMaker Clarify, Bedrock model evaluation, open source models, disclosure of data and licenses, and AWS AI Service Cards.
Human-centered design values mechanisms for collecting user feedback and showing the basis of AI decisions.

Domain 5 Security, Compliance, and Governance for AI Solutions (14%)

The tools for securing AI are IAM roles and policies, encryption, Macie, PrivateLink, the shared responsibility model, and Bedrock Guardrails.
AgentCore Identity manages agent identities and permissions, and AgentCore Policy defines rules for the actions agents are allowed to take.
Record data provenance with data lineage, data catalogs, and Model Cards, and cite sources in answers.
Secure data handling means assessing quality, using privacy-enhancing technologies, controlling access, and ensuring data integrity.
AI-specific threats are prompt injection, data leakage, and harmful outputs, which call for output filtering and validation and logging of interactions.
Reduce hallucinations with grounding through RAG, output validation, and confidence scores.
For governance, use Config (configuration compliance), Inspector (vulnerabilities), Artifact (reports), CloudTrail (API records), and Trusted Advisor (checks).
Run data governance by defining the lifecycle, logging, residency, monitoring, and retention period.
Governance procedures are policies, review cadence, frameworks such as the Generative AI Security Scoping Matrix, transparency standards, and team training.
The Generative AI Security Scoping Matrix has five scopes from using a consumer app to training your own model, and the more you own, the more responsibility falls on the customer.

Services in one line

Name	In one line
Amazon Bedrock	Use multiple foundation models through a serverless API
Amazon Bedrock Knowledge Bases	Build RAG on your own data
Amazon Bedrock Guardrails	Block harmful content and personal data in inputs and outputs
Amazon Bedrock AgentCore	Platform for running AI agents in production
Amazon SageMaker AI	Where you build, train, and deploy ML models
SageMaker JumpStart	Catalog of pre-trained models and solution templates
SageMaker Clarify	Bias detection and prediction explanations
SageMaker Model Monitor	Monitoring production model quality and data drift
SageMaker Model Cards	Record of a model's intended use, evaluation, and risks
Amazon Comprehend	Extract sentiment and entities from text
Amazon Textract	Read text and tables from documents
Amazon Rekognition	Image and video analysis
Amazon Nova	Family of foundation models provided by AWS
Amazon Quick	AI assistant and BI for business users (including Quick Sight)
Amazon Q	Q Business answers from company data, and Q Developer helps with development
Strands Agents	Open source SDK for building AI agents
AWS Transform	AI agents that migrate and modernize legacy apps and infrastructure
Kiro	AI development environment that builds code from specs

Top 15 traps

RAG is not retraining, and it only passes retrieved information without changing the model weights.
MCP is not a training method but a standard protocol that connects agents to external tools and data.
Lowering the temperature does not eliminate hallucinations.
ROUGE and BLEU are metrics for generated text and are not used to evaluate classification models.
On imbalanced data, high accuracy can still hide many missed cases, so look at recall.
AgentCore Identity does not inspect outputs but manages agent identities and access to tools.
Prompt caching does not reuse answers but skips reprocessing a repeated preamble to cut cost and latency.
Human-in-the-loop means people review predictions, while labeling training data is SageMaker Ground Truth.
Prompt injection slips in instructions from outside, while jailbreaking makes the model drop its safety limits.
Macie finds sensitive data in S3, while Guardrails is what blocks personal data in generative AI outputs.
PrivateLink keeps the network path private and is separate from data encryption (AWS KMS).
Continued pre-training adds knowledge from unlabeled domain data, while fine-tuning adjusts behavior with labeled examples.
Distillation is not a way to build a larger model but a way to transfer a large model's capability to a smaller one.
Token-based pricing is pay-as-you-go, and Provisioned Throughput is for securing steady, high-volume throughput.
Running the Bedrock infrastructure is the responsibility of AWS, while managing prompts, data, and access permissions is the customer's responsibility.

Last-minute checklist

I can state the number of questions, the exam time, the number of scored questions, and the passing score.
I can name the five domains and their weights.
I can name the four question types and which ones score only when fully correct.
I can explain how AI, ML, deep learning, generative AI, and agentic AI relate.
I can name the three types of learning with an example of each.
I can explain when to use regression, classification, and clustering.
I can name the four inference methods and when each fits.
I can name two situations where AI should not be used.
I can explain the differences among accuracy, precision, recall, and F1 score.
I can name three business metrics.
I can explain what tokens, chunking, embeddings, and vectors mean.
I can name four weaknesses of generative AI.
I can explain how answers change when the temperature goes up or down.
I can explain how RAG works and name the AWS feature used to build it.
I can name four AWS services usable as vector databases.
I can rank the customization methods from cheapest to most expensive.
I can name four prompting techniques.
I can name four attacks on prompts.
I can explain the differences among ROUGE, BLEU, BERTScore, and LLM-as-a-judge.
I can name five features of responsible AI.
I can explain how bias and variance relate to underfitting and overfitting.
I can explain the roles of Clarify, Model Monitor, Model Cards, and human-in-the-loop review.
I can explain the difference between transparency and explainability and the trade-off with performance.
I can name three legal risks of generative AI.
I can name both ends of the five scopes in the Generative AI Security Scoping Matrix.
I can name five AWS services used for governance.
I can name three ways to reduce hallucinations.
I can name four building blocks of AI agents.
I can name three customer responsibilities when using Bedrock under the shared responsibility model.