AIF-C01 · CHEAT SHEET

AWS Certified AI Practitioner

Key concepts grouped by official exam weight.

Download as PDF (free, no sign-up)

1. Fundamentals of AI and ML

20–20%

Explain basic AI concepts and terminologies

An ML model is trained on historical data to learn patterns and then performs inference, making predictions on new data. A good fit generalizes: an overfit model is accurate on training data but fails on new data, while an underfit model fails on both. Bias produces unequal outcomes across groups and harms fairness; an LLM is a very large deep learning model pre-trained on huge volumes of text.
AI is the umbrella term; ML is one of its branches and learns patterns from data without explicit instructions, and deep learning is a type of ML that uses neural networks with many layers. Generative AI uses deep learning models to create new content instead of only classifying it. Agentic AI combines those models with autonomy to plan and take actions through tools toward a goal.
Real-time inference serves interactive, low-latency requests, and serverless inference removes instance management and suits intermittent traffic that tolerates cold starts. Asynchronous inference queues requests with large payloads or long processing and can scale to zero instances. Batch inference, such as batch transform or Amazon Bedrock batch inference, processes large volumes of data without a persistent endpoint.
Structured data fits into tables with a fixed schema; unstructured data, such as audio, images or free text, follows no predefined rules. Labeled data pairs each input with its correct answer and is required for supervised learning, while unlabeled data is used to discover patterns and to pre-train LLMs. Time-series data adds a timestamp to each value so that future values can be forecast.
Supervised learning trains on inputs paired with labeled outputs (classification and regression); unsupervised learning finds patterns in unlabeled data (clustering, anomaly detection); semi-supervised learning mixes a little labeled data with a lot of unlabeled data. Reinforcement learning learns by trial and error to maximize a reward, and RLHF applies it with human preferences to align LLMs.

Identify practical use cases for AI

AI/ML adds value by automating routine tasks such as extracting and classifying data from documents, by helping people decide better and faster by surfacing patterns in large data volumes, and by scaling services such as 24/7 customer support. It fits best with problems whose patterns are complex or keep changing, such as fraud or personalized recommendations.
ML is probabilistic: it provides predictions, not guaranteed exact results, so when the logic is known and a specific, reproducible outcome is required, a rules-based system is the better fit. Before investing, compare ML with a simpler baseline and estimate its total cost: data preparation and labeling, training, hosting, and maintenance.
Regression predicts continuous values, such as a house price, and classification assigns each item to predefined classes: binary with two classes, such as spam or not spam, and multiclass with several. Without labels, clustering discovers groups of similar items, anomaly detection flags rare observations, and topic modeling organizes documents into topics not known in advance.
Computer vision analyzes images and video, such as defect inspection; NLP extracts sentiment or entities from text, and speech recognition turns audio into text. Recommendations come from user interactions, fraud is detected as an anomaly, forecasting uses time series, knowledge bases answer from proprietary data, and agentic AI acts on its own toward a goal.
Amazon Transcribe converts speech to text, Amazon Polly turns text into lifelike speech, and Amazon Translate translates text between languages. Amazon Comprehend extracts entities, key phrases and sentiment and builds custom classifiers with AutoML; Amazon Lex builds voice and text bots that understand intent. Amazon SageMaker AI is the managed service to build, train and deploy your own models.

Describe the AI/ML development lifecycle

The ML lifecycle runs through business goal, problem framing, data processing (collection, preprocessing and feature engineering), model development (training, tuning and evaluation), deployment and monitoring, with feedback loops. Data is split into training, validation and test sets, and the same preprocessing applied to training data is applied to inference requests.
Building a foundation model from scratch is expensive, needs enormous resources and can take months, so the usual path is to start from a pre-trained FM. SageMaker JumpStart offers publicly available and proprietary third-party FMs, each under its own license that the customer must review; Bedrock Marketplace extends the catalog, and Custom Model Import brings models customized elsewhere into Bedrock.
A model can be consumed as a managed API service such as Amazon Bedrock (serverless, per-token billing, with Provisioned Throughput for a fixed level of throughput) or self-hosted. SageMaker AI managed endpoints give more control than Bedrock with less overhead than EC2; full self-hosting gives the most control but the owner must maintain endpoints and model performance.
Amazon Bedrock gives serverless access to foundation models; SageMaker AI covers building, training and deploying your own models (Feature Store, Pipelines, Model Registry). Kiro is a spec-driven agentic development environment for developers, and Amazon Quick gives business users AI-powered dashboards, research and automations with no infrastructure to manage.
MLOps applies DevOps principles to ML: it automates training and deployment in CI/CD pipelines, versions code and data so results are reproducible and auditable, and limits technical debt by starting from a simple baseline model. In production, teams monitor for drift and retrain when metrics fall below defined KPIs rather than on a fixed schedule.

2. Fundamentals of GenAI

24–24%

Explain the basic concepts of generative AI (GenAI)

A token is the unit of text a model processes; chunking splits documents so they fit the context window, and embeddings turn each chunk into a vector that represents its meaning. Transformer-based LLMs rely on self-attention, multimodal models understand text, images or video, and diffusion models generate images by removing noise step by step.
Generative AI creates new content: text, code, images, video and speech. Common uses are summarizing documents, translating, suggesting code, drafting marketing copy, assistants and conversational customer service agents, knowledge search, personalized recommendations, and synthetic data to test rare cases.
An FM is pre-trained with self-supervised learning on large volumes of unlabeled data, carefully filtered to avoid bias. Then a model is chosen by modality, context window, latency and cost, customized with prompts, RAG or fine-tuning on labeled data, evaluated, deployed, and continuously improved with metrics and user feedback.
With on-demand pricing you pay for input and output tokens, so longer prompts and responses make each request more expensive and also add latency. To save, shorten prompts, cap response length, use prompt caching for repeated context and batch inference when no immediate answer is needed; Provisioned Throughput is billed hourly instead.
Context engineering designs systems that dynamically assemble the information an LLM needs: system prompt, user query, user profile, memory, tools or MCP servers and documents retrieved with RAG. It goes beyond prompt engineering; too little context causes hallucinations, while an overstuffed context window increases latency and cost.

Understand the capabilities and limitations of GenAI for solving business problems

Generative AI responds naturally to human conversation, which enables assistants that resolve requests on first contact, and it generates new content: text, prototypes, code, summaries or synthetic labeled data. Foundation models are adaptable: one model performs very different tasks depending on the prompt, whereas traditional ML is trained for a single task.
Generative AI can hallucinate or give inaccurate but plausible information, and it reflects biases in its data. Its models are black boxes that are hard to interpret, and output is nondeterministic: each token is sampled from a distribution, so the same request can vary (a low temperature makes it steadier). RAG and the contextual grounding checks in Amazon Bedrock Guardrails help mitigate this.
When choosing a model, weigh modality, size, accuracy, training data, price, context window, latency, infrastructure compatibility and the provider's data usage policies. AWS recommends setting the minimum quality bar, testing smaller models first and moving up only if needed, comparing candidates with Amazon Bedrock evaluations, and routing simple requests to cheaper models.
Every technical metric should trace back to a business outcome, for example through the OGSM framework. ROI is tracked continuously on a dashboard combining costs (cost per interaction, infrastructure) and value (hours saved, revenue, CSAT). By use case: conversion rate for marketing, customer lifetime value for retention, and time to market or code review time for efficiency.

Describe AWS infrastructure and technologies for building GenAI applications

Amazon Bedrock gives managed API access to foundation models; SageMaker AI builds, trains and deploys your own models, and JumpStart offers pretrained models to fine-tune and deploy. Strands Agents is an open-source SDK for coding agents, and Amazon Bedrock AgentCore runs and operates them at scale. Kiro is a spec-driven agentic IDE, and Amazon Quick is an AI workspace for business users.
AWS managed generative AI services lower the barrier to entry: Amazon Bedrock gives API access to models from many providers with no infrastructure, and Amazon Quick brings agents and automation to non-technical users. They remove the undifferentiated heavy lifting of hosting models, speed up time to market, and let you use your own data with Knowledge Bases without training models.
Under the shared responsibility model, AWS protects the infrastructure and the customer controls its data, IAM permissions and configuration. Amazon Bedrock encrypts data, supports AWS KMS keys and AWS PrivateLink, does not share data with model providers, and has compliance reports in AWS Artifact. Amazon Bedrock Guardrails adds safety: denied topics, content filters and PII filters.
In Amazon Bedrock, on-demand inference is paid per input and output token; Provisioned Throughput gives fixed hourly-billed capacity, cheaper with longer commitments. The Priority, Standard and Flex tiers trade speed for price. Cross-Region inference absorbs bursts (geographic profiles for data residency), and distillation creates smaller, cheaper models.

3. Applications of Foundation Models

28–28%

Describe design considerations for applications that use foundation models (FMs)

Choosing an FM starts from requirements: modality (text, image, audio), languages, context window, latency and per-token cost. AWS advises testing smaller models first and moving up in size only if quality falls short, and routing simple requests to cheaper models. Prompt caching cuts latency and input cost when a long context is reused across requests.
Inference parameters change the response without changing the model. Low temperature gives more deterministic answers and high temperature more varied ones; Top K and Top P limit how many candidates are considered for the next token. Maximum tokens, stop sequences and penalties control output length and repetition, which also affects cost.
RAG retrieves relevant information from company data sources and adds it to the prompt before the FM answers, without retraining the model. It suits private or frequently changing data and lets answers cite their sources. Amazon Bedrock Knowledge Bases provides managed RAG: ingestion, chunking, embeddings, vector search and prompt augmentation.
Embeddings live in vector stores so they can be searched by similarity. Amazon OpenSearch Service offers high throughput and full-text plus vector search; Amazon Aurora and Amazon RDS for PostgreSQL with pgvector combine SQL and vectors; Amazon Neptune Analytics joins graphs and vectors for GraphRAG; Amazon S3 Vectors lowers cost for large, rarely queried sets.
AWS recommends starting with prompt engineering (in-context learning), then RAG, and moving to fine-tuning only when needed; training an FM from scratch costs the most. Examples in the prompt are paid as tokens on every request, while fine-tuning costs training and storage. Distillation produces a smaller model that is cheaper to serve.

Choose effective prompt engineering techniques

A prompt combines an instruction (what to do), context (guiding information), input data, examples and output indicators (format or length). Through the API, the model does not recall earlier requests unless they are included in the prompt. For image generation, a negative prompt lists what to exclude, because these models do not understand negation words such as "no" or "without".
Zero-shot gives no examples and relies on pre-training; single-shot gives one example and few-shot several, which the model learns in context without changing its weights. Examples should resemble real data. Chain-of-thought asks for step-by-step reasoning on complex tasks, and prompt templates fix a reusable structure with placeholders for the variable parts.
Prompt engineering improves the relevance and accuracy of responses without retraining the model and gives more control over what users can ask for. Best practices: clear, concise instructions, an explicit output format, a default answer when the model does not know, iterating and testing against a hold-out set, and brief guardrails in the template, because over-engineered ones cost more and reduce accuracy.
Prompt injection tries to override the developer's instructions (goal hijacking), a jailbreak tries to bypass the model's native safety, and prompt leakage tries to expose the system prompt, which enables further attacks. Poisoning manipulates training data or a knowledge base. Mitigations include validating input, controlling who can edit the data and using Amazon Bedrock Guardrails.
Amazon Bedrock Prompt Management stores reusable prompts with their model and inference configuration. Variables in double curly braces adapt one prompt to each case, and variants let you compare alternatives. Saving updates the draft; for production you create a version, a fixed snapshot that the application references and that lets you switch configurations or go back to an earlier one.

Describe the training and fine-tuning process for FMs

An FM is pre-trained on huge volumes of unlabeled data through self-supervised learning. It can then be adapted: continued pre-training uses unlabeled domain text, fine-tuning uses labeled input-output examples for a task, and distillation trains a smaller, cheaper student model on the responses of a larger teacher model.
Transfer learning reuses a pre-trained model and fine-tunes it for a related task with less data. Instruction-based fine-tuning uses labeled prompt-response pairs, while domain adaptation and continued pre-training use unlabeled domain text to learn jargon and keep knowledge current without starting from scratch.
Fine-tuning data must be clean, representative and appropriately sized: remove duplicates, toxic content and personal information, and gather labels from several experts to reduce bias. RLHF trains a reward model on human preferences, and reinforcement fine-tuning uses reward functions when labeling is expensive.

Describe methods to evaluate FM performance

Amazon Bedrock Evaluations offers programmatic evaluations with built-in or custom datasets, evaluations with human workers for subjective criteria, LLM-as-a-judge evaluations and RAG evaluations. Public benchmarks complement, but do not replace, a golden dataset of prompts and expected responses from your own use case.
ROUGE measures n-gram overlap between a generated summary and a reference summary, and BLEU compares machine translations with human translations. BERTScore compares embeddings and captures paraphrases that ROUGE penalizes. When there are no reference answers, an LLM as a judge scores criteria such as helpfulness or tone and explains each score.
To know whether an FM meets business objectives, measure a baseline before rollout and link every technical metric to a measurable business goal, for example with the OGSM framework. Productivity and feature adoption reflect value and user engagement, and ROI is tracked continuously because costs and usage change.
An FM application is evaluated in layers, not only by its final answer. For RAG, Amazon Bedrock Evaluations measures retrieval (context relevance and coverage) and generation (faithfulness, correctness, citations). For agents, Amazon Bedrock AgentCore Evaluations scores sessions, traces and tool calls, and for batch workflows throughput and cost per task come first.
Infrastructure metrics show whether an AI application is running, not whether it delivers value. To align it with the business, measure task completion and escalation rates, user satisfaction (NPS, CSAT, feedback) and cost per interaction or per task. These metrics are compared with a baseline, watched with anomaly alerts and reviewed when objectives change.

4. Guidelines for Responsible AI

14–14%

Explain the development of AI systems that are responsible

AWS defines responsible AI through several dimensions: fairness (impacts on different groups), explainability, privacy and security, safety against harmful output and misuse, controllability, veracity and robustness (correct outputs even with unexpected or adversarial inputs), governance, and transparency (letting people make informed choices about using the system).
Amazon Bedrock Guardrails evaluates inputs and responses with configurable safeguards: content filters, denied topics, word filters, sensitive information filters (block or mask PII), and contextual grounding checks and Automated Reasoning checks against hallucinations. It applies during model inference or through the ApplyGuardrail API without invoking any foundation model.
To select a model sustainably, AWS recommends considering managed AI services and pre-trained models first, starting with simple algorithms as a baseline and not deciding on accuracy alone. For generative AI, RAG instead of retraining large models, parameter-efficient fine-tuning (LoRA), distillation into smaller models and quantization all reduce resource use.
Legal risks of generative AI include intellectual property claims when outputs reproduce protected content, biased outputs, hallucinations and loss of customer trust. The AWS Responsible AI Policy notes that outputs are probabilistic and that customers must evaluate them and apply human oversight to consequential decisions; customers must also review the licenses of the models they use.
A good dataset represents the real users and contexts of the use case, with enough examples from each demographic group to assess fairness; gaps are fixed with more data or stratified sampling. It should also cover the relevant variations and be documented in a datasheet with intended uses, sources, collection method and known biases.

Recognize the importance of transparent and explainable models

Simple models such as linear models or decision trees are interpretable: you can see how each input affects the output. Complex models such as deep neural networks or foundation models are not interpretable on their own and need post hoc tools such as SHAP or LIME, confidence scores or source citations. Explainability lets people understand outputs; transparency lets them make informed choices about using the system.
Amazon SageMaker Model Cards document the intended use, risk, training and evaluation of your own models in one place, with immutable versions and PDF export. AWS AI Service Cards describe the intended uses, limitations and design choices of AWS services, but they do not replace your own testing, which you can run with Amazon Bedrock Evaluations.
Interpretable models such as decision trees are easy to understand but may lose accuracy, so compare their performance with complex models. More transparency can conflict with privacy or security: AWS recommends meeting the minimum for each criterion, documenting trade-off decisions and measuring explainability with user studies.
Human-centered design lets users understand, challenge and give feedback on AI outputs, tailors explanations to each group's expertise and gives human reviewers contributing factors, confidence scores and similar cases. For Amazon Bedrock agents, the trace shows the steps and reasoning behind each response.

5. Security, Compliance, and Governance for AI Solutions

14–14%

Explain methods to secure AI systems

AWS protects the infrastructure while the customer controls its data, least-privilege IAM permissions and AWS KMS encryption. Amazon Macie discovers PII in Amazon S3, AWS PrivateLink connects a VPC to Amazon Bedrock without crossing the internet, and Guardrails filters content. AgentCore Identity stores agent credentials and Policy in AgentCore authorizes agent tool calls at the gateway.
Documenting the origin of data and models makes results auditable and reproducible. SageMaker ML Lineage Tracking records which data, jobs and pipelines produced each model; the AWS Glue Data Catalog indexes data location and schema; and SageMaker Model Cards keeps intended uses, risk and evaluation in immutable versions. Bedrock Knowledge Bases adds citations to responses.
Secure data engineering starts by identifying and classifying sensitive data and minimizing it with masking or tokenization. AWS Glue Data Quality measures quality with rules, AWS Lake Formation restricts access by column, row and cell, AWS Clean Rooms applies differential privacy, and S3 Object Lock protects integrity by preventing data from being overwritten or deleted.
Prompt injection is an application risk owned by the customer: validate input and use the Guardrails prompt attack filter, which also filters toxicity and masks PII in responses. GuardDuty detects threats, Inspector finds vulnerabilities, and Bedrock model invocation logging, disabled by default, keeps prompts and responses for audits.
RAG reduces hallucinations by grounding responses in data retrieved for each query. The Guardrails contextual grounding check scores each response for grounding and relevance and blocks those below the threshold. Automated Reasoning checks validate responses with formal logic against rules and return findings without blocking.

Recognize governance and compliance regulations for AI systems

AWS Artifact provides AWS compliance reports (ISO, SOC, PCI) for auditors. AWS Config evaluates resource configurations with rules and keeps their history; AWS CloudTrail records who took each action and when; Amazon Inspector continually scans for vulnerabilities; and AWS Trusted Advisor recommends cost, performance and security improvements.
Data governance defines how long data is kept, where it is processed and how it is watched. S3 Lifecycle moves and deletes objects by rule, CloudWatch Logs keeps logs indefinitely unless a retention period is set and can mask sensitive data, geographic inference profiles keep processing within one geography, and CloudWatch observes generative AI workloads.
The Generative AI Security Scoping Matrix classifies use cases into five scopes, from consumer app to self-trained model, and guides the controls for each. The AI policy should set periodic evaluations against ground truth data, including for fine-tuned models. AWS AI Service Cards provide transparency, and each customer must train its own staff.

Practice this certification