AWS AI services and sub-services for AIF-C01: what each one does and how to use it

The AIF-C01 exam does not ask you to build anything. It asks whether you can pick the right AWS service for a problem and explain what it does. That sounds easy until you realize AWS has a dozen AI services, and the two big ones - Amazon Bedrock and Amazon SageMaker AI - each contain a dozen sub-services of their own. Most people study the headline names and get caught out by a scenario that hinges on Knowledge Bases vs. fine-tuning, or Ground Truth vs. Data Wrangler.

This is the deep version: every service and every sub-service in scope for AIF-C01, with what it is, why it exists, and how you would actually use it. It is long on purpose - treat it as a reference you scan by section, not a story you read start to finish. If you want the shorter "which service maps to which exam domain" map first, read the companion post Every AWS service on the AIF-C01 exam, then come back here for the sub-service detail.

The mental model: five ways to use AI on AWS

Before the service list, hold this framing in your head. Every AI question on the exam is really asking "which of these five layers fits?"

• Managed AI APIs (Comprehend, Rekognition, Textract, and so on) - pre-trained, single-purpose. You send data, you get a result. No model choice, no training.
• Amazon Bedrock - managed access to foundation models plus the tooling to build generative AI apps (RAG, agents, guardrails). You choose a model and compose behavior; you do not manage servers.
• Amazon SageMaker AI - the full build-your-own-model platform. You prepare data, train, tune, deploy, and monitor custom models.
• Amazon Q - finished AI assistant products. You do not choose a model or wire up a vector store; AWS already did.
• PartyRock - a no-code playground for learning and prototyping generative AI apps.

The exam constantly tests the boundaries between these. Keep the layer in mind as you read each service below.

Amazon Bedrock

What it is. A fully managed service that gives you one API to call foundation models from many providers, plus a set of built-in tools for building generative AI applications on top of them. No infrastructure, no model hosting - you send prompts, you get responses.

Why it exists. Generative AI apps need more than a raw model: they need private-data grounding, safety controls, orchestration, and evaluation. Bedrock bundles all of that behind serverless APIs so you build the app, not the plumbing.

How to use it. Pick a model, test it in a playground, then add the sub-services you need (Knowledge Bases for RAG, Guardrails for safety, Agents for actions) and call it all from your app through the API or SDK.

Foundation models

The catalog of models you can call. You are not locked to one vendor - selecting the right model for cost, latency, and capability is a core exam skill.

• Amazon Titan - Amazon's first-party models: Titan Text, Titan Text Embeddings (for RAG vector search), Titan Image Generator, and Titan Multimodal Embeddings.
• Amazon Nova - Amazon's newer model family (covered in its own section below).
• Anthropic Claude - strong general-purpose reasoning and long-context models.
• Meta Llama, Mistral, Cohere, AI21 Labs (Jamba), Stability AI (image generation), and others.

Know that a "foundation model" is a large pre-trained model you adapt through prompting, RAG, or fine-tuning - not something you train from scratch.

Bedrock playgrounds

What/why. Console chat, text, and image consoles for trying models and prompts with no code. This is where you compare models and tune inference parameters (temperature, top-p, max tokens) before wiring anything up.

How to use. Open the playground, pick a model, iterate on a prompt, then carry the winning model and settings into your app.

Knowledge Bases

What/why. The managed retrieval-augmented generation (RAG) capability. It connects a foundation model to your private documents so answers are grounded in your data - without fine-tuning.

How to use. Point it at a data source (for example S3), and it chunks the documents, generates embeddings, stores them in a vector store, and retrieves relevant passages to inject into the prompt at query time. It tracks citations automatically. The exam's canonical answer to "ground a model in private data without retraining" is Knowledge Bases (RAG).

Agents

What/why. Give a foundation model the ability to take actions - call APIs, run multi-step workflows, query a Knowledge Base - instead of just returning text.

How to use. Define "action groups" (the tools and APIs the model may call, often backed by Lambda), and Agents runs the reason-act-observe loop. If a scenario needs the assistant to look something up and then do something (check an order, update a record), the answer is Agents, not plain inference.

Guardrails

What/why. A configurable policy layer between the user and the model that enforces safety and responsible-AI requirements.

How to use. Turn on the protections you need: content filters (hate, violence, and so on), denied topics you define, word filters, sensitive-information (PII) redaction on inputs and outputs, and contextual grounding checks to reduce hallucination. It is declarative - no custom code. Guardrails apply across models, Agents, Knowledge Bases, and Flows.

Flows

What/why. A visual, low-code builder for chaining Bedrock pieces into one named workflow.

How to use. Drag prompts, Knowledge Bases, Agents, Guardrails, Lambda functions, and conditional logic into a flow, then invoke the whole thing with one API call. Use it when you want orchestration without writing the glue code.

Prompt Management

What/why. A versioned catalog for prompts so teams reuse and compare them instead of pasting strings into code.

How to use. Store prompts as versioned resources, test variants, and reference them from apps and Flows. Pairs with Model Evaluation to measure which version performs better.

Model Evaluation

What/why. Built-in evaluation of model and RAG quality, so model choice is measured rather than guessed.

How to use. Run automatic evaluations (metrics like accuracy, robustness, toxicity) or bring human reviewers, and compare models or prompt versions on your own data.

Custom models (fine-tuning and continued pre-training)

What/why. When prompting and RAG are not enough, adapt a model to your domain.

How to use. Fine-tuning teaches a model from labeled examples; continued pre-training adapts it to a large body of unlabeled domain text. The result is a private custom model you serve through Bedrock. The exam wants you to know fine-tuning changes the model, whereas RAG leaves the model alone and supplies context at query time.

Provisioned Throughput

What/why. A pricing and capacity mode. On-demand pricing charges per token with no commitment; Provisioned Throughput reserves dedicated capacity for steady, high-volume, or custom-model workloads.

How to use. Choose Provisioned Throughput when you need guaranteed throughput or to serve a custom model; otherwise on-demand is the default.

Other Bedrock features to recognize

• Batch inference - run large jobs offline at lower cost.
• Cross-Region inference - route requests across Regions for capacity and resilience.
• Model invocation logging - send prompt and response logs to CloudWatch or S3 for audit.
• Bedrock Data Automation - extract and transform unstructured content (documents, images, audio, video) into structured output.
• Bedrock Marketplace - access specialized and emerging models beyond the core catalog.

Amazon Bedrock AgentCore

What it is. A newer set of services for running AI agents in production, regardless of which framework or model built them. Bedrock Agents helps you build an agent; AgentCore helps you operate one at scale.

Why it exists. Prototype agents break in production - they lack memory, secure tool access, identity, and observability. AgentCore supplies those as managed building blocks you adopt independently.

How to use. Pick the components you need:

• Runtime - serverless hosting for the agent or tool code, with session isolation.
• Memory - persistent context across sessions so the agent remembers.
• Gateway - exposes your APIs and Lambda functions to the agent as callable tools.
• Identity - gives each agent a workload identity and manages its access to external services (integrates with Cognito, Okta, Entra ID).
• Code Interpreter - a sandbox for the agent to run code.
• Browser - headless web browsing for the agent.
• Observability - tracing and monitoring of agent behavior.

For AIF-C01, recognize AgentCore as the "take an agent to production" layer; you do not need to build one.

Amazon Nova

What it is. Amazon's own family of foundation models, served through Bedrock.

Why it exists. To offer strong capability at low, predictable price and latency across text, image, and video.

How to use. Choose the tier that fits the job:

• Nova Micro - text-only, lowest cost and latency.
• Nova Lite - low-cost multimodal (text, image, video input).
• Nova Pro - balanced multimodal model for most tasks.
• Nova Premier - the most capable, and a good "teacher" for distilling smaller custom models.
• Nova Canvas - image generation.
• Nova Reel - video generation.
• Nova Sonic - speech-to-speech; Nova Act - agentic browser actions.

Amazon SageMaker AI

What it is. The end-to-end platform for building, training, deploying, and monitoring your own machine learning models. This is the "build a custom model" layer - the opposite end from the managed APIs.

Why it exists. Some problems need a model trained on your data, with full control of the lifecycle. SageMaker AI provides every stage as a managed capability so teams do not stitch together their own ML infrastructure.

How to use. For AIF-C01 you use SageMaker conceptually: know what each sub-service does and when it applies. You are not expected to write training code.

SageMaker Studio

The web-based IDE and control center - notebooks, experiments, and access to every other SageMaker capability in one place.

SageMaker JumpStart

A catalog of pre-trained foundation models and task-specific models with one-click deploy and fine-tuning notebooks. Think of it as deploying a ready-made model into your own account and endpoint - distinct from calling a hosted API through Bedrock.

SageMaker Canvas

No-code ML for analysts. Upload a dataset, pick a target, and Canvas builds a model and generates predictions. It can also tap Bedrock foundation models for generative tasks - all without writing code.

Data Wrangler

Visual data preparation. Import from many sources, apply hundreds of built-in transforms, analyze, and export the prepared data into a training pipeline. This is the "clean and shape the data" step.

Feature Store

A central repository for the features (model inputs) your teams engineer, with an online store for low-latency inference and an offline store for training. Prevents every team from recomputing the same features.

Ground Truth (and Ground Truth Plus)

Data labeling. Ground Truth manages human and automated labeling workflows to build training datasets; Ground Truth Plus is the fully managed, AWS-run version with an expert workforce. Labeled data is what supervised training needs.

Clarify

Bias detection and explainability. It measures bias in data and in model predictions across groups, and computes feature attributions (which inputs drove a prediction, using SHAP). This is the named answer for "detect bias" and "explain the model" in responsible-AI scenarios.

Model Monitor

Continuous monitoring of models in production. It detects data drift, model-quality drift, bias drift, and feature-attribution drift, and raises alerts - so you catch a model degrading before users do.

Autopilot

Automated machine learning (AutoML). Point it at a dataset and it explores algorithms, trains, and tunes candidate models automatically, then surfaces the best one. The no-heavy-lifting path to a trained model.

Pipelines

MLOps orchestration. Define the repeatable workflow - prepare, train, evaluate, register, deploy - as a pipeline so model delivery is automated and reproducible.

Model Registry and Model Cards

Model Registry versions trained models and manages approval for deployment. Model Cards document a model's purpose, data, performance, and limitations for governance. Both support responsible, auditable ML.

Inference options

How you serve a trained model, matched to the traffic pattern:

• Real-time endpoints - persistent, low-latency, always-on serving.
• Serverless inference - scales to zero for intermittent traffic; no capacity to manage.
• Asynchronous inference - queues large or long-running requests.
• Batch transform - scores a whole dataset offline, no endpoint needed.

Knowing which inference option fits a scenario (steady traffic vs. spiky vs. bulk) is a recurring exam pattern.

The managed AI services (task-specific APIs)

Pre-trained services that solve one kind of problem with no training or model management. The exam tests picking the right one for a use case.

Amazon Comprehend

What/why/how. Natural language processing over text: sentiment, entities, key phrases, language detection, topic modeling, and PII detection, plus custom classification and custom entity recognition. Send text, get structured results. Answer for "analyze customer feedback" or "find PII in text."

Amazon Textract

What/why/how. Document understanding beyond basic OCR - it extracts text, key-value pairs, tables, and fields from PDFs and scans, with specialized analysis for invoices, receipts, and identity documents. Answer for "extract structured data from forms."

Amazon Rekognition

What/why/how. Computer vision for images and video: object and scene detection, facial analysis and comparison, content moderation, text in images, and Custom Labels for your own categories. Answer for "analyze images without an ML team."

Amazon Transcribe

What/why/how. Automatic speech-to-text with speaker identification, custom vocabulary, custom language models, PII redaction, real-time streaming, and a medical variant. Often paired with Comprehend: transcribe a call, then analyze sentiment.

Amazon Polly

What/why/how. Text-to-speech with neural and generative voices in many languages, controlled with SSML for pronunciation and pacing. Answer for "convert text to natural speech."

Amazon Translate

What/why/how. Neural machine translation across many languages, with custom terminology for domain phrasing, real-time and batch. Answer for "translate content without training a model."

Amazon Lex

What/why/how. Build conversational bots with automatic speech recognition and natural-language understanding: define intents and slots, fulfill with Lambda, speak back through Polly. It is the structured, intent-based chatbot framework - contrast it with Bedrock Agents for open-ended, tool-using assistants.

Amazon Personalize

What/why/how. Managed real-time recommendations and personalization built from the same technology behind Amazon.com. Feed it interaction data; get recommendations, personalized ranking, and user segments - no recommender model to build.

Amazon Q

What it is. AWS's finished AI-assistant products. The key idea for the exam: Q is a product, not a toolkit. You do not select a model, manage prompts, or build a vector store - AWS did that.

• Amazon Q Developer - a coding and AWS assistant in your IDE and the console: code generation, debugging, reference tracking, AWS resource questions, and cost or CLI help (formerly CodeWhisperer).
• Amazon Q Business - an enterprise assistant that answers questions grounded in your company data (for example SharePoint, Confluence, Salesforce), with permission-aware responses.
• Embedded flavors you may see named: Q in QuickSight (business intelligence), Q in Connect (contact center), and Q Apps.

How to use / exam angle. Choose Q when the requirement is "use AI without building anything"; choose Bedrock when you need to customize and build your own app.

PartyRock

What it is. A no-code, web-based playground (powered by Bedrock) for building generative AI apps by describing them in plain language. It runs outside the AWS Console and does not require an AWS account.

Why it exists. To let anyone learn generative AI and prompt engineering hands-on, and to prototype app ideas fast.

How to use. You compose an app from widgets - user input, static text, and AI-powered text generation, image generation, or chatbot widgets. You chain them by referencing one widget's output inside another's prompt with the @ sign (dynamic variables), so outputs flow from widget to widget. Great for building intuition before you touch Bedrock proper.

Supporting services in scope

Not AI-specific, but tested in security, governance, cost, and data scenarios. Know what each is for; you do not need deep configuration detail.

• Security, identity, and compliance: IAM (least-privilege access, for example restricting which Bedrock models a role can invoke), KMS (encryption keys for data, model artifacts, and logs), Macie (ML-based sensitive-data and PII discovery in S3), Amazon Inspector (vulnerability scanning), Secrets Manager (managed secrets), AWS Artifact (on-demand compliance reports).
• Management and governance: CloudTrail (API audit trail), CloudWatch (metrics, logs, alarms; Bedrock invocation logs and endpoint metrics), AWS Config (resource configuration and compliance), Trusted Advisor (best-practice checks), AWS Well-Architected Tool (review workloads against the framework).
• Storage: Amazon S3 (the data lake for training data, documents, and artifacts), S3 Glacier (archival).
• Compute: Amazon EC2 (including GPU and accelerator instance families; AWS Trainium for training and AWS Inferentia for inference), AWS Lambda (serverless glue for pipelines and agent actions).
• Containers: Amazon ECS and Amazon EKS for containerized workloads.
• Databases and vector stores: Amazon OpenSearch Service (k-NN vector search for RAG), Amazon Aurora and Amazon RDS (relational, with pgvector for embeddings), Amazon DynamoDB (key-value), Amazon Neptune (graph), Amazon DocumentDB, Amazon ElastiCache.
• Analytics and data prep: AWS Glue and Glue DataBrew (ETL and prep), Amazon EMR (big-data processing), AWS Lake Formation (data-lake governance), Amazon Redshift (warehouse), AWS Data Exchange (third-party data), Amazon Quick / QuickSight (business intelligence).
• Networking: Amazon VPC (including PrivateLink patterns to keep traffic off the internet), Amazon CloudFront (content delivery).
• Cost: AWS Budgets and AWS Cost Explorer (track and forecast spend - relevant because generative AI cost management is on the exam).

New in-scope additions to recognize

AWS refreshed the in-scope list, so a few names are newer than most study material:

• Amazon Bedrock AgentCore and Amazon Nova (covered above).
• Kiro - an agentic, spec-driven development environment (IDE and CLI) that turns prompts into specs, code, and docs.
• Strands Agents - an open-source SDK for building and orchestrating AI agents on AWS.
• AWS Transform - agentic workbench for cloud migration and application modernization.
• The exam guide now writes Amazon SageMaker AI (not just "SageMaker"). Same platform, updated name.

Explicitly out of scope: skip these

AWS lists these as out of scope for AIF-C01, so do not spend study time on them:

• Machine learning: AWS DeepComposer, AWS HealthImaging, AWS HealthOmics, Amazon Monitron, AWS Panorama.
• Security and networking: Amazon GuardDuty, AWS Security Hub, Amazon Cognito, AWS WAF, AWS Shield, Amazon Route 53.
• Management: AWS Control Tower, AWS Organizations.
• Plus most Internet of Things, Media, and Migration and Transfer services.

If a practice question hinges on one of these, it is almost certainly a distractor.

A note on naming and scope drift

A handful of services show up in older AIF-C01 study material (and in some question banks) but are not on the current official in-scope list. You may still see them referenced; here is what each is for, briefly:

• Amazon Q - the AI assistant products above; ubiquitous in practice even though it is not in the formal in-scope service table.
• Amazon Kendra - ML-powered enterprise search (retrieval only, versus Knowledge Bases which retrieves and generates).
• Amazon Augmented AI (A2I) - adds human review for low-confidence AI predictions.
• AWS Audit Manager - automates collection of compliance evidence.
• AWS Fargate, Amazon EventBridge, Amazon API Gateway - serverless compute, eventing, and API plumbing that AI apps are often built on.

Understand what they do at a sentence level and move on.

The 30-second decision guide

Given a scenario, match it to a layer:

• Call a foundation model by API -> Bedrock
• Ground a model in private data, no retraining -> Bedrock Knowledge Bases (RAG)
• Let the model take actions -> Bedrock Agents
• Add safety and PII controls -> Bedrock Guardrails
• Run agents in production (memory, tools, identity) -> Bedrock AgentCore
• Use AI without building anything -> Amazon Q
• Learn or prototype with no code -> PartyRock
• Train or tune a custom model -> SageMaker AI
• Deploy a ready-made model yourself -> SageMaker JumpStart
• Label data -> Ground Truth; prepare data -> Data Wrangler; store features -> Feature Store
• Detect bias or explain a model -> SageMaker Clarify
• Monitor a model in production -> SageMaker Model Monitor
• Analyze text / images / documents / speech / translation / chatbots / recommendations -> the matching managed API
• Secure, encrypt, scan, audit, or monitor -> IAM / KMS / Macie / CloudTrail / CloudWatch

Related certifications

This guide maps to several AWS certifications - it appears under Related Study Guides on each of their hubs:

• AWS Certified AI Practitioner (AIF-C01) - the primary target; every service here is in its scope.
• AWS Certified Machine Learning Engineer Associate (MLA-C01) - goes deeper on SageMaker AI, data prep, and MLOps.
• AWS Certified Generative AI Developer Professional (AIP-C01) - builds on Bedrock, Agents, Knowledge Bases, and Guardrails.
• AWS Certified Cloud Practitioner (CLF-C02) - foundational; touches these AI services lightly.
• AWS Certified Data Engineer Associate (DEA-C01) - the data-prep and analytics services (Glue, Data Wrangler, Feature Store) overlap directly.

How to study this

Do not memorize every feature. The exam rewards pattern recognition: read a scenario, identify the layer, name the service. Run practice questions with the 30-second decision guide open, and each time you miss one, come back to that service's section here and read the what-why-how until the boundary is clear. Get the Bedrock sub-services and the SageMaker sub-services straight - that is where the exam separates people who skimmed from people who understood.

Source: AWS Certified AI Practitioner (AIF-C01) exam guide (in-scope and out-of-scope service lists) and AWS service documentation, as of September 2026.