Chapter 3

AWS AI Services in Practice

AWS Certified AI Practitioner (AIF-C01) Study GuideBased on the AWS AIF-C01 exam objectives

If the last chapter was the map, this one is the vehicle. AWS offers dozens of AI and machine learning services, and the AIF-C01 exam expects you to know what each one does, when to use it, and how they differ from each other. Don't panic — most of them follow a simple pattern: give them input, they give you intelligence back. This chapter walks through the ones that matter most, one story at a time.

Why AWS Has So Many AI Services

Imagine you run a bakery. You need flour, sugar, eggs, and an oven. You could buy a pre-made cake (someone else did the work), or you could buy the raw ingredients and bake it yourself. AWS's AI services work the same way. Some services are pre-trained cakes — you call them with your data and get answers immediately, no machine learning expertise required. Others are ingredients and ovens — full machine learning platforms where you build, train, and deploy your own custom models.

The official service categories you'll see on the exam are:

• Machine learning services (SageMaker — build your own models)
• AI-powered services (Rekognition, Polly, Transcribe, Comprehend, Translate, Lex, Personalize, Textract, Amazon Q — pre-trained, ready to use)
• Generative AI services (Amazon Bedrock — build with foundation models)

Let's meet them all.

Amazon SageMaker: The Machine Learning Workshop

If AWS's AI services were a kitchen, SageMaker would be the entire professional kitchen — ovens, mixers, prep stations, and all. Amazon SageMaker is a fully managed service for building, training, and deploying machine learning models. "Fully managed" means AWS handles the servers, the infrastructure, the scaling — you focus on the machine learning work itself.

Here's the typical journey in SageMaker:

1. Prepare your data — SageMaker provides notebooks (Jupyter) where you explore and clean your data.
2. Choose an algorithm — SageMaker includes built-in algorithms (for common tasks like classification and forecasting) or you can bring your own code.
3. Train the model — SageMaker spins up the compute you need, runs the training, and shuts it down when done. You only pay for what you use.
4. Deploy the model — SageMaker hosts your trained model behind an endpoint so applications can send it data and get predictions back in real time.

A real use case: A bank wants to predict which loan applicants are likely to default. Their data science team uses SageMaker notebooks to explore ten years of loan data, trains a classification model on SageMaker's managed training infrastructure, and deploys it as an endpoint. Now, when a new application arrives, the bank's system calls the endpoint and gets a risk score in milliseconds.

When to pick SageMaker: When you need a custom model trained on your own data for your specific problem — and you have (or are building) machine learning expertise. If a pre-trained service already does what you need, don't use SageMaker. Building custom models is expensive and time-consuming; the exam loves scenarios where the "right" answer is the simpler pre-trained service instead.

SageMaker has several sub-features the exam may mention:

• SageMaker Studio — the web-based development environment (IDE) for ML work.
• SageMaker Autopilot — automatically builds and trains models for you, picking the best algorithm. Think of it as "SageMaker for people who don't want to hand-tune everything."
• SageMaker JumpStart — pre-built models and solutions you can deploy with a few clicks.

Let's talk about when each makes sense, because this is exactly the kind of judgment the exam tests. SageMaker Studio is for data scientists who want full control — they'll write their own training code, experiment with hyperparameters, and iterate. Autopilot is for teams that have tabular data and a prediction goal but limited ML expertise: upload a CSV, tell it what column to predict, and Autopilot tries dozens of algorithm and preprocessing combinations, then shows you a leaderboard. JumpStart is for the fastest path — one-click deployment of popular open-source models (like language models or image classifiers) that you can then fine-tune.

There's also a cost dimension worth understanding. Training a large model on SageMaker can run for hours or days on expensive GPU instances. That's why SageMaker offers Spot Instances for training (discounted spare capacity — fine for experimentation, risky for deadline-driven jobs) and SageMaker Savings Plans. The exam won't ask you to calculate costs, but it may present a scenario where the "best" answer considers cost-effectiveness — and "use the pre-trained service instead of training a custom model" is often the cost-effective choice.

Finally, understand what "fully managed" really buys you. Before services like SageMaker, a data science team needed to provision EC2 instances, install frameworks (TensorFlow, PyTorch), configure networking, manage storage for datasets, and build their own deployment infrastructure. SageMaker absorbs all of that. The team in our bank example never touches a server — they work in notebooks, click "train," and get an endpoint. That operational simplicity is the product.

Amazon Rekognition: Eyes for Your Applications

Amazon Rekognition performs image and video analysis. Show it a photo, and it can tell you what's in it: objects, people, text, scenes, activities — and even faces, including facial attributes and whether two faces match.

A real use case: A social media company receives millions of uploaded photos per day. They use Rekognition's content moderation to automatically detect and flag inappropriate images before any human ever sees them. Another classic: a smart doorbell company uses Rekognition to detect packages on doorsteps and alert homeowners.

What Rekognition can detect:

• Objects and scenes — "car," "dog," "beach," "birthday party"
• Faces — detect faces, compare faces (is this the same person?), search a collection of faces
• Text in images — license plates, street signs, product labels
• Celebrities — recognize well-known people (with appropriate use restrictions)
• Unsafe content — flag explicit or suggestive imagery

Rekognition works on both still images and video. For video, it can track people across frames, detect activities ("person entering a room"), and even summarize segments. A retail chain uses Rekognition Video to count foot traffic and analyze how shoppers move through stores — information that used to require expensive manual studies.

It's worth understanding how Rekognition charges and scales, because the exam occasionally touches on operational thinking: you pay per image analyzed or per minute of video processed. There's no infrastructure to manage and no model to train. Upload an image via the API, get back a structured JSON response with labels and confidence scores. That confidence score matters — Rekognition doesn't just say "dog," it says "dog, 98.2% confident." Your application decides what threshold counts as a match.

Amazon Polly: Giving Applications a Voice

Amazon Polly turns text into lifelike speech. Feed it written text, and it returns natural-sounding audio — in dozens of languages and voices, including neural voices that sound remarkably human.

A real use case: A news website wants an "listen to this article" button. Instead of hiring voice actors to record thousands of articles, they send article text to Polly and embed the returned audio. A navigation app uses Polly to speak turn-by-turn directions in the driver's preferred language.

When to pick Polly: Any time an application needs to speak — accessibility features, voice assistants, audiobooks, announcements, IVR phone systems. If the direction is text → speech, it's Polly.

Polly's voices come in two flavors: standard voices and neural voices. Neural voices use deep learning to produce remarkably natural speech — with appropriate intonation, pauses, and emphasis — while standard voices sound more robotic. There's also a conversational speaking style for voices designed to sound like natural dialogue rather than narration. And for brands that want a unique identity, Polly offers Brand Voice: a custom voice built exclusively for one company (think of a famous brand's unmistakable spokesperson voice, available on demand).

Polly supports Speech Synthesis Markup Language (SSML), which lets developers fine-tune pronunciation — emphasizing words, adding pauses, spelling out acronyms, or whispering. A banking app might use SSML to make Polly read "$1,234.56" as "one thousand two hundred thirty-four dollars and fifty-six cents" rather than stumbling over the symbols.

Amazon Transcribe: Ears for Your Applications

Amazon Transcribe converts speech to text. Give it an audio file (or a live audio stream), and it returns a transcript — with timestamps, speaker identification, and even custom vocabulary for industry terms.

A real use case: A call center records thousands of customer calls daily. Transcribe converts them all to text so the company can search for complaint patterns ("how many callers mentioned 'refund' this week?"). A video platform uses Transcribe to auto-generate subtitles for uploaded videos. Doctors use Transcribe Medical — a specialized version — to turn dictated notes into clinical documentation.

Key capabilities:

• Batch transcription — process recorded audio files
• Streaming transcription — real-time captions for live audio
• Speaker diarization — label who said what ("Speaker 1," "Speaker 2")
• Custom vocabulary — teach it product names, jargon, or unusual terms
• Transcribe Medical — tuned for medical terminology and HIPAA-eligible workflows

Think about why each of these exists. Batch transcription handles the backlog — the thousands of hours of recorded calls sitting in storage. Streaming transcription handles the live moment — a webinar happening right now needs captions now, not tomorrow. Speaker diarization solves a real readability problem: a raw transcript of a two-person conversation without speaker labels is nearly useless. Custom vocabulary solves the accuracy problem every specialized industry hits — a generic speech model will mangle drug names, legal terms, or product codenames unless you teach it.

Transcribe also handles the messy realities of audio: background noise, different accents, multiple languages. It can automatically detect the spoken language, redact PII from transcripts (useful for compliance — a call center doesn't want credit card numbers sitting in searchable text), and produce output in formats like SRT and VTT that video players understand natively.

Amazon Comprehend: Understanding What Text Means

Amazon Comprehend performs natural language processing (NLP) — it doesn't just read text, it understands it. Give it a pile of documents, reviews, or social media posts, and Comprehend can tell you the sentiment, extract key phrases, identify entities (people, places, organizations), detect the language, and even find personally identifiable information (PII).

A real use case: An e-commerce company has 200,000 product reviews. Reading them all is impossible. Comprehend analyzes every review and reports: overall sentiment is 72% positive, the most-mentioned complaint is "shipping speed," and here's every review that mentions a competitor. A law firm uses Comprehend's PII detection to find and redact Social Security numbers before sharing documents.

What Comprehend detects:

• Sentiment — positive, negative, neutral, or mixed
• Entities — names, dates, locations, organizations, quantities
• Key phrases — the most important phrases in a document
• Language — which of 100+ languages the text is written in
• PII — Social Security numbers, credit card numbers, addresses
• Syntax — parts of speech (nouns, verbs, adjectives)

Comprehend Medical is the healthcare-tuned variant — it extracts medical conditions, medications, dosages, and test results from clinical text.

There's an important subtlety worth understanding: Comprehend comes in two modes. The standard API calls analyze text on demand — you send text, you get results. But Comprehend also supports custom models: you can train Comprehend on your own labeled data to recognize entity types specific to your business. A financial firm might train a custom entity recognizer to find "account numbers" and "transaction IDs" in support tickets. This is a middle ground between a fully pre-trained service and building your own model in SageMaker — easier than SageMaker, more tailored than the base API.

Amazon Translate: Breaking Language Barriers

Amazon Translate performs neural machine translation — fast, high-quality translation between dozens of languages. It's the engine behind "translate this page" features.

A real use case: A travel booking site operates in 15 countries. When a hotel in Tokyo writes its listing in Japanese, Translate automatically renders it in English, Spanish, French, and eleven other languages. A customer support team uses Translate to convert incoming tickets to English so agents can respond, then translates the reply back.

When to pick Translate: Any time content needs to move between languages — websites, apps, documents, chat. It's real-time and scales to millions of characters.

Amazon Lex: Building Conversational Chatbots

Amazon Lex is the service for building conversational interfaces — chatbots and voice assistants. It's the same technology that powers Alexa: automatic speech recognition (understanding spoken words) plus natural language understanding (figuring out what the user means).

A real use case: A bank's website has a chatbot. A customer types "I lost my card." Lex recognizes the intent (report lost card), asks follow-up questions ("Was it a debit or credit card?"), and either completes the task or hands off to a human agent with full context. An airline uses Lex for a voice bot: "Book me a flight to Chicago on Friday" becomes a real reservation.

Key concepts:

• Intents — what the user wants to accomplish ("book a flight," "check balance")
• Slots — the details needed to fulfill the intent (destination, date)
• Utterances — the different ways users might phrase the same request

Let's make this concrete. Suppose you're building a pizza-ordering bot with Lex. You define an intent called OrderPizza. It needs slots: size, crust, and toppings. You provide sample utterances: "I want a large pepperoni pizza," "Order me a medium thin-crust with mushrooms," "Get a small cheese pizza." Lex's natural language understanding learns from these examples, so when a real customer types "Can I get a large pizza with extra cheese and pepperoni?", Lex identifies the intent (OrderPizza) and fills the slots (size=large, toppings=[extra cheese, pepperoni]). If a slot is missing — the customer said "pizza" but no size — Lex prompts: "What size would you like?" That back-and-forth slot-filling is what makes Lex feel conversational rather than robotic.

Lex bots can be deployed to websites, mobile apps, messaging platforms (like Facebook Messenger or Slack), and even telephony through Amazon Connect. And because Lex handles both text and voice input, the same bot can power a website chat widget and a phone system.

Amazon Personalize: Recommendations Like the Big Players

Amazon Personalize delivers personalized recommendations — the "customers who bought this also bought" experience. It uses the same machine learning technology that powers Amazon.com's recommendation engine, but trained on your data.

A real use case: A streaming service wants "recommended for you" rows. They feed Personalize their catalog and user viewing history; Personalize returns tailored recommendations for each user. An online retailer uses it for personalized product rankings and "frequently bought together" suggestions.

When to pick Personalize: When you want recommendations but don't have a machine learning team. You provide user interaction data (clicks, purchases, views); Personalize handles the ML. No ML expertise required.

Amazon Textract: Reading Documents Like a Human

Amazon Textract extracts text, handwriting, and data from scanned documents — going beyond simple optical character recognition (OCR). It doesn't just read characters; it understands document structure: tables, forms, key-value pairs.

A real use case: An insurance company receives thousands of claim forms as PDFs. Textract extracts every field — name, policy number, claim amount, dates — including data inside tables, and even reads handwritten entries. A bank uses Textract to pull income figures from scanned pay stubs during loan processing.

What makes Textract more than basic OCR:

• Tables — extracts rows and columns with their relationships intact
• Forms — identifies key-value pairs ("Name: John Smith")
• Handwriting — reads handwritten text, not just print
• Queries — ask natural-language questions about a document ("What is the total amount?")
• Expense analysis — specialized for receipts and invoices

Consider what "understanding structure" actually means in practice. Basic OCR on a table gives you a soup of disconnected words — "Item," "Quantity," "Price," "Widget," "3," "$12.99" — with no indication of which price belongs to which item. Textract preserves the geometry: it knows that "$12.99" sits in the same row as "Widget" and under the "Price" column. For forms, it pairs each label with its value even when they're laid out in complex multi-column designs. The Queries feature goes further — instead of parsing the whole document yourself, you ask "What is the invoice date?" and Textract finds it, like asking a careful assistant.

Amazon Q: The Generative AI Assistant for Work

Amazon Q is a generative AI-powered assistant designed for work — it can answer questions, summarize content, generate drafts, and take actions, grounded in your company's data.

There are two main flavors the exam may reference:

• Amazon Q Business — a chat assistant employees can ask about company policies, documents, and data. It connects to your enterprise data sources (like SharePoint, S3, or Confluence) and answers with citations, respecting existing access permissions.
• Amazon Q Developer — a coding assistant inside your IDE that suggests code, explains code, debugs, and even performs upgrades.

A real use case: A company's HR team spends hours answering the same policy questions. They deploy Q Business connected to the employee handbook and internal wiki. Now employees ask "How many vacation days do I get in my second year?" and get an instant, cited answer. A development team uses Q Developer to generate unit tests and explain legacy code.

When to pick Q: When you need a ready-to-use AI assistant for employees — not a model to build with, but a product to deploy. Contrast with Bedrock (where you build the generative AI application) — Q is the application.

One detail the exam may probe: Q Business respects your existing access controls. If an employee doesn't have permission to view the finance folder, Q won't surface finance documents in its answers to that employee. This matters enormously in enterprise settings — an assistant that leaks restricted information would be worse than no assistant at all. Q also provides citations, so users can verify where an answer came from rather than blindly trusting it.

Q Developer, meanwhile, lives where developers work — inside IDEs like VS Code and JetBrains. It can explain unfamiliar code ("what does this function do?"), generate code from comments, find security vulnerabilities, and even help upgrade Java versions across a codebase. Think of it as a pair programmer that has read most of the internet's public code.

How These Services Work Together

In the real world, these services rarely work alone. The exam sometimes presents scenarios where the right answer involves combining services, so it's worth seeing how they fit together.

The customer support pipeline. A company receives support calls as audio recordings. First, Transcribe converts the audio to text. Then Comprehend analyzes the transcripts for sentiment — flagging angry customers for priority callback. If a caller speaks Spanish, Translate converts the transcript to English for the agent. The whole pipeline runs automatically, turning thousands of hours of audio into actionable insight.

The voice assistant. A smart speaker needs to hear, understand, and respond. Transcribe (or Lex's built-in speech recognition) converts the user's voice to text. Lex figures out the intent and decides what to do. Polly speaks the response aloud. Three services, one seamless conversation.

The document intelligence workflow. A mortgage company processes applications with pay stubs, bank statements, and tax forms. Textract extracts the structured data from each document. Comprehend checks the free-text fields for PII that needs redaction. The cleaned, structured data flows into the company's database — no manual data entry.

The global storefront. Product listings arrive in English. Translate renders them into twenty languages. Polly generates audio versions for accessibility. Personalize decides which products each visitor sees first. Rekognition auto-tags product photos so search works ("show me red dresses").

Notice the pattern: each service does one thing well, and applications chain them together. You don't need to memorize every possible combination — just understand each service's job, and the combinations become obvious.

The Build-vs-Buy Decision

Step back and look at the whole chapter through a decision-maker's eyes, because the AIF-C01 exam rewards this kind of thinking. For almost any AI problem, you face a spectrum of choices:

1. Use a pre-trained AI service (Rekognition, Comprehend, Polly...) — fastest, cheapest, no ML expertise needed. The trade-off: you're limited to what the service does. If Rekognition doesn't detect the specific defect your factory cares about, you're stuck.

1. Customize a pre-trained service — some services (like Comprehend custom entities or Rekognition Custom Labels) let you train on your own examples while keeping the managed simplicity. Middle ground: more tailored than option 1, easier than option 3.

1. Build with SageMaker — full control, custom algorithms, your own data. The trade-off: you need ML expertise, and training costs time and money.

1. Build with Bedrock — use foundation models via API for generative AI tasks. Less ML expertise than SageMaker, more flexibility than pre-trained services for language and generation tasks.

The exam loves scenarios that test whether you can place a requirement on this spectrum. "A startup with no ML engineers needs image moderation next week" → option 1 (Rekognition). "A manufacturer needs to detect a defect type no pre-trained model recognizes" → option 2 or 3. "A team wants a custom chatbot grounded in their documentation" → Bedrock with retrieval, or Q Business if they want it ready-made.

When in doubt, prefer the simpler option. AWS designed these services so that most business problems don't require building models from scratch. The exam reflects that philosophy.

Putting It All Together: The Input Test

Here's the fastest way to keep these services straight on exam day — ask what goes IN:

Summary

• SageMaker is the full ML platform: build, train, and deploy your own custom models. Use it when pre-trained services aren't enough.
• Rekognition analyzes images and video (objects, faces, text in images, moderation).
• Polly converts text to lifelike speech.
• Transcribe converts speech to text (captions, transcripts, call analytics).
• Comprehend understands written text (sentiment, entities, key phrases, PII).
• Translate translates text between languages.
• Lex builds conversational chatbots and voice assistants (intents and slots).
• Personalize generates product/content recommendations from user behavior.
• Textract extracts text and structured data from documents (tables, forms, handwriting).
• Q is the ready-to-use generative AI assistant for employees (Business) and developers (Developer).
• The exam's favorite trick is making you choose between similar services — the input type and the direction of conversion almost always reveal the answer.

---

End-of-Chapter Test

1. A company wants to automatically generate subtitles for thousands of uploaded training videos. Which AWS service should they use?

A. Amazon Polly B. Amazon Transcribe C. Amazon Comprehend D. Amazon Rekognition

2. A retail website wants to add a "customers who viewed this also viewed" feature but has no machine learning team. Which service is the best fit?

A. Amazon SageMaker B. Amazon Personalize C. Amazon Lex D. Amazon Comprehend

3. A bank needs to build a custom fraud-detection model trained on ten years of its own transaction data using a proprietary algorithm. Which service should it use?

A. Amazon Bedrock B. Amazon Rekognition C. Amazon SageMaker D. Amazon Q Business

4. An insurance company receives claim forms as scanned PDFs and needs to extract names, policy numbers, and amounts from tables within the forms. Which service should it use?

A. Amazon Rekognition B. Amazon Textract C. Amazon Comprehend D. Amazon Translate

5. A mobile app needs to read news articles aloud to users in natural-sounding voices. Which service converts the article text to speech?

A. Amazon Transcribe B. Amazon Polly C. Amazon Lex D. Amazon Connect

6. A business wants to analyze 50,000 customer reviews to determine overall sentiment and identify the most common complaints. Which service should it use?

A. Amazon Rekognition B. Amazon Transcribe C. Amazon Comprehend D. Amazon Personalize

7. A developer is building a website chatbot that can understand when a customer wants to "track my order" versus "return an item" and ask follow-up questions. Which service provides this conversational capability?

A. Amazon Q Business B. Amazon Lex C. Amazon Polly D. Amazon SageMaker

8. A travel company needs to translate hotel descriptions from Japanese into twelve other languages automatically. Which service should it use?

A. Amazon Comprehend B. Amazon Translate C. Amazon Transcribe D. Amazon Textract

9. A company's employees waste time searching wikis and policy documents for answers. The company wants a ready-to-use AI assistant that answers questions with citations from internal documents. Which service should it deploy?

A. Amazon Bedrock B. Amazon SageMaker C. Amazon Q Business D. Amazon Lex

10. A social media platform needs to automatically detect and flag inappropriate images in user uploads before they are published. Which service performs this image analysis?

A. Amazon Textract B. Amazon Rekognition C. Amazon Comprehend D. Amazon Personalize

---

Answer Key

1. B — Transcribe converts speech to text, which is exactly what subtitles/captions require. Polly goes the opposite direction (text to speech).

2. B — Personalize delivers recommendations without requiring ML expertise — you supply user interaction data and it handles the modeling. SageMaker would require building a custom model from scratch.

3. C — SageMaker is the platform for training custom models on your own data with your own algorithms. Bedrock works with existing foundation models rather than training traditional ML models from scratch.

4. B — Textract extracts structured data (tables, forms, key-value pairs) from documents. Rekognition analyzes photos/video for objects and faces, not document structure.

5. B — Polly turns text into lifelike speech. Transcribe does the reverse.

6. C — Comprehend performs sentiment analysis and key-phrase extraction on text. It is designed for exactly this "understand large volumes of written feedback" scenario.

7. B — Lex builds conversational interfaces with intents (what the user wants) and slots (details needed). Q Business is a deployed assistant, not a chatbot-building platform.

8. B — Translate performs neural machine translation between languages. Comprehend understands text but does not translate it.

9. C — Q Business is the ready-to-deploy AI assistant for employees, grounded in company data sources. Bedrock is for developers building their own generative AI applications.

10. B — Rekognition analyzes images and video, including content moderation to detect inappropriate imagery. Textract is for documents, not photo content analysis.