Imagine you’re handed four business problems on a Monday morning. A legal team has thousands of scanned contracts. A call center has a year of recorded phone calls. A social platform needs to screen uploaded photos. A retailer has piles of customer emails to analyze. Each problem involves “AI,” and each has a different best-fit AWS service.
Picking the right service is one of the most practical skills in the AWS Certified AI Practitioner (AIF-C01) conversation, and it’s a common source of mistakes. The names sound like they overlap, and several scenarios deliberately blur the lines. Let’s build a clear decision framework.
For the wider exam context behind AWS AI service selection, see the free AWS AIF-C01 study guide and its domain-by-domain service comparisons.
What These Four Services Have in Common
Amazon Rekognition, Amazon Textract, Amazon Comprehend, and Amazon Transcribe are managed, pretrained AI services. You call them through an API (or the console) and send in your data. AWS operates the underlying models and service infrastructure. You do not need to build or host a model for their standard capabilities; some services also offer customization options.
That is a meaningful contrast with Amazon SageMaker AI, which provides tools to build, train, customize, and deploy machine learning models when a project needs more control or a model tailored to its data. The choice is not simply “managed service versus SageMaker”; it depends on the required capability and level of customization.
The cleanest way to separate the four is by input type:
| Service | Primary input | Core question it answers |
|---|---|---|
| Amazon Rekognition | Images and video | “What’s in this picture or video?” |
| Amazon Textract | Scanned documents and forms | “What does this document say, and how is it structured?” |
| Amazon Comprehend | Text | “What does this text mean?” |
| Amazon Transcribe | Audio | “What was said in this audio?” |
Memorize that table and you’ll answer many service-selection questions. The rest of this guide covers the details and the traps.
Amazon Rekognition: Understanding Images and Video
Amazon Rekognition analyzes images and videos to identify objects, scenes, activities, faces, and potentially inappropriate content.
Common capabilities include:
- Detecting objects and scenes, such as a car, beach, or kitchen.
- Face detection and analysis, including face locations and visual attributes. Face analysis is not a reliable basis for inferring a person’s identity, intent, or character.
- Face comparison and face search, matching detected faces against faces in a collection you manage.
- Content moderation, flagging potentially unsafe or inappropriate imagery.
- Text detection in images, such as words on a street sign or storefront.
- Custom Labels, which can be trained to recognize objects specific to your business, such as your own product line.
- Detection of personal protective equipment (PPE) in images.
- Analysis of supported stored and streaming video workflows.
Practical scenario: A neighborhood marketplace app lets sellers upload photos of items. Before a listing goes live, the app sends each photo to Rekognition’s content moderation capability to screen for potentially inappropriate images, reducing the number of uploads that need manual review.
Face-related capabilities involve privacy, fairness, and governance considerations. Responsible AI is a major theme in AIF-C01, so recognize that face-related use cases require careful policy decisions, appropriate safeguards, and human oversight where warranted.
Amazon Textract: Extracting Data From Documents
Amazon Textract goes beyond simple OCR (optical character recognition). It extracts printed text and handwriting while also identifying document structure and relationships.
Common capabilities include:
- Extracting printed text and handwriting.
- Identifying forms as key-value pairs, such as “Policy Number: 48291.”
- Extracting tables while preserving rows, columns, and cell relationships.
- Using Queries to ask targeted questions about document content, such as “What is the invoice total?”
- Specialized analysis for expenses (invoices and receipts), identity documents, and lending-related documents through dedicated APIs and workflows.
- Detecting the locations of signatures in supported document analysis workflows.
Practical scenario: An insurance company receives thousands of scanned claim forms each week. Staff retype each one into a claims system. By sending the scans to Textract and extracting the policy number, claimant name, and itemized charges table, the company can turn paperwork into structured data for downstream processing.
The easiest way to remember this: Rekognition looks at scenes. Textract reads paperwork. If the input resembles a form, invoice, receipt, ID document, or table, Textract is the likely answer.
Amazon Comprehend: Finding Meaning in Text
Amazon Comprehend is a natural language processing (NLP) service. It analyzes text you already have in digital form and extracts insights from it.
Common capabilities include:
- Sentiment analysis: positive, negative, neutral, or mixed.
- Entity recognition: identifying people, organizations, locations, dates, and other named items.
- Key phrase extraction: finding the main talking points in a passage.
- Language detection: identifying the dominant language of text.
- PII detection and redaction: identifying personally identifiable information, such as names, addresses, and account-related data, in supported languages and workflows.
- Custom classification and custom entity recognition: adapting NLP models to your own categories and terminology.
Healthcare text has a specialized service, Amazon Comprehend Medical, designed to extract medical information such as conditions, medications, and clinical entities from supported clinical text.
Practical scenario: A regional airline collects thousands of post-flight survey comments. Comprehend analyzes sentiment and extracts entities and key phrases, helping the customer experience team spot recurring themes such as baggage claim without reading every comment.
Common exam trap: Comprehend works on text. It does not directly transcribe audio or perform OCR on scanned images. If the information starts as a phone call or a paper form, convert it to text first with Transcribe or Textract.
Amazon Transcribe: Converting Speech to Text
Amazon Transcribe is an automatic speech recognition (ASR) service. It converts spoken audio into written text.
Common capabilities include:
- Batch transcription of recorded audio files and streaming transcription for supported live audio workflows.
- Speaker diarization, distinguishing speakers and labeling their utterances in a transcript.
- Custom vocabulary, helping recognition of industry-specific terms and product names.
- Language identification in supported batch and streaming scenarios.
- PII redaction in supported transcription workflows.
- Call Analytics for customer-service conversations, including conversation characteristics and post-call insights.
- Amazon Transcribe Medical, a specialized offering for healthcare speech-to-text workflows. Check current regional and feature availability before selecting it for a production solution.
Practical scenario: A utility company records customer support calls for quality review. Transcribe converts each call into a searchable transcript with speaker labels, allowing supervisors to review conversations without listening to hours of recordings.
Chaining Services Together
Real solutions often combine these services. The input type changes at each step, which is exactly how you can reason through them.
Scenario: Analyzing customer call sentiment
Audio
→
Transcript text
→
Sentiment and entities
→
Insights
The audio is transcribed first. Comprehend then analyzes the resulting text; it does not analyze the original audio file.
Scenario: Processing scanned mortgage applications
Image or PDF
→
Fields and tables
→
Text insights
→
Structured data
The pattern is consistent: convert the input into the format the next service understands. Audio becomes text through Transcribe. Scanned pages become text and structure through Textract. Text then becomes insight through Comprehend.
A Decision Framework You Can Use
- What is the input?
- Photo or video → Rekognition
- Scanned form, invoice, receipt, or ID → Textract
- Audio or speech → Transcribe
- Plain text → Comprehend
- What’s the desired output?
- Objects, faces, or moderation labels → Rekognition
- Structured fields, key-value pairs, or tables → Textract
- A transcript → Transcribe
- Sentiment, entities, key phrases, or supported PII analysis → Comprehend
- Is it a pipeline? If the input type does not match the service that produces the final insight, chain services.
- Does the scenario require a custom model or deeper control? Consider SageMaker AI for broader model development and customization, or service-specific customization such as Rekognition Custom Labels and Comprehend custom models.
Comparison: Easily Confused Pairs
| Confusing pair | The distinction |
|---|---|
| Rekognition vs. Textract | Both can find text in images. Rekognition detects text in visual scenes such as signs and storefronts. Textract is designed for documents and extracts structure such as forms and tables. |
| Textract vs. Comprehend | Textract extracts text and structure from document images. Comprehend analyzes the meaning and characteristics of text. They often work in sequence. |
| Transcribe vs. Comprehend | Transcribe turns audio into text. Comprehend interprets text. Neither replaces the other. |
| Rekognition vs. Transcribe | Both can be used in video-related workflows, but Rekognition analyzes visual content while Transcribe processes the audio track. |
Related Services Worth Recognizing
AIF-C01 candidates should recognize that AWS offers other purpose-built AI services beyond these four:
- Amazon Translate: translates text between languages.
- Amazon Polly: converts text into lifelike speech (the reverse direction from Transcribe).
- Amazon Lex: builds conversational interfaces such as chatbots.
- Amazon Kendra: provides intelligent enterprise search.
- Amazon Bedrock: provides access to foundation models for generative AI applications.
A helpful pairing is Transcribe (speech to text) and Polly (text to speech). They work in opposite directions, and service-selection questions may test whether you can tell them apart.
Where Do Foundation Models Fit?
Generative AI has changed the landscape. A foundation model available through Amazon Bedrock can summarize text, classify content, and answer questions about documents—tasks that may overlap with some traditional NLP use cases. So which should you use?
Think in terms of fit and efficiency. A purpose-built service such as Comprehend provides specific, well-defined outputs through dedicated APIs, which can be an efficient and predictable choice for standard text analysis. A foundation model offers flexibility for open-ended tasks, such as summarizing a long report in a particular tone or generating a response from context. In exam scenarios, look for clues: a standard, clearly defined analysis task often points toward a purpose-built service, while open-ended generation or flexible language tasks may point toward a foundation model.
For a related foundation-model comparison, read our guide to AI, machine learning, deep learning, and generative AI.
Common Mistakes and Misconceptions
- Using Comprehend on audio or images. Comprehend analyzes text; convert audio or scanned documents first.
- Choosing Rekognition for forms and invoices. Use Textract when document structure matters.
- Assuming Textract is only OCR. It can extract forms, tables, query responses, signatures, and layout information.
- Assuming you must train models. Standard capabilities are available from managed pretrained services; customization is optional and service-specific.
- Overlooking privacy and language support. When data includes faces, voices, or personal information, consider governance, retention, access, and supported languages.
Exam-Focused Insights for AIF-C01
The AIF-C01 exam guide includes AWS AI services within its scope. Service selection is therefore a practical skill worth practicing. We cannot predict specific exam questions, but these habits help:
- Start with the input type. It eliminates several incorrect answers quickly.
- Notice the output wording. “Key-value pairs and tables” signals Textract. “Sentiment” signals Comprehend.
- Watch for pipelines. A scenario that starts with recordings and ends with sentiment analysis usually needs more than one service.
- Beware of neighboring services. Polly sits near Transcribe in the speech domain, while Rekognition and Textract both work with images but solve different problems.
- Read qualifiers carefully. Language support, input format, and whether the content is a document or a natural scene can change the correct choice.
Key Takeaways
- Rekognition analyzes images and video; Textract extracts text and structure from documents; Comprehend analyzes text; Transcribe converts speech to text.
- Match the input type first, then the desired output.
- Many real solutions chain services: Transcribe → Comprehend, or Textract → Comprehend.
- Rekognition’s text detection suits scenes. Textract suits documents.
- These are managed services with pretrained capabilities. Consider SageMaker AI when you need broader model development and customization, and Bedrock when the task calls for foundation-model generation.
- Always check feature-specific language, region, and workflow support before designing a production solution.
Original AIF-C01 Practice Questions
These are original, unofficial educational practice questions written by CloudExamPro. They are not actual AWS exam questions.
Question 1 — Extracting invoice data
An accounting firm receives scanned vendor invoices as PDFs. It needs the invoice number, vendor name, and line-item table extracted into a database automatically. Which AWS service is the best fit?
A. Amazon Rekognition
B. Amazon Textract
C. Amazon Comprehend
D. Amazon Transcribe
Correct answer: B — Amazon Textract.
- B is correct. Textract extracts text, key-value pairs, and tables from scanned documents. Its expense analysis is designed for invoices and receipts.
- A is incorrect. Rekognition analyzes visual content in images and video; it is not designed to preserve document structure such as tables.
- C is incorrect. Comprehend analyzes text that is already available as text. It does not read the scanned PDF image.
- D is incorrect. Transcribe converts speech to text. There is no audio in this scenario.
Question 2 — Sentiment analysis of recorded calls
A telecommunications company stores thousands of recorded support calls. It wants to determine whether customers sounded satisfied or frustrated by analyzing what was said. Which combination of services accomplishes this?
A. Amazon Rekognition, then Amazon Textract
B. Amazon Textract, then Amazon Comprehend
C. Amazon Transcribe, then Amazon Comprehend
D. Amazon Comprehend, then Amazon Transcribe
Correct answer: C — Amazon Transcribe, then Amazon Comprehend.
- C is correct. Transcribe converts audio into text, and Comprehend then performs sentiment analysis on the transcript.
- A is incorrect. Rekognition analyzes images and video, while Textract extracts text from documents. Neither is the right audio-to-text path.
- B is incorrect. Textract expects document images, not audio. The first step would not solve the input problem.
- D is incorrect. The order is reversed. Comprehend cannot take the original audio as input, so Transcribe must come first.
Question 3 — Moderating uploaded images
A photo-sharing community wants to automatically flag user-uploaded images that contain inappropriate content before they are published. Which service should the team use?
A. Amazon Comprehend
B. Amazon Textract
C. Amazon Transcribe
D. Amazon Rekognition
Correct answer: D — Amazon Rekognition.
- D is correct. Rekognition includes image content moderation capabilities for detecting potentially inappropriate or unsafe imagery.
- A is incorrect. Comprehend analyzes text, not image content.
- B is incorrect. Textract extracts text and structure from documents; it does not perform image content moderation.
- C is incorrect. Transcribe processes audio and is unrelated to image moderation.
Question 4 — Detecting and redacting PII in text
A property management company has tenant emails stored as plain text. It needs to find and redact names, phone numbers, and account numbers before sharing the data with an analytics team. Which service is most appropriate, assuming the text language is supported for the required PII feature?
A. Amazon Comprehend
B. Amazon Rekognition
C. Amazon Polly
D. Amazon Transcribe
Correct answer: A — Amazon Comprehend.
- A is correct. Comprehend can detect and redact PII in supported text workflows and languages.
- B is incorrect. Rekognition analyzes images and video, not email text.
- C is incorrect. Polly converts text into speech; it does not detect or redact PII.
- D is incorrect. Transcribe converts audio to text. The data is already text, so transcription is unnecessary.
Question 5 — Reading text in roadside scenes
A city transportation department captures photographs of roadside signs and wants to read the text on each sign. The photographs are outdoor scenes, not scanned forms. Which service is the most appropriate choice?
A. Amazon Textract, because it extracts text
B. Amazon Comprehend, because it analyzes text
C. Amazon Rekognition, because it can detect text in images of real-world scenes
D. Amazon Transcribe, because it converts media into text
Correct answer: C — Amazon Rekognition.
- C is correct. Rekognition text detection is designed for text appearing in visual scenes, such as signs and storefronts.
- A is incorrect. Textract is the better choice for documents, forms, and tables. A roadside scene is not a structured document.
- B is incorrect. Comprehend works on text that has already been extracted; it does not read text from a photograph.
- D is incorrect. Transcribe processes spoken audio, not visual text.
Frequently Asked Questions
1. Can Amazon Rekognition and Amazon Textract both extract text from images?
Yes, but for different situations. Rekognition detects text in real-world scenes such as signs, while Textract is built for documents and can extract forms, tables, and other structure.
2. Do I need machine learning experience to use these services?
Not for their standard pretrained capabilities. They are managed services available through APIs. Some customization options, such as Rekognition Custom Labels or Comprehend custom models, require preparing training data and evaluating the resulting model.
3. Can I use Amazon Comprehend directly on a recorded phone call?
No. Comprehend requires text input. Use Amazon Transcribe first to convert supported audio into text, then send the transcript to Comprehend.
4. When should I use Amazon Bedrock instead of Amazon Comprehend?
Use Comprehend for standard, well-defined text analysis such as sentiment or entity extraction. Consider Bedrock when you need open-ended generation, flexible summarization, or language tasks that do not fit a fixed output format.
5. Which service is the opposite of Amazon Transcribe?
Amazon Polly. Transcribe converts speech to text, while Polly converts text to speech.
Official AWS References
- AWS Certified AI Practitioner (AIF-C01) Exam Guide
- Amazon Rekognition Developer Guide
- Amazon Textract Developer Guide
- Amazon Comprehend Developer Guide
- Amazon Comprehend supported languages
- Amazon Transcribe Developer Guide
- Amazon Transcribe feature matrix, including medical and call analytics options
- Amazon Bedrock User Guide
- Amazon SageMaker AI Developer Guide
Practice choosing the right AWS AI service
Service selection is easier to learn by doing. Try our free AIF-C01 diagnostic exam to see how you handle scenario-based questions like these. If you’d like to go deeper, the AIF-C01 Premium practice exams include original questions with detailed explanations for every option. No dumps. No shortcuts.
Keep reading: AI vs. Machine Learning vs. Deep Learning vs. Generative AI: Key Differences Explained.