AWS has a pre-trained service for speech, one for documents, one for images and several for text, plus Amazon Bedrock and Amazon SageMaker AI. A scenario gives you a requirement and four plausible names. The quickest way through is to read the requirement as two questions. What form is the data in? And what exactly must come out the other end?
Start with what goes in
A service that cannot accept the input fails before it does anything, so the input’s form narrows the choice first.
- Audio goes to Amazon Transcribe, an automatic speech recognition service. It transcribes files in common formats such as WAV and MP3, or a live stream.
- Scanned documents go to Amazon Textract, which extracts text and data from scans, handwriting included.
- Images and video go to Amazon Rekognition, which identifies objects, people, text, scenes and activities, and detects inappropriate content.
- Text goes to Amazon Comprehend for insights, or Amazon Translate for another language.
Order follows from this. Spanish call recordings that must become English text need Transcribe first and Translate second, because Translate takes text, not audio. Amazon Polly runs the other way, turning text into lifelike speech, so it never belongs where the input is audio.
Then read what must come out
| The requirement asks for | Service |
|---|---|
| Speech as text, each speaker labelled | Transcribe, with speaker partitioning |
| Text in another language | Translate |
| Form fields and tables from scans | Textract |
| Words and lines in a photo, sign or number plate | Rekognition text detection |
| Sentiment, entities and personal data in text | Comprehend |
| Text read aloud | Polly |
| A chatbot or voice bot that recognises intent | Amazon Lex |
| Recommendations from users’ activity | Amazon Personalize |
| Search across enterprise content | Amazon Kendra (closed to new customers) |
| An assistant for employees over company data | Amazon Q Business (closed to new customers; Amazon Quick is its successor) |
| Generated content inside your own application | Amazon Bedrock |
| A prediction learned from your own labelled records | Amazon SageMaker AI |
The details in the requirement matter. Speaker partitioning in Transcribe labels up to 30 speakers, spk_0 to spk_29, in batch and streaming jobs, but only if you turn it on. Comprehend’s standard insights cover entities, key phrases, personal data (PII), language, sentiment, targeted sentiment and syntax.
The near misses
Most wrong answers in this area are a real service doing a neighbouring job.
- Rekognition for forms. Text detection is built for words and lines: it returns them and where they sit in the image, up to 100 words per image, for JPEG or PNG images with text within 90 degrees of horizontal. Textract is the service that “goes beyond simple optical character recognition (OCR) to also identify the contents of fields in forms and information stored in tables”.
- Kendra for analysis. Kendra is intelligent search: it helps people find content, though it has been closed to new customers since July 30, 2026. Measuring sentiment in that content is Comprehend’s job.
- A custom Comprehend classifier for standard insights. Comprehend’s insights use a pre-trained model, “so that there is no need for you to provide training data”. Custom classification and custom entity recognition build models from data you already have, which only makes sense when your categories are your own.
- Translate for new copy. Translate translates text that exists. It writes nothing.
- Q Business for an application. Q Business answers questions, summarises and generates content for employees, from the data in enterprise systems; AWS has closed it to new customers and points to Amazon Quick. An application that must generate copy calls a model through the Amazon Bedrock API.
- Personalize for a score. Personalize takes an activity stream and an inventory of items and produces individualised recommendations. A yes-or-no prediction for each booking is a different problem.
Where building your own begins
The pre-trained services are the low-effort answer whenever they produce what is asked. AWS’s overview describes Kendra, for example, as a fully managed service with “no ML models to build, train, or deploy”. You move to SageMaker AI, which lets you “build, train, and deploy ML models for any use case”, when the answer has to be learned from your own records, such as a cancellation score from years of bookings with known outcomes.
If auditors also want the features behind each score, that is feature attribution. SageMaker Clarify computes it with the SHAP library, as per-instance SHAP values, but AWS now says Clarify “is no longer open to new customers”, and points new work to the SHAP library itself, the same engine.
What the exam guides ask
The AIF-C01 guide asks you to “explain basic AI concepts and terminologies” (Task 1.1), “identify practical use cases for AI” (Task 1.2) and “describe AWS infrastructure and technologies for building GenAI applications” (Task 2.3). The MLA-C02 guide asks you to “compare and select appropriate ML models, generative AI (GenAI) models, algorithms, and solution templates to meet business needs” (Skill 2.1.3), “evaluate tradeoffs between custom solutions, managed services, pre-trained models, and FMs” (Skill 2.1.4) and “apply AWS AI services to solve specific business problems (for example, Amazon Textract, Amazon Rekognition, Amazon Comprehend, Amazon Transcribe)” (Skill 2.1.8). The AIP-C01 guide asks you to “create data processing workflows to handle complex data types including text, image, audio, and tabular data” for FM consumption (Skill 1.3.2) and to “enhance input data quality to improve FM response quality and consistency” (Skill 1.3.4).
Scenario questions on this topic often add “with the least operational overhead” and offer four kinds of wrong answer: a service that cannot take the input, a service that runs in the opposite direction, a neighbour that returns the wrong structure, and a custom model where a pre-trained feature already does the job.
Play it
The free stage below is a contact centre’s to-do list. Larkspur Travel has no machine learning team and five requests waiting: Spanish and English call recordings that must become searchable English text with speakers labelled, scanned refund forms with handwriting, chat transcripts for the CRM, a cancellation score the auditors can explain, and campaign copy in five languages this week. You drop services into each row’s slots and run the rows. Each record stops at the first service that cannot take it, and the debrief explains why not the others.