MLA-C01 asks "do you even need to build a model here?" more often than you might expect. Task 2.1 of the exam guide expects you to know how to solve business problems with AI services such as Amazon Translate, Amazon Transcribe, Amazon Rekognition, and Amazon Bedrock, and in our question bank about 14% of questions use managed AI services like Comprehend, Transcribe, Polly, Textract, Rekognition, and Personalize as the answer or as distractors. Knowing the service names is not enough for these questions, because they make you pick a specific feature: "the brand name is mispronounced," "we need a copy with the personal data removed."
This post sets the criteria for deciding whether a managed AI service is enough, then covers the features that decide the answer for each data type, human review with A2I, and the services that have reached end of support or closed to new customers. Choosing between SageMaker built-in algorithms and AI services was covered first in Choosing Modeling Approaches.
---
Decide First: Is a Managed AI Service Enough?
When you cook, you do not have to make every sauce from scratch. Some dishes are fine with a store-bought sauce, while a restaurant's signature dish deserves its own. ML works the same way. Features everyone needs, such as translation, speech recognition, and text extraction from documents, are a matter of calling a model AWS has already trained through an API. Predictions that set your company apart are the ones you train yourself in SageMaker AI.
| Clue in the question | First choice to consider | |------|------| | No ML specialists, and a fast launch is needed | Managed AI service | | A standard capability such as translation, speech recognition, or text extraction | Managed AI service | | Recognizing labels only your company has (part defects, contract types) | Custom features of AI services (Rekognition Custom Labels, Comprehend custom classification) | | You must control the algorithm and the training process | SageMaker AI built-in algorithms or your own model | | Generative tasks such as writing or summarizing | Amazon Bedrock |
Exam tip: When you see phrases like "minimize operational overhead," "a team with no ML experience," or "as quickly as possible," a managed AI service is probably the answer. When the question says "tune the model architecture directly" or "custom training code," lean toward SageMaker AI.
The figure below is a map that starts from the data type and leads to a service. The rest of this post follows the same order.
!Map of AWS AI services chosen by data type: text, speech, translation, documents, images, recommendations, and conversation
---
Understanding Text: Amazon Comprehend and Comprehend Medical
Comprehend is a natural language processing service that reads text and tells you what it is about, what sentiment it carries, and who or what appears in it. Its features split into those that use pre-trained models and custom features that you train on your own data.
| Feature | What it does | Clue in the question | |------|------|------| | Sentiment analysis | Classifies the whole document as positive, negative, neutral, or mixed | "whether customer reviews are positive" | | Targeted sentiment | Sentiment toward specific entities in the document (such as product features) | "only the opinions about the battery" | | Entities and key phrases | Extracts names of people, places, organizations, and key expressions | "company names mentioned in the documents" | | PII detection and redaction | Locates personal data such as addresses and account numbers, and produces a redacted copy | "analyze after removing personal data" | | Custom classification | Sorts documents into your company's own categories | "classify contracts into company-specific types" | | Custom entity recognition | Extracts terms only your company uses | "extract internal part numbers" |
There are two ways to process text. A single document can be analyzed in real time with an immediate result, and large document sets are gathered in S3 and processed with asynchronous batch jobs. This is where a common exam trap lies. Locating PII works in real time, but getting a copy with the PII redacted requires an asynchronous batch job. PII detection also supports only English and Spanish documents.
Comprehend Medical: For Medical Text Only
For clinical text such as physician notes and discharge summaries, the answer is Comprehend Medical, not regular Comprehend. It finds medical conditions, medications and dosages, test results, and protected health information (PHI), and links what it finds to standard medical codes such as RxNorm (medications) and ICD-10-CM (diagnoses). It supports only English (US).
Exam tip: "Extract medication names and dosages from physician notes" is Comprehend Medical, "automatically route customer emails into the company's eight department categories" is Comprehend custom classification, and "store a copy with personal data redacted in S3" is an asynchronous PII redaction job.
---
Speech: Transcribe and Polly
Transcribe turns speech into text (STT), and Polly turns text into speech (TTS). For these two services that work in opposite directions, the custom features that improve accuracy are the heart of the exam questions.
Amazon Transcribe: Speech to Text
| Feature | What it does | Clue in the question | |------|------|------| | Custom vocabulary | Gives hints so specific words (brand names, acronyms) are recognized | "the product name keeps being transcribed wrong" | | Custom language model | Trains on domain text so it learns the context in which words are used | "accuracy across an entire specialized field, such as conference talks" | | Speaker diarization | Separates the transcript by who spoke | "separate agent and customer speech" | | Automatic language identification | Detects the language of the audio automatically | "recordings in several languages are mixed" | | PII redaction | Masks personal data in the transcript | "card numbers must not remain in the transcript" | | Streaming transcription | Turns live speech into text immediately | "live captions during a call" |
A custom vocabulary is a hint that helps recognize words, while a custom language model also learns the context in which they are used. With a custom language model you must specify the language code yourself, so it cannot be combined with automatic language identification.
Amazon Polly: Text to Speech
Polly has four voice engines: generative, long-form, neural, and standard. There are two ways to change pronunciation and reading style, which makes them easy to confuse.
| Method | What it does | Best fit | |------|------|------| | SSML tags | Specifies pauses, emphasis, speed, and reading style inside the text | Pause only in this sentence, or read a number as a date | | Pronunciation lexicon | Stores per-word pronunciations or substitutions and applies them every time | Always read a brand name or acronym the same way | | Asynchronous synthesis task | Synthesizes text beyond the real-time limit (3,000 characters) with StartSpeechSynthesisTask and saves it to S3 | Long scripts such as audiobooks |
Exam tip: "Pronounce the company name correctly in every announcement" is a pronunciation lexicon, "pause for two seconds after a particular sentence" is SSML, and "turn a one-hour script into an audio file" is an asynchronous synthesis task.
---
Translation: Amazon Translate
Translate is a neural machine translation service. Short text goes through real-time translation (TranslateText), and large document sets collected in S3 go through batch translation jobs. What separates the answers is the two ways of tailoring translations to your company.
| Aspect | Custom terminology | Parallel data (Active Custom Translation) | |------|------|------| | What it does | Fixes specific words to the translation you defined | Shows examples of source text and desired translations to match style, tone, and word choice | | Where it applies | Real-time and batch translation | Batch translation jobs | | Example | Keep the product name "Amazon Photos" untranslated | Legal documents in the style of your legal team | | Cost | No extra charge | Higher rate than regular jobs |
Exam tip: "Brand and product names must not be translated" is custom terminology. "Translations must resemble the style of the company's existing translations" is parallel data.
---
Documents and Images: Textract and Rekognition
Amazon Textract: Pulling Text and Structure Out of Documents
Textract goes beyond simple character recognition (OCR) and also extracts the "field: value" pairs of forms and the structure of tables.
| API or feature | What it does | |------|------| | DetectDocumentText | Extracts printed and handwritten text | | AnalyzeDocument | Forms (key-value), tables, queries (finding a value by asking a question such as "What is the contract start date?"), signatures | | AnalyzeExpense | Fields such as amounts and vendors on invoices and receipts | | AnalyzeID | US government-issued IDs (driver's licenses, passports) |
Single-page images go through the synchronous API, and multi-page PDFs through asynchronous jobs that notify you through SNS. Textract stops at extracting text; to analyze its meaning (sentiment, entities), pass the result to Comprehend.
Amazon Rekognition: Seeing Images and Video
| Feature | What it does | Clue in the question | |------|------|------| | Label detection | Recognizes objects, scenes, and activities | "automatically tag objects in photos" | | Face detection, comparison, and search | Analyzes face attributes, compares two faces, searches a face collection | "match against registered employee faces" | | Content moderation | Detects inappropriate or offensive images and video | "screen photos uploaded by users" | | PPE detection | Checks whether hard hats, masks, and other protective gear are worn | "verify workers' safety equipment" | |