Rubric Edu
Contents — domains, guide and mocks

When AI fits, when it doesn't, and which AWS service does the job

AIF-C01 1.215 min read · checked 6 November 2026

Task statementIdentify practical use cases for AI.

Where AI/ML adds value

The guide names three kinds of value. Recognise them in a scenario and you can usually eliminate half the options.

  • Assisting human decisions. The model narrows, ranks or flags; a person decides. Fraud alerts for an analyst, a shortlist of likely diagnoses for a clinician to review.
  • Scalability. The task is easy for a person but impossible at volume: reading every product review, tagging every uploaded image, translating every support article.
  • Automation. A repetitive judgement — extracting fields from invoices, transcribing calls — done without a person in the loop for the routine cases.

When AI/ML is not appropriate

Should this be an ML problem at all?

What does the business need?
  • An exact, deterministic result that must be right every time
    Use rules or ordinary code
  • The cost of building and running the model exceeds the value of better predictions
    Don't build it yet
  • Too little relevant data, or no way to tell right from wrong
    Collect data first
  • A useful prediction or generated draft at scale
    AI/ML is a candidate

Matching the technique to the problem

TechniqueOutputExampleLearning type
RegressionA numberPredict next month's sales; estimate a house priceSupervised
ClassificationA categorySpam or not; which product line a ticket belongs toSupervised
ClusteringGroups the data forms on its ownSegment customers by behaviour with no predefined segmentsUnsupervised

Real-world applications the guide names

ApplicationWhat it doesTypical AWS fit
Computer visionDetects objects, faces, text and activity in images and videoAmazon Rekognition
NLPFinds sentiment, entities, key phrases and topics in textAmazon Comprehend
Speech recognitionTurns speech into textAmazon Transcribe
Recommendation systemsSuggests items based on behaviourAmazon Personalize
Fraud detectionFlags likely fraudulent activityAmazon Fraud Detector, or a custom model in SageMaker AI
ForecastingPredicts future values from time-series historyAmazon Forecast, or a custom model in SageMaker AI
Knowledge basesAnswers questions from an organisation's documentsAmazon Kendra for search; Amazon Bedrock Knowledge Bases for generative answers
Agentic AIPursues a goal across steps, calling toolsAgents built on Amazon Bedrock

The managed services named in the objective

The AWS decision guide describes three tiers. Pre-built AI services solve a specific task with no ML expertise. Amazon SageMaker AI is for building, training and deploying custom models. Specialised infrastructure (for example AWS Trainium and Inferentia chips) is for teams that want to manage the underlying resources themselves. AIF-C01 lives almost entirely in the first two tiers.

ServiceInput → outputPick it when the scenario says…
Amazon TranscribeAudio → text“transcribe calls”, “captions”, “speech to text”
Amazon PollyText → lifelike speech“read aloud”, “voice for an app”, “text to speech”
Amazon TranslateText → text in another language“localise”, “translate support articles”
Amazon ComprehendText → sentiment, entities, key phrases, topics“analyse reviews”, “extract entities”, “detect sentiment”
Amazon LexBuilds conversational bots (voice and text)“chatbot”, “IVR”, “conversational interface”
Amazon SageMaker AIYour data → your own trained model“custom model”, “our own algorithm”, “full control over training”

Traditional ML or a foundation model?

The guide asks you to choose based on regulatory concerns, explainability requirements and operational constraints. Foundation models are flexible and need no task-specific training, but they are large, their reasoning is hard to inspect, and their output can vary from run to run. Traditional models are narrow but small, cheap to run, and — for model types like linear models and decision trees — far easier to explain.

Choosing between the two

Lean towards traditional ML

  • A regulator or customer must be told why a decision was made
  • A narrow, well-defined prediction on tabular data
  • Tight latency or cost per prediction
  • Results must be consistent run to run

Lean towards a foundation model

  • Open-ended language, images or conversation
  • Many different tasks from one model
  • Little or no labelled data for the task
  • Generating drafts, summaries or answers

Traps the wrong answers are built from

Tempting but wrongDo this instead
Using ML where a fixed rule gives the exact answer.If the outcome must be deterministic and correct every time, write the rule.
Building a custom SageMaker AI model for a task a managed service already does.Check the pre-built services first: Transcribe, Translate, Comprehend, Rekognition, Polly, Lex.
Choosing an LLM for a regulated decision that must be explained.Prefer an interpretable traditional model when explainability is a requirement.
Calling customer segmentation “classification”.With no predefined labels it is clustering — unsupervised.

You should now be able to

  • Recognise when AI/ML adds value through assisted decisions, scale or automation.
  • Recognise when AI/ML is inappropriate: exact outcomes required, or costs outweigh benefits.
  • Choose regression, classification or clustering from a problem description.
  • Match real-world applications to the right kind of AI.
  • Pick the right managed service: SageMaker AI, Transcribe, Translate, Comprehend, Lex, Polly.
  • Choose between a traditional ML model and a foundation model using regulatory, explainability and operational constraints.

Practice questions

Original questions written for this lesson, in the exam’s style. Answer first, then open the reasoning — every option is explained, including why the wrong ones are tempting.

  1. Question 1

    A travel company wants its mobile app to read booking confirmations aloud to visually impaired customers in a natural-sounding voice.

    Which AWS service should it use?

    1. AAmazon Polly
    2. BAmazon Transcribe
    3. CAmazon Comprehend
    4. DAmazon Lex
    Show answer and reasoning
    1. ACorrect. Polly converts text into lifelike speech.
    2. BIncorrect. Transcribe does the opposite — it converts speech into text.
    3. CIncorrect. Comprehend analyses text for sentiment and entities; it does not produce audio.
    4. DIncorrect. Lex builds conversational bots. Reading a confirmation aloud needs text-to-speech, not a conversation.

  2. Question 2

    An online shop has a year of purchase history but no predefined customer segments. Marketing wants to discover natural groups of customers with similar buying behaviour.

    Which technique fits?

    1. AClustering
    2. BClassification
    3. CRegression
    4. DReinforcement learning
    Show answer and reasoning
    1. ACorrect. With no labels, finding natural groups is clustering — an unsupervised technique.
    2. BIncorrect. Classification assigns examples to categories that are already defined and labelled. Here the segments don't exist yet.
    3. CIncorrect. Regression predicts a number, not a grouping.
    4. DIncorrect. There is no environment giving rewards for actions; this is pattern discovery in historical data.

  3. Question 3

    A payroll team asks for an AI model to calculate each employee's statutory tax deduction from salary and published tax bands.

    What is the best recommendation?

    1. AImplement the published rules directly in code; ML is not appropriate.
    2. BTrain a regression model on past payslips.
    3. CUse a foundation model with a prompt describing the tax bands.
    4. DUse Amazon Comprehend to read the tax rules.
    Show answer and reasoning
    1. ACorrect. The outcome is deterministic and must be exactly right every time. A prediction — even a very accurate one — is the wrong tool.
    2. BIncorrect. A regression model would approximate the answer and occasionally be wrong, which is unacceptable for a statutory calculation.
    3. CIncorrect. Foundation models can be nondeterministic and can make arithmetic errors; neither is acceptable here.
    4. DIncorrect. Comprehend analyses text; it does not perform calculations.

  4. Question 4

    A regulated lender must give every declined applicant the main reasons for the decision. It has ten years of labelled, tabular application data.

    Which approach is MOST appropriate?

    1. AA traditional, interpretable supervised model trained on the tabular data
    2. BA large language model prompted to approve or decline each application
    3. CClustering applicants into groups
    4. DAmazon Polly
    Show answer and reasoning
    1. ACorrect. Explainability is a stated requirement, the data is labelled and tabular, and the decision must be consistent — the cases the guide associates with traditional ML.
    2. BIncorrect. LLM reasoning is hard to inspect and output can vary between runs, which conflicts with the explanation and consistency requirements.
    3. CIncorrect. Clustering finds groups but does not make an approve/decline prediction, and the data is already labelled.
    4. DIncorrect. Polly converts text to speech and has nothing to do with credit decisions.

  5. Question 5

    A support centre records customer calls in Spanish. It wants English transcripts and then a sentiment score for each call.

    Which THREE services, used together, meet the requirement? (Select THREE.) (Select 3.)

    1. AAmazon Transcribe
    2. BAmazon Translate
    3. CAmazon Comprehend
    4. DAmazon Polly
    5. EAmazon Lex
    Show answer and reasoning
    1. ACorrect. Converts the Spanish audio into Spanish text.
    2. BCorrect. Translates the Spanish transcript into English.
    3. CCorrect. Detects sentiment in the English text.
    4. DIncorrect. Polly produces speech from text. Nothing here needs to be spoken.
    5. EIncorrect. Lex builds bots; the requirement is to analyse recordings, not hold a conversation.

Practise every question in Domain 1 →

Sources

Drafted with AI assistance and checked against the sources above; expert review is in progress. Spotted an error? Tell us and it gets fixed, dated and listed on how this is written.