AIF-C01 fundamentals of AI and ML explained
Fundamentals of AI and ML is worth 20% of AIF-C01’s scored content — around 10 of the 50 scored questions. It is the vocabulary layer, and everything in the other four domains is expressed in it. Get this wrong and you will misread questions elsewhere.
The nesting
Learn this as concentric circles, because questions test the distinction:
- Artificial intelligence — the broad field of systems performing tasks that need human-like intelligence.
- Machine learning — a subset where systems learn patterns from data rather than following written rules.
- Deep learning — a subset of ML using multi-layered neural networks, behind most modern vision and language capability.
- Generative AI — applications producing new content, built on deep learning.
The three learning types
| Type | Data | Typical use |
|---|---|---|
| Supervised | Labelled — inputs with known answers | Classification, regression |
| Unsupervised | Unlabelled | Clustering, anomaly detection, grouping |
| Reinforcement | Rewards from an environment | Sequential decisions, robotics, game playing |
The exam’s usual framing: a scenario describes the data available, and you identify the learning type. If the historical data includes known outcomes, it is supervised. If you are looking for structure nobody has labelled, it is unsupervised.
Classification versus regression
Both supervised, and the most reliably tested distinction in the domain:
- Classification predicts a category — spam or not, which of five defect types, will this customer churn.
- Regression predicts a continuous number — next quarter’s revenue, tomorrow’s temperature, expected lifetime value.
If the answer is a number on a scale, it is regression. If it is a label from a set, it is classification.
Training and inference
Two distinct stages:
- Training — the model learns from data. Expensive, occasional.
- Inference — the trained model produces output for new input. Cheaper per call, continuous.
Much of the cost and architecture discussion in later domains rests on this split.
The terms worth knowing cold
- Features — the inputs. Labels — the known answers in supervised learning.
- Overfitting — the model memorised the training data and performs poorly on new data. Underfitting — it never learned the pattern.
- Bias — systematic error, including unfair error between groups. Variance — sensitivity to the particular training sample.
- Model — the learned artefact you deploy.
MLOps concepts
Task 1.3 asks you to describe the AI/ML development lifecycle, and one of its objectives is the fundamental ideas of MLOps: experimentation, repeatable processes, scalable systems, managing technical debt, production readiness, model monitoring and model re-training. The exam guide puts building pipelines out of scope for this candidate, so expect “which concept or service fits” questions, not configuration.
MLOps is DevOps practice applied to models. The reason it exists: a model can get worse without anyone changing its code, because the world it predicts keeps changing.
| Concept | What it means in practice |
|---|---|
| Experimentation | Trying data, algorithms and settings, and recording each run so the best one can be identified and reproduced |
| Repeatable processes | The same steps (prepare data, train, evaluate, deploy) run as an automated pipeline instead of by hand |
| Scalable systems | Training and inference that grow with data and traffic without redesign |
| Technical debt | Shortcuts such as manual steps, untracked data or one-off scripts that make the system fragile later |
| Production readiness | Versioned, tested, approved and secured before it serves real users |
| Model monitoring | Watching live inputs and predictions for drift and falling quality |
| Model re-training | Training again on fresh data when monitoring shows decay, then re-evaluating before release |
Read the last two as a loop: monitoring detects that live data no longer looks like the training data, re-training produces a new version, and that version is checked before it replaces the old one.
On AWS, these ideas map to Amazon SageMaker AI features: experiment tracking (managed MLflow), SageMaker Pipelines for repeatable workflows, the Model Registry for versions and approval, and Model Monitor for drift in production. At this level, knowing that SageMaker AI is where the MLOps tooling lives is usually enough.
Note that AWS has closed SageMaker Model Monitor and SageMaker Clarify to new customers (the closure took effect July 30, 2026). Existing customers keep using them, but AWS adds no new features. For new accounts AWS documents a replacement built from open-source monitoring solutions on SageMaker AI with MLflow, Amazon CloudWatch and Amazon QuickSight. Older practice questions may still name Model Monitor as the answer; the concept they test, monitoring for drift, is unchanged.
Exam pitfall: a scenario says a model’s accuracy has slowly dropped over months although nobody changed the code or the infrastructure. Answers such as “a larger instance” or “more endpoint capacity” fix nothing. The concept being tested is model monitoring plus re-training on recent data.
For the registry, CI/CD and both kinds of drift in more depth, see the MLOps basics guide.
Matching problems to AWS services
A substantial share of this exam is recognising which service addresses a described need. At foundational level you need purpose, not configuration:
- Amazon SageMaker AI — the platform for building, training and deploying models.
- Amazon Bedrock — managed access to foundation models.
- Pre-trained AI services for vision, speech, language and document processing, which solve common problems without training anything.
The instinct AWS rewards: if a managed service already does it, use the managed service. Building a custom model when a pre-trained service exists is rarely the expected answer at this level.
The managed AI services, by input and output
Task 1.2 asks you to explain the capabilities of AWS managed AI/ML services, and the exam guide names SageMaker AI, Transcribe, Translate, Comprehend, Lex and Polly. Its real-world applications (computer vision, NLP, speech recognition) pull in Rekognition and Textract as well. You never configure these services on AIF-C01; you recognise them from a one-line need. The reliable method is to name the input and the output first:
| Service | Input | Output | Typical scenario |
|---|---|---|---|
| Amazon Transcribe | Speech (audio) | Text | Subtitles, call transcripts, meeting notes |
| Amazon Translate | Text | Text in another language | Localising product pages or support tickets |
| Amazon Comprehend | Text | Sentiment, entities, key phrases, language, PII | Analysing reviews or tickets at scale |
| Amazon Polly | Text | Speech (audio) | Reading content aloud, voice prompts |
| Amazon Lex | Voice or text from a user | A conversation that recognises intent and fulfils it | Chatbots and voice bots for self-service |
| Amazon Textract | Scanned documents, forms | Text, forms and tables | Invoices, claims, ID documents |
| Amazon Rekognition | Images, video | Objects, faces, text, unsafe content | Moderating uploads, tagging photos |
Transcribe and Polly are mirror images (speech to text versus text to speech), and that pairing is a favourite distractor. Many scenarios chain services: recorded calls go through Transcribe, then Translate, then Comprehend for sentiment. Each step has one job, and none involves training a model.
Exam pitfall: Lex and Comprehend both “understand” text, but they do different jobs. Comprehend analyses text you already have and returns insights; Lex holds a back-and-forth conversation with a user to complete a task. A scenario about a customer-facing bot that books appointments is Lex, not Comprehend, and not Polly either, which only speaks.
Some AI services and features have been closed to new customers since the exam guide was written; the AWS AI services guide lists them with dates and gives the full input-to-output cheat sheet.
Sample questions
Question 1. A retailer has transaction data with no labels and wants to discover natural groupings of customers for marketing. Which approach fits?
- A. Supervised classification
- B. Supervised regression
- C. Unsupervised clustering
- D. Reinforcement learning
Show answer
Answer: C
Finding natural groupings in unlabelled data is clustering, an unsupervised technique. Classification and regression are supervised and require labelled outcomes, and reinforcement learning trains an agent through rewards from an environment.
Want more questions like this? Full AIF-C01 practice tests →
Question 2. A model scores 99% on the data it was trained on but 62% on new data it has never seen. What does this indicate?
- A. Overfitting
- B. Underfitting
- C. Insufficient inference capacity
- D. A context window that is too small
Show answer
Answer: A
Excellent performance on training data with poor performance on unseen data is the definition of overfitting: the model memorised the training set rather than learning a generalisable pattern. Underfitting would show poor performance on both, and neither insufficient inference capacity nor a small context window explains the gap.
Want more questions like this? Full AIF-C01 practice tests →
Question 3. A company needs to extract text from scanned invoices. It has no machine learning staff and wants the fastest route to a working solution. What should it use?
- A. Train a custom model in Amazon SageMaker AI
- B. A pre-trained AWS AI service for document text extraction
- C. Hire a machine learning team to build a solution
- D. Build a neural network from scratch on Amazon EC2
Show answer
Answer: B
A pre-trained managed AWS service for document text extraction solves a common problem with no model training and no ML expertise, which is what the constraints require. Training a custom model, hiring a team or building a neural network from scratch all add cost and time the scenario rules out.
Want more questions like this? Full AIF-C01 practice tests →
Question 4. A demand forecasting model was accurate at launch. A year later its forecasts are noticeably worse, although its code and infrastructure have not changed. Which MLOps practice addresses this?
- A. Move the endpoint to a larger instance type
- B. Track more experiments during the original training
- C. Freeze the current model version so it cannot change
- D. Monitor the model in production and re-train it on recent data
Show answer
Answer: D
Quality falling while code and infrastructure stay the same points to the live data drifting away from the training data. Model monitoring detects that drift and re-training on recent data restores accuracy. More compute does not change what the model learned, experiment tracking records training runs but does not watch production, and freezing the model keeps the decay.
Want more questions like this? Full AIF-C01 practice tests →
Question 5. A clinic wants a customer-facing assistant on its website that lets patients book, move and cancel appointments by typing or speaking. Which AWS service is designed for this?
- A. Amazon Comprehend
- B. Amazon Lex
- C. Amazon Polly
- D. Amazon Transcribe
Show answer
Answer: B
Amazon Lex builds conversational interfaces for voice and text that recognise the user intent and fulfil it, which is what a booking assistant needs. Comprehend analyses existing text for sentiment or entities but does not hold a conversation, Polly only turns text into speech, and Transcribe only turns speech into text.
Want more questions like this? Full AIF-C01 practice tests →
What to practise
Write one sentence for each term above in your own words, then for each of five business problems name the learning type and the AWS service you would reach for. That exercise is most of this domain, and the service-matching habit carries into every other domain on the exam.