Fundamentals of AI and ML

Chapter 01 — How AIF-C01 Thinks, and the Language of AI and ML

Certification Blueprint

Field	Coverage
Exam	AWS Certified AI Practitioner (AIF-C01), exam guide v1.1
Domain	Content Domain 1 — Fundamentals of AI and ML
Exam weight	20% of scored content
Task statement	1.1 Explain basic AI concepts and terminologies
Objectives	1.1.1 terms · 1.1.2 nesting · 1.1.3 inferencing types · 1.1.4 data types · 1.1.5 learning types

What This Chapter Covers

This chapter does two jobs. First, it gives you the exam's operating model — how it is scored, what the four question types punish, and the order in which to read a question. Second, it builds the vocabulary that every later chapter assumes you already have.

Domain 1 is 20% of scored content, but it is closer to 100% of the language. A learner who is shaky on the difference between a model and an algorithm will misread questions in Domain 3, where the stakes are higher and the wording is denser. That is the reason this chapter comes first and the reason it is worth over-learning.

Only Task 1.1 is covered here. Task 1.2 — choosing techniques, recognising when AI is the wrong answer, and the AWS managed services — is Chapter 02. Task 1.3, the lifecycle and MLOps, is Chapter 03.

How the Exam Is Scored

Four facts change how you should study:

• 65 questions are presented; 50 are scored. The 15 unscored questions are not identified anywhere, so there is no way to spot them and no benefit in trying.
• The scale is 100–1,000 and the passing score is 700. This is a scaled score, not a percentage. Scaling equates forms of slightly different difficulty, so 700 does not correspond to a fixed number of correct answers, and AWS does not publish one.
• Scoring is compensatory. There is no per-domain pass bar; you pass the exam as a whole. This licenses proportional effort, not abandonment — Domains 4 and 5 together are 28% of scored content, more than any single domain including Domain 3.
• An unanswered question is scored as incorrect and there is no guessing penalty. A blank is strictly worse than a guess. Never leave one.

The four question types each punish a different weakness. Multiple choice punishes picking a true statement instead of the best one. Multiple response has no partial credit — two of three correct scores zero. Ordering punishes knowing the components of a process but not their sequence. Matching punishes a vague sense of which service does what.

The Course Map

What to remember from this diagram: chapters are how you study; domains are how the exam is weighted. They are deliberately not the same shape. Domain 3 is the heaviest at 28% and is spread across four chapters; Domains 4 and 5 are 14% each and get two chapters apiece. Every chapter declares which domain and task statement it serves, so you can always map your progress back to the official blueprint.

How to Read an AIF-C01 Question

Work these four layers in order, and do not jump to the service:

Layer	What to extract
1. Business constraint	Regulator, budget, ownership, risk of being wrong
2. Data constraint	Labeled or not? How much? Structured or not?
3. Operational constraint	Who waits? What volume? What arrival pattern?
4. AWS mechanism	Only now: which service or approach

Three of the four options will usually be true statements about AWS. Being true is not the test. Being best for the stated constraint is the test.

The Nesting Model

What to remember from this diagram: each layer sits strictly inside the one before it. Two sentences carry the whole idea, and they run in opposite directions — every agentic system is generative, but not every AI system is machine learning.

Layer	The boundary that defines it
AI	Performs tasks that normally need human intelligence — including hand-written rules
Machine learning	Behaviour is learned from data, not explicitly programmed
Deep learning	Multi-layer neural networks that learn features from raw data
Generative AI	Produces new content rather than only classifying or scoring
Agentic AI	Plans, calls tools, and acts over multiple steps toward a goal

The boundary most often missed is the first one. A spam filter built from a hand-written keyword list is AI — it performs a task that would otherwise need human judgment — but it is not machine learning, because nothing was learned from data. Rebuild the same filter by training it on a corpus of labelled mail and it becomes machine learning. The system's purpose did not change; the source of its behaviour did, and that is the boundary.

Agentic AI was added to Objectives 1.1.1 and 1.1.2 in version 1.1 of the exam guide. If you are studying from older material, you have not seen it.

Algorithm, Training, Model, Inferencing

What to remember from this diagram: four words that answer options routinely swap for each other. The algorithm is the recipe. The model is the cake. Training bakes it; inferencing serves a slice. Or, if you prefer a second angle: the algorithm is the driving lesson, the model is the licensed driver, and inferencing is the journey they actually make.

Overfitting and underfitting branch off the training step, but you only notice them at inference time — which is exactly why exam questions describe them as a gap between two numbers rather than as a property you could see directly.

Detailed Concepts

Types of data

Three axes describe data, but only one of them changes the answer to an exam question.

Axis	Values	Why it matters
Labeled vs unlabeled	Answer known per example, or not	Decides which learning type is available
Structured vs unstructured	Defined schema, or not	Decides storage and preparation
Modality	Tabular · time-series · image · text	Decides technique family

A scenario that never mentions recorded outcomes has quietly ruled out supervised learning. That absence is a deliberate signal, not an oversight in the question.

The three learning types

Type	Requires	Produces	Signal phrase
Supervised	Labeled data	Prediction of a known target	"historical outcomes", "past examples marked as…"
Unsupervised	Unlabeled data	Structure already in the data	"we don't know what groups exist"
Reinforcement	Environment + reward	A policy for acting	"trial and error", "maximise reward over time"

Reinforcement learning is over-selected because it sounds advanced. It requires an environment to act in and a reward signal to learn from. Most business scenarios have neither.

The four inferencing modes

Training is when the model learns. Inferencing is when it is used — and it is where the ongoing cost lives. Learners consistently assume training dominates cost; for a system serving traffic every day, it usually does not.

Mode	Use when	Deciding constraint
Real-time	Someone is waiting	Low predictable latency
Batch	Bulk set on a schedule	Throughput and cost; nobody waits
Asynchronous	Large payloads, long processing	Payload size and duration
Serverless	Intermittent, unpredictable traffic	No idle cost; cold start acceptable

Version 1.1 expanded this objective from two modes to four. Older material covers only real-time and batch.

Decision Rules and Exam Signals

Rule 1 — the label signal decides the learning type. Ask one question: do we know the answer for past examples? If yes, supervised is available. If no, you are choosing between discovering structure (unsupervised) and learning from consequences (reinforcement). Volume of data is not a label. Neither is the industry, and neither is how modern the company sounds.

Rule 2 — the inferencing mode is decided by who waits, then by schedule, then by payload. Work the questions in that order. The closest pair on the whole exam is asynchronous versus batch, and the separator is what decides the mode: payload and duration for asynchronous, bulk and schedule for batch. A large file processed nightly is still batch.

Rule 3 — diagnose fit by comparing two numbers. Overfitting is high on training data and low on new data. Underfitting is low on both. If a question gives you both numbers, it is giving them to you for a reason.

Rule 4 — bias is a model property, fairness is an outcome property. A biased model can be corrected downstream into fair outcomes; an unbiased model deployed into an unequal process can still produce unfair ones. They are separate objectives in Domain 4 for exactly this reason.

Distractor Patterns

Pattern	What it looks like	How to defuse it
Right layer, wrong boundary	Calling a rules engine "machine learning"	Ask what is learned from data
Training/inferencing swap	"Reduce training cost" when the volume is at serving time	Ask when the cost is incurred
Fit inversion	Offering "underfitting" for a model that aced training and failed production	Compare training vs production performance
Bias for fairness	Using the two terms interchangeably	Bias = model property; fairness = outcome property
Volume as the label signal	"Millions of records, so supervised"	Volume is not a label; ask if outcomes were recorded
Async for batch	Large payload wrongly steering away from a nightly job	Schedule beats payload size

These six account for most of the Task 1.1 questions that are answered incorrectly.

Scenario Walkthrough

A logistics operator holds four years of sensor readings from its fleet. Nobody has ever recorded which readings preceded a breakdown. The team wants to know whether the vehicles fall into natural groups that behave differently, so engineers can investigate each group. Readings arrive continuously, but the analysis is run once a month over the whole archive.

Constraint	Reading	Decision
No recorded outcomes	Labels absent — supervised unavailable	Unsupervised
"Natural groups nobody has defined"	Discovering structure	Confirms unsupervised
Whole archive, once a month	Bulk on a schedule, nobody waiting	Batch inferencing
Continuous arrival	⚠️ Describes ingestion, not inference	Does not make it real-time

The outcome is unsupervised learning with batch inferencing.

The trap is the phrase "readings arrive continuously". It is planted to pull you toward real-time, and it describes the arrival pattern of data, not of inference requests. Those are different things, and separating them is the whole exercise.

Key Concepts

Term	Definition
Artificial intelligence	Systems performing tasks that normally require human intelligence, including systems built from hand-written rules with no learning at all
Machine learning	The subset of AI whose behaviour is learned from data rather than explicitly programmed
Deep learning	The subset of machine learning using multi-layer neural networks that learn features directly from raw data
Generative AI	The subset of deep learning that produces new content rather than only classifying or scoring existing content
Agentic AI	The subset of generative AI that plans, calls tools, and acts over multiple steps toward a goal
Algorithm	The procedure used to produce a model from data — the recipe, not the result
Model	The trained artefact that makes predictions — the result, not the recipe
Training	The process in which a model learns patterns from data
Inferencing	The process in which a trained model produces a result from new input; where ongoing cost lives
Overfitting	The model memorised its training data and fails on new data — high training score, low live score
Underfitting	The model is too simple to capture the pattern — low score on both training and live data
Bias	A systematic error in a model that skews its outputs in a consistent direction
Fairness	Whether the outcomes a system produces are equitable across groups of people
Supervised learning	Learning from labeled data to predict a known target
Unsupervised learning	Learning structure already present in unlabeled data
Reinforcement learning	Learning a policy for acting by taking actions and receiving rewards

Revision Flashcards

Say the answer aloud before revealing it.

1. What single question decides the learning type? → Do we know the answer for past examples? If yes, supervised is available; if no, choose between unsupervised and reinforcement.

2. Is a hand-written rules engine artificial intelligence? → Yes. It performs a task that would otherwise need human judgment. It is not machine learning, because nothing is learned from data.

3. What is the difference between an algorithm and a model? → The algorithm is the procedure that produces the model. The model is the trained artefact that makes predictions. Recipe versus cake.

4. How do you tell overfitting from underfitting? → Compare two numbers. Overfitting is high on training and low on new data. Underfitting is low on both.

5. What separates bias from fairness? → Bias is a property of the model — a systematic skew in its outputs. Fairness is a property of the outcome — whether results are equitable across groups.

6. Name the four inferencing modes and the constraint that selects each. → Real-time (someone is waiting), batch (bulk on a schedule, nobody waits), asynchronous (large payload or long processing), serverless (intermittent traffic, no idle cost).

7. What is the closest pair in the inferencing modes, and what separates them? → Asynchronous versus batch. Asynchronous is decided by payload size and duration; batch by bulk and schedule. A large file processed nightly is still batch.

8. Which axis of data actually changes the answer to an exam question? → Labeled versus unlabeled. It decides which learning type is available at all.

9. Does 700 mean 70% of questions correct? → No. It is a scaled score on a 100–1,000 scale that equates forms of differing difficulty. No fixed number of correct answers maps to it.

10. What does compensatory scoring mean, and what does it not license? → There is no per-domain pass bar; you pass overall. It does not license skipping domains — Domains 4 and 5 are 28% combined.

11. Why should you never leave a question blank? → An unanswered question is scored as incorrect and there is no guessing penalty, so a blank is strictly worse than a guess.

12. What is the boundary that defines agentic AI? → It plans, calls tools, and acts over multiple steps toward a goal, rather than producing a single response.

The Four-Beat Answer

The core question this chapter prepares you for: "Walk me through how you would classify an AI problem before choosing any technology."

Four beats. Interviewers and exam questions check for the same four, in this order. Missing a beat is a failure state — you will be probed on whichever one you skipped.

1. Layer — place the system in the AI → ML → deep learning → generative → agentic nesting, and name the boundary that put it there. Say whether anything is actually learned from data.
2. Data — state whether outcomes were recorded for past examples. That single fact decides whether supervised learning is even available, before any technique is discussed.
3. Learning type — name supervised, unsupervised or reinforcement, and justify it from the label signal rather than from the size of the dataset or the industry.
4. Serving — state the inferencing mode and the constraint that selected it: who is waiting, whether the work is scheduled in bulk, and how large or slow each request is.

Only after all four beats does a service name belong in the answer. Naming a service first is the single most common way to lose a question that you actually knew.

Why This Helps You

On the job: these four beats are the opening move of every real AI project scoping conversation. Teams that skip beat 2 discover halfway through a build that nobody ever recorded the outcomes they intended to predict, and the project quietly becomes an unsupervised one.

In interviews: "what is the difference between AI and machine learning" is asked constantly, and most candidates answer it as a synonym pair. Naming the boundary — learned from data versus programmed — separates you immediately.

On the exam: every Domain 1 question and a large share of Domains 2 and 3 assume this vocabulary is automatic. The reading habit taught here — constraint before service — is the single highest-value thing in the chapter.

Chapter Checklist

• I can place any system in the AI → ML → deep learning → generative → agentic nesting and name the boundary that put it there
• I can explain why a hand-written rules engine is AI but not machine learning
• I can separate a model from an algorithm, and training from inferencing
• I can tell overfitting from underfitting using the training-versus-live gap
• I can distinguish bias as a model property from fairness as an outcome property
• I can name the three learning types and pick one from the label signal alone
• I can name all four inferencing modes and the constraint that selects each
• I can separate asynchronous from batch and say which fact decides it
• I can name the six Task 1.1 distractor patterns and the question that defuses each
• I know that 50 of 65 questions are scored, that 700 passes on a scaled 100–1,000 range, and that scoring is compensatory

After the Chapter

1. Complete student/project.md — the AI/ML Decision Sheet. Keep it; every later chapter adds to it.
2. Take student/quiz.md closed-book, then review the reasoning for every question you guessed, including the ones you got right.
3. Open the official v1.1 exam guide's Domain 1 page and confirm you can attach a concept from this chapter to each of the five bullets under Task 1.1. If you cannot attach one, that is where to revise.
4. Next chapter: Chapter 02 — Where AI Fits, and Where It Does Not (Domain 1, Task 1.2). Choosing regression, classification or clustering from the output shape; recognising when AI/ML is the wrong answer entirely; the AWS managed AI services and the direction each one runs; and the v1.1 objective that decides between traditional ML and a foundation model by constraint rather than by task.

Chapter 02 — Where AI Fits, and Where It Does Not

Certification Blueprint

Field	Coverage
Exam	AWS Certified AI Practitioner (AIF-C01), exam guide v1.1
Domain	Content Domain 1 — Fundamentals of AI and ML
Exam weight	20% of scored content
Task statement	1.2 Identify practical use cases for AI
Objectives	1.2.1 value · 1.2.2 when not appropriate · 1.2.3 techniques · 1.2.4 applications · 1.2.5 AWS managed services · 1.2.6 traditional ML vs FM

What This Chapter Covers

Chapter 01 taught classification — what a system is, what its data permits, how it will be served. This chapter teaches selection: whether to use AI at all, which technique the output shape requires, which AWS service already solves the problem, and whether the answer is a traditional model or a foundation model.

The unusual part is that selection includes saying no. Objective 1.2.2 asks you to identify when AI/ML is not an appropriate solution, and the exam genuinely marks the non-AI option correct when the scenario calls for it. Most of this exam rewards recognising the right answer. Task 1.2 rewards recognising that three plausible, technically accurate answers are wrong.

Chapter 01 answered	Chapter 02 answers
What is this system?	Should we build it with AI at all?
Which learning type does the data permit?	Which technique does the output shape require?
How will it be served?	Which AWS service already does this?
—	Traditional ML or a foundation model?

Where AI/ML Provides Value

Objective 1.2.1 names three patterns, and the exam expects the names.

Pattern	What it looks like	The tell in a scenario
Assisted decision making	Surface a recommendation; a human decides	"help the team prioritise", "flag for review"
Solution scalability	Handle volume no team could review manually	"millions of items", "cannot hire enough reviewers"
Automation	Remove repetitive judgment work	"currently done manually for every case"

These overlap in real systems — a claims triage tool is arguably all three. On the exam, pick the one the scenario spends its words on. If two sentences describe the size of the backlog, it is scalability. If the scenario stresses that a person still signs off, it is assisted decision making even though work is clearly being saved.

When AI Is the Wrong Answer

What to remember from this diagram: three gates, and two of them are exits. Most people picture AI adoption as a funnel ending in "yes". This one ends in "no" three ways out of four, and that is the shape Objective 1.2.2 is testing.

Reject AI/ML in these cases:

• A specific guaranteed outcome is required, not a prediction. Tax owed, a statutory benefit entitlement, interest accrued, a bill computed from a published tariff. The correct answer is defined by rule, so a prediction is wrong by construction. A model that is 99.7% accurate is a defect here — ask what the other 0.3% means when it is somebody's tax bill.
• Cost-benefit fails. Labeling, re-labeling, training, re-training on drift, monitoring and on-call are all recurring costs.
• Data is insufficient or unstable. Either no outcomes were ever recorded — which removes supervised learning entirely, per Chapter 01's label signal — or the underlying process changes so often that historical data no longer describes the current world. The second is subtler and is usually signalled by a phrase about the business changing frequently.

The cost-benefit test has a baseline problem

The trap is comparing the model against nothing, where it always looks good. The honest comparison is against the alternative that actually exists, which is usually a rules table maintained by two people.

Cost that recurs	Often forgotten
Labeling and re-labeling	Yes
Training and re-training on drift	Yes
Monitoring, alerting, on-call	Yes
Inference at production volume	Sometimes
A rules table maintained by two people	This is the comparison

If a deterministic rules table reaches the same outcome, it wins — it is cheaper, exact, and explainable at no extra cost. That last property is worth naming: explainability is free in a rules table and expensive in a model.

Choosing the Technique

What to remember from this diagram: classification and clustering both produce groups. The difference is that classification's groups already have names.

Output shape	Technique	Example
A continuous number	Regression	Beds occupied, price, demand
One of N known categories	Classification	Fraud/legitimate, defect type
Groups nobody defined	Clustering	Customer segments

Ask what the output looks like, not what the domain is. Healthcare is not a technique. Neither is "text" and neither is "images".

The scenario says…	Technique	Why
"Predict how many beds will be occupied next Tuesday"	Regression	A continuous number
"Decide whether this transaction is fraudulent"	Classification	Two known categories
"Find natural groupings of customers we haven't defined"	Clustering	Groups do not exist yet
"Sort defects into the six fault codes we already use"	Classification	The categories are already named
"Forecast next quarter's demand"	Regression	Continuous, over time

Rows three and four are a deliberate pair — same shape of sentence, opposite answers, and a single clause decides it. Named categories → classification. Unnamed groups → clustering.

Applications and Their Technique Families

What to remember from this diagram: both v1.1 additions — knowledge bases and agentic AI — hang off language models. If you are revising from material written before 2026-04-30, neither appears in your copy of this objective.

Application	Technique family
Computer vision	Deep learning on images
NLP	Language models
Speech recognition	Audio → text
Recommendation systems	Collaborative/content filtering
Fraud detection	Supervised classification
Forecasting	Time-series regression
Knowledge bases (new in v1.1)	Retrieval over a corpus
Agentic AI (new in v1.1)	Multi-step tool-using systems

AWS Managed AI Services

What to remember from this diagram: read the arrow, not the service name. Transcribe and Polly are the same edge travelled in opposite directions, and swapping them is the single most common distractor in this domain.

These services are pre-built: no training, no labeled data, no model to manage.

Service	Does	Reverse service
Amazon Transcribe	Speech → text	Polly
Amazon Polly	Text → speech	Transcribe
Amazon Translate	Language → language	—
Amazon Comprehend	Text → entities, sentiment, topics	—
Amazon Lex	Conversational interfaces	—
Amazon SageMaker AI	Build, train, deploy your own	—

The scenario says…	Service	The tell
"Generate an audio version of each article"	Polly	Producing speech
"Search the call archive by what was said"	Transcribe	Producing text from audio
"Detect the sentiment of support tickets"	Comprehend	Extracting meaning from text
"Serve the help centre in nine languages"	Translate	Language to language
"Build a phone menu that understands intent"	Lex	Conversational interface
"We need a custom model on our own data"	SageMaker AI	Nothing pre-built fits

SageMaker AI is correct only in that last row, and it is correct because nothing pre-built fits — not because it is more powerful. If a managed service already solves the stated problem, building a custom model is the wrong answer, and it will be offered, because it is a true statement about AWS and therefore hard to eliminate on knowledge alone.

Traditional ML or Foundation Model

What to remember from this diagram: notice what is absent — the industry, the data type, and how modern the company sounds. Every one of those appears in distractors.

Objective 1.2.6 was added at v1.1 and names exactly three deciding factors:

1. Regulatory concerns
2. Explainability requirements
3. Operational constraints

The instinct is to choose by task — text implies a foundation model, tabular implies traditional ML. The objective is explicitly constraint-driven. A text classification task under a regulator, with plenty of labeled examples and a tight per-inference cost budget, is traditional ML, and the fact that it involves language changes nothing.

Constraint	Choose	Why
Must be explained to a regulator	Traditional ML	Coefficients are inspectable and reproducible
Auditable, reproducible outputs	Traditional ML	Deterministic, fixed feature set
Narrow task, plenty of labels	Traditional ML	Cheaper, faster, usually more accurate
Broad language understanding, many tasks	Foundation model	Generalises without task-specific data
Very little labeled data	Foundation model	In-context learning needs no training set
Content must be generated	Foundation model	Traditional models score; they don't generate
Tight latency + cost at high volume	Traditional ML	Far cheaper per inference

Explainability is not articulacy

A foundation model asked why it declined an application will produce a fluent, plausible, well-structured paragraph. That paragraph is generated text about a decision. It is not a faithful account of the computation that produced the decision, and it is not reproducible — ask twice and you may get two different paragraphs.

A regulator requires the second thing. The test to apply: could an auditor re-run the process and obtain the same factor weights? For a linear model, yes. For a generated explanation, no.

This single distinction decides the hardest Domain 1 questions, and it returns in Chapters 12 and 13.

Decision Rules and Exam Signals

Rule 1 — screen before selecting. Run the three suitability gates before naming any technique or service. Two of the three are exits.

Rule 2 — the output shape picks the technique. Not the industry, not the data type. Continuous number, named category, or undefined group.

Rule 3 — named groups mean classification. If the scenario tells you what the buckets are called, clustering is wrong.

Rule 4 — read the arrow, not the service name. Verbalise every managed service as "from X to Y".

Rule 5 — a managed service beats a custom model when it already fits. SageMaker AI is the escape hatch, not the default.

Rule 6 — Objective 1.2.6 is constraint-driven. Regulatory, explainability, operational. Data type is not on the list.

Distractor Patterns

Pattern	What it looks like	How to defuse it
AI offered where it is forbidden	A prediction proposed for a statutory calculation	Ask whether an exact answer is legally required
Real service, wrong direction	Polly offered for transcription	Read the arrow, not the name
Custom model over managed service	SageMaker AI where Comprehend fits	Ask if something pre-built already does it
Technique chosen by industry	"Healthcare, so classification"	Ask what the output looks like
Clustering for named categories	Clustering offered for six known fault codes	If the groups have names, it is classification
FM chosen by data type	"It's text, so a foundation model"	Objective 1.2.6 is constraint-driven
Fluent explanation accepted as explainability	An FM that "explains its reasoning" for a regulator	Plausible text is not a reproducible account

The last two separate a pass from a fail on this task statement.

Scenario Walkthrough

A utility must calculate each customer's bill from meter readings and a published tariff. The regulator publishes the tariff and audits the calculation. The company also wants to reduce the 900,000 support emails it receives each year by routing them automatically to the right team, and to publish an audio version of every bill explanation for accessibility.

Requirement	Reading	Decision
Calculate the bill from a published tariff	Exact answer legally defined	Not AI/ML — deterministic calculation
Route 900,000 emails to the right team	Text → one of N known teams	Classification; Comprehend fits
Publish an audio version	Text → speech	Amazon Polly

Three requirements, three different answers — and one of them is "do not use AI".

The failure mode is answering all three with a single service. Real exam questions do a smaller version of this by embedding one disqualifying clause inside an otherwise ordinary scenario.

Key Concepts

Term	Definition
Assisted decision making	An AI value pattern where the system surfaces a recommendation and a human makes the final decision
Solution scalability	An AI value pattern where the system handles a volume of work no team could review manually
Automation	An AI value pattern where repetitive judgment work is removed from people entirely
Regression	A technique producing a continuous numeric output, such as price or demand
Classification	A technique assigning an input to one of N categories that already exist and have names
Clustering	A technique discovering groupings that nobody defined in advance; a human interprets what they mean afterwards
Knowledge base	An application pattern retrieving over a corpus of documents; added to Objective 1.2.4 at v1.1
Amazon Transcribe	Managed service converting speech to text
Amazon Polly	Managed service converting text to speech — the reverse of Transcribe
Amazon Translate	Managed service converting text from one language to another
Amazon Comprehend	Managed service extracting entities, sentiment and topics from text
Amazon Lex	Managed service for building conversational interfaces
Amazon SageMaker AI	Service for building, training and deploying your own models when nothing pre-built fits
Traditional ML	A model with a fixed feature set and inspectable, reproducible behaviour; preferred under regulatory, explainability or tight cost constraints
Foundation model	A large pre-trained model that generalises across tasks; preferred when labels are scarce, the task is broad, or content must be generated
Explainability	A faithful, reproducible account of how a decision was computed — not a fluent description of it

Revision Flashcards

Say the answer aloud before revealing it.

1. Name the three value patterns in Objective 1.2.1. → Assisted decision making, solution scalability, and automation.

2. What are the two ways a scenario disqualifies AI/ML? → A specific guaranteed outcome is required rather than a prediction, or the cost-benefit analysis fails against the honest alternative.

3. Why is a 99.7% accurate model a defect for a statutory calculation? → Because the correct answer is defined by rule. A prediction is wrong by construction, and the 0.3% is somebody's real entitlement.

4. What is the correct baseline for a cost-benefit comparison? → The honest alternative that already exists — usually a deterministic rules table — not "doing nothing".

5. What decides the technique: regression, classification or clustering? → The shape of the required output. A continuous number, one of N named categories, or groups nobody has defined.

6. What separates classification from clustering? → Classification's groups already exist and have names. Clustering discovers groupings nobody defined, and a human interprets them afterwards.

7. Which direction does Amazon Transcribe run, and what is its reverse? → Speech to text. Its reverse is Amazon Polly, which runs text to speech.

8. When is Amazon SageMaker AI the right answer? → Only when nothing pre-built fits and a custom model on your own data is genuinely required — not because it is more capable.

9. Name the three deciding factors in Objective 1.2.6. → Regulatory concerns, explainability requirements, and operational constraints.

10. The scenario says the data is text. Does that select a foundation model? → No. Data type is not one of the three deciding factors. A text task under a regulator with plenty of labels is traditional ML.

11. Why is a foundation model's fluent explanation not explainability? → It is generated text about a decision rather than a faithful, reproducible account of how the decision was computed. Asking twice may produce two different answers.

12. Which two applications were added to Objective 1.2.4 at v1.1, and what family do they belong to? → Knowledge bases and agentic AI. Both hang off language models.

The Four-Beat Answer

The core question this chapter prepares you for: "How do you decide whether — and how — to apply AI to a business problem?"

Four beats, checked in this order. Missing a beat is a failure state.

1. Suitability — is an exact answer required by rule, is the data sufficient and stable, and does the value beat the honest alternative? Say out loud that "don't use AI" is a real outcome of this step, because that is what separates a considered answer from an enthusiastic one.
2. Value pattern — name which of the three the problem is: assisted decision making, scalability, or automation. This tells you where the human sits.
3. Technique — derive it from the shape of the required output, not from the industry or the data type, and say which shape you saw.
4. Build versus buy, and which kind of model — check whether a managed service already solves it before proposing a custom model, then apply the three constraints to choose traditional ML or a foundation model.

A strong answer names the constraint that decided each beat. A weak answer names services.

Why This Helps You

On the job: beat 1 is the conversation nobody has. Teams commit to a model for a problem where a rules table would have been exact, cheaper and auditable — and then spend a year maintaining it. Being the person who asks whether an exact answer is required is disproportionately valuable.

In interviews: "when would you not use machine learning?" is a standard senior screening question, and enthusiasm is the wrong answer. Naming the two disqualifiers and the correct cost-benefit baseline reads as experience.

On the exam: Task 1.2 questions are built to make the non-AI or managed-service answer feel too simple. Three of four options will be true statements about AWS. The habit of screening before selecting is what keeps you from picking a capable answer to a question that wanted a correct one.

Chapter Checklist

• I can name the three value patterns and pick the one a scenario spends its words on
• I can recognise both ways a scenario disqualifies AI/ML
• I can state why a highly accurate model is a defect for a legally defined calculation
• I can compare a model against the honest alternative rather than against nothing
• I can choose regression, classification or clustering from the output shape alone
• I can separate classification from clustering using whether the groups already have names
• I can map each Objective 1.2.4 application to its technique family, including both v1.1 additions
• I can name the direction each AWS managed AI service runs in, and its reverse where one exists
• I can say when SageMaker AI is the right answer and when it is a distractor
• I can apply Objective 1.2.6's three factors instead of choosing by task or data type
• I can explain why a fluent explanation is not explainability, and give the auditor test

After the Chapter

1. Complete student/project.md — parts 7, 8 and 9 of the AI/ML Decision Sheet you began in Chapter 01. Bring the same sheet; do not start a new one.
2. Take student/quiz.md closed-book, then review the reasoning for every question you guessed, including the ones you got right.
3. Open the official v1.1 exam guide's Domain 1 page and confirm you can attach a concept from this chapter to each of the six bullets under Task 1.2.
4. Next chapter: Chapter 03 — The AI/ML Lifecycle and MLOps (Domain 1, Task 1.3). The pipeline from business goal to monitoring and what each stage produces; where models come from and how they are served; what MLOps adds beyond "the model works"; and choosing between precision, recall, F1 and a business metric from the cost of being wrong. Chapter 03 completes Domain 1.

Chapter 03 — The AI/ML Lifecycle and MLOps

Certification Blueprint

Field	Coverage
Exam	AWS Certified AI Practitioner (AIF-C01), exam guide v1.1
Domain	Content Domain 1 — Fundamentals of AI and ML
Exam weight	20% of scored content
Task statement	1.3 Describe the AI/ML development lifecycle
Objectives	1.3.1 pipeline components · 1.3.2 FM sources · 1.3.3 production methods · 1.3.4 services per stage · 1.3.5 MLOps · 1.3.6 metrics

What This Chapter Covers

Chapters 01 and 02 stopped at the decision — what a system is, and whether AI is the right tool. This chapter is everything after the decision: the stages of building it, where models come from, how they are served, and what keeps them working once real traffic arrives.

Chapters 01-02 answered	Chapter 03 answers
What is this system?	What are the stages of building it?
Should we use AI at all?	Where do models come from?
Which technique and which service?	How is a model served in production?
—	What keeps it working after launch?

The practice focus is unusual. Task 1.3 questions describe a symptom — error climbing in production, a model that passed evaluation and failed live, a team that cannot reproduce last month's result — and the answer depends on locating when in the lifecycle it happened. The distractors are all real pipeline activities, performed at the wrong point.

This chapter completes Domain 1.

The Pipeline

What to remember from this diagram: the shape is a sequence, but Objective 1.3.1's verb is "describe and differentiate". Sequencing is half the objective. Separating the stages that sit next to each other is the other half, and it is the half that gets tested.

The diagram groups nine named activities into five stages so it stays readable. Learn the nine.

What each stage produces

Name the output, not the activity. The exam asks what a stage hands to the next one.

#	Stage	Produces
1	Data collection	A stored, accessible dataset
2	Exploratory data analysis (EDA)	Understanding — distributions, gaps, outliers
3	Data pre-processing	Clean, consistent, usable data
4	Feature engineering	Model-ready input variables
5	Model training	A candidate model
6	Hyperparameter tuning	A better-configured candidate model
7	Evaluation	A pass/fail verdict against the success metric
8	Deployment	A callable model in production
9	Monitoring	Drift, quality and cost signals

Two rows repay extra attention. EDA produces understanding, not a changed dataset — that is the whole distinction from pre-processing. And tuning produces a better-configured model, not a better-trained one: training learns parameters from data, while hyperparameters are settings chosen before learning that govern how it happens.

Differentiating adjacent stages

These four pairs are what the exam swaps. Each has a one-word separator.

Pair	Separator
EDA vs pre-processing	EDA looks; pre-processing changes
Pre-processing vs feature engineering	Pre-processing fixes what is wrong; feature engineering creates what is useful
Training vs hyperparameter tuning	Training learns the parameters; tuning sets the knobs that govern learning
Evaluation vs monitoring	Evaluation is before deployment, once; monitoring is after, continuously

Removing duplicate rows is pre-processing. Deriving "days since last purchase" from a timestamp is feature engineering. The last pair decides more questions than the other three combined: when a scenario describes a live production symptom and offers "evaluate the model", that option is on the wrong side of launch.

The Two Feedback Loops

What to remember from this diagram: a pipeline drawn as a straight line is wrong. Loop 1 fires before launch; loop 2 fires years after it.

Loop	From → to	Trigger
1	Evaluation → model development	The model fails the success bar
2	Monitoring → data preparation	Data drift in production

Loop 1 runs many times during the build and nobody outside the team sees it. Loop 2 is the one that gets tested, because it is the one teams forget to build.

Drift is not a bad model

The code has not changed. The model has not been redeployed. Error is climbing anyway.

Evidence	Reading
Code and model artefact unchanged	Rules out a deployment or algorithm defect
Error growing over time	Rules out a training bug — that would be wrong from day one
Input distribution has shifted	Data drift

The fix is loop 2: return to the data and re-train. Changing the algorithm is the intuitive move and it is wrong, because nothing in the evidence points at the algorithm.

Compare with Chapter 01's overfitting. Overfitting is high training score and low live score from the start. Drift is good performance that decays. Same symptom shape, different time signature, different fix — and any clause establishing a timeline is what separates them.

AWS Services for Each Stage

What to remember from this diagram: Objective 1.3.4 names five services explicitly — Bedrock, Amazon Q, Amazon Quick, Kiro, SageMaker AI — and expects you to know which stage each serves, not just what each does.

The scenario says…	Service	Stage
"Store and catalogue the raw training data"	S3, Lake Formation	Data collection
"Clean and normalise it without writing code"	AWS Glue DataBrew	Pre-processing
"Train and deploy our own model"	SageMaker AI	Training, deployment
"Start from a pre-built model we can adapt"	SageMaker JumpStart, Bedrock	Sourcing
"Call a foundation model through an API"	Amazon Bedrock	Deployment/serving
"Help our developers write the code"	Amazon Q, Kiro	Assisted development
"Build the dashboard the business reads"	Amazon Quick	Analysis and reporting
"Alert us when live quality degrades"	SageMaker Model Monitor, CloudWatch	Monitoring

Amazon Q and Kiro are assisted-development tools: they help the people building the pipeline rather than transforming the data flowing through it.

The wrong-stage distractor is this objective's whole strategy — a real service, correctly described, offered for a stage it does not serve. Amazon Quick genuinely builds dashboards; offering it for data pre-processing is not a false statement about the service, it is a false statement about the stage. Asking "which stage?" first is what makes it visible.

Where Models Come From

What to remember from this diagram: "train a custom model" is the answer far less often than learners expect. It requires labeled data and the skills to train, and the scenario has to say so.

Objective 1.3.2 is worded as sources of FM models — the sourcing question on this exam is framed around foundation models.

Source	Speed	Control	Use when
Open source pre-trained	Fastest	Least	Something already does the task well
Adapt a pre-trained model	Middle	Middle	Close but not exact; domain vocabulary differs
Train a custom model	Slowest	Most	Nothing existing fits, and you have data and skills

The trade-off is a straight line: speed and cost at one end, control and specificity at the other.

Serving a Model in Production

Objective 1.3.3 names exactly two methods.

Method	What it is	You give up	You gain
Managed API service	Call a model AWS operates, such as Bedrock	Control over the runtime and model internals	No servers, no scaling, no patching
Self-hosted API	Run the model on infrastructure you operate	Nothing about control	Every operational burden

The axis is operational burden versus control, and nothing else. Cost, performance and scale follow from the choice rather than driving it.

The scenario says…	Method	Why
"Small team, no ML operations staff"	Managed API	Operational burden is binding
"Get to production quickly"	Managed API	No infrastructure to build
"The model weights must stay on our own infrastructure"	Self-hosted	Control is mandated
"We need a specific model version pinned indefinitely"	Self-hosted	Managed services move underneath you
"Unpredictable traffic, no idle cost"	Managed API	Elasticity without capacity planning

"We want control" is not enough. Preference is not a constraint. The scenario must say why control is required — a regulator, a data residency rule, a contractual obligation. Without a stated reason, managed is the better answer.

MLOps

MLOps is what makes a model a maintained system rather than a successful experiment.

Three questions MLOps answers that a merely working model does not:

• Can we reproduce last month's result?
• Can we redeploy without a person remembering the steps?
• Will we know when it stops working?

A model that scores well and cannot answer these is a demo.

Objective 1.3.5 names exactly seven concepts, and you should be able to recite all seven.

Concept	What it means in practice
Experimentation	Trying variants with results that can be compared and found again
Repeatable processes	The same inputs produce the same model, without manual steps
Scalable systems	Training and serving grow with data and traffic
Managing technical debt	Paying down glue code, dead features and untracked datasets
Achieving production readiness	Testing, rollback, monitoring and ownership before launch
Model monitoring	Watching live quality, not just uptime
Model re-training	A defined trigger and process for refreshing the model

Technical debt earns its place because ML systems accumulate it faster than ordinary software. Every dataset, engineered feature, experiment and model version is a dependency, and unlike code, the data underneath them keeps changing.

Metrics

What to remember from this diagram: the first branch is the one people skip — is this a model question or a business question at all?

Objective 1.3.6 names accuracy, precision, recall and F1 score.

Metric	Optimise when	Failure it protects against
Precision	A false positive is expensive	False alarms
Recall	A false negative is dangerous	Misses
F1	Both matter; classes are imbalanced	—
Accuracy	Classes are balanced	⚠️ Misleading when imbalanced

On a dataset that is 99% one class, a model that always guesses that class scores 99% accuracy and has learned nothing. This is the single most testable fact in the objective.

Model metrics versus business metrics

Model metrics	Business metrics
Accuracy, precision, recall, F1	Cost per user, development costs, customer feedback, ROI
Tells you is the model good?	Tells you was it worth building?

Objective 1.3.6 names both families in the same bullet, deliberately. A model can be excellent and the project still a failure — a 97% accurate model that costs more to operate than the losses it prevents is a technical success and a business failure, and no confusion matrix reveals that.

Expect a question whose correct answer is a business metric while every distractor is a valid model metric. Those distractors are not false statements; they answer a question that was not asked.

Decision Rules and Exam Signals

Rule 1 — name the output, not the activity. Every pipeline stage is identified by what it hands to the next one.

Rule 2 — locate the stage before choosing the fix. Task 1.3 questions give a symptom. The distractors are real activities at the wrong point in the lifecycle.

Rule 3 — drift has a time signature. Unchanged code, plus error growing over time, plus shifted inputs. Overfitting is bad from day one; drift decays.

Rule 4 — ask "which stage?" before accepting a service. The distractor is a real service correctly described and wrongly placed.

Rule 5 — custom training needs data and skills, both stated. Otherwise it is a distractor.

Rule 6 — self-hosting needs a stated reason. Preference is not a constraint.

Rule 7 — check class balance before accepting accuracy. A stated imbalance means accuracy is wrong.

Rule 8 — ask what the question is about. Model quality, or project worth.

Distractor Patterns

Pattern	What it looks like	How to defuse it
Wrong-stage service	Amazon Quick offered for data pre-processing	Ask which stage the service serves
Drift read as a bad model	"Change the architecture" for a drift symptom	Code unchanged + error growing = drift
Evaluation/monitoring swap	"Evaluate the model" for a live production symptom	Evaluation is once, before launch
Custom model over-selected	Training from scratch where a pre-trained model fits	Does the scenario supply data and skills?
Accuracy on imbalanced data	Accuracy offered where one class is 0.2%	Ask for the class balance
Model metric for a business question	Precision offered for "did it pay for itself?"	Ask what the question is about
"We want control"	Self-hosted chosen on preference alone	The scenario must say why control is required

The first and the last separate a pass from a fail on this task statement.

Scenario Walkthrough

A retailer deployed a demand-forecasting model fourteen months ago. It passed evaluation at 94% and performed well for a year. Since a range refresh three months ago, forecast error has risen steadily. The code and the deployed model artefact are unchanged. The team asks whether it should switch to a different algorithm.

Evidence	Reading	Decision
Passed evaluation, performed well for a year	The model was correct for its original data	Not an algorithm defect
Code and artefact unchanged	Rules out deployment and code faults	Not a release problem
Error rising since a range refresh	The input distribution changed	Data drift
Error rising steadily, not from day one	Rules out overfitting	Confirms drift

The outcome is loop 2 — return to data preparation and re-train on data reflecting the new range. Changing the algorithm addresses a problem the scenario never described.

The load-bearing fact is "since a range refresh". The decoy is "94%", which invites an argument about model quality that has nothing to do with the failure.

Key Concepts

Term	Definition
Exploratory data analysis (EDA)	Examining data to understand its distributions, gaps and outliers; it produces understanding, not a changed dataset
Data pre-processing	Correcting what is wrong with data — missing values, inconsistencies, duplicates
Feature engineering	Creating useful model input variables from data that is already correct
Hyperparameter tuning	Searching for good values of the settings that govern how learning happens, chosen rather than learned
Evaluation	The pass/fail verdict against the success metric, performed once before deployment
Monitoring	Continuous observation of live quality, drift and cost after deployment
Data drift	Degrading production performance caused by the input distribution shifting away from the training data
Managed API service	Calling a model the cloud provider operates; minimum operational burden, less control
Self-hosted API	Running a model on infrastructure you operate; maximum control, every operational burden
MLOps	The practices that make a model a maintained system rather than a successful experiment
Technical debt (in ML)	Accumulated glue code, dead features and untracked datasets; accrues faster than in ordinary software because the data keeps changing
Model re-training	Refreshing a model on newer data against a defined trigger and process
Precision	The metric protecting against false alarms; optimise when a false positive is expensive
Recall	The metric protecting against misses; optimise when a false negative is dangerous
F1 score	The metric for when both error types matter and the classes are imbalanced
Business metric	A measure of whether the project was worth building — cost per user, development costs, customer feedback, ROI

Revision Flashcards

Say the answer aloud before revealing it.

1. What does EDA produce, and how does that separate it from pre-processing? → Understanding — distributions, gaps and outliers. EDA looks; pre-processing changes.

2. What separates pre-processing from feature engineering? → Pre-processing fixes what is wrong. Feature engineering creates what is useful from data that is already correct.

3. What separates training from hyperparameter tuning? → Training learns the model's parameters from data. Tuning sets the knobs that govern how learning happens; they are chosen, not learned.

4. What separates evaluation from monitoring? → Evaluation happens once, before deployment. Monitoring happens continuously, after it.

5. Name the two feedback loops and what triggers each. → Loop 1: evaluation back to model development, triggered by failing the success metric. Loop 2: monitoring back to data preparation, triggered by data drift.

6. Give the three-part evidence signature of data drift. → Code and model unchanged, error growing over time, and the input distribution has shifted.

7. How does drift differ from overfitting? → By time signature. Overfitting is bad from day one. Drift is good performance that decays.

8. Which five services does Objective 1.3.4 name explicitly? → Amazon Bedrock, Amazon Q, Amazon Quick, Kiro, and Amazon SageMaker AI.

9. What do Amazon Q and Kiro do in a pipeline context? → Assisted development — they help the people building the pipeline, not the data flowing through it.

10. When is training a custom model the right answer? → Only when nothing pre-trained fits and the scenario supplies both labeled data and the skills to train.

11. What is the only axis separating a managed API from a self-hosted one? → Operational burden versus control. Self-hosting requires a stated reason, not a preference.

12. Name the seven MLOps concepts. → Experimentation, repeatable processes, scalable systems, managing technical debt, achieving production readiness, model monitoring, and model re-training.

The Four-Beat Answer

The core question this chapter prepares you for: "A model that worked is now failing in production. How do you diagnose and fix it?"

Four beats, checked in this order. Missing a beat is a failure state.

1. Locate the stage — say where in the lifecycle the symptom lives. A live production symptom is monitoring, not evaluation, and naming that first rules out half the plausible fixes.
2. Read the time signature — was it wrong from day one, or did it decay? Bad from the start points at the model or its training. Decay points at the data.
3. Check what changed — code, model artefact, or input distribution. Unchanged code plus shifted inputs plus growing error is drift, and drift is a data problem, not an algorithm problem.
4. Name the loop and the fix — loop 2, back to data preparation, re-train on data reflecting the new reality, and add a re-training trigger so the next occurrence is caught by monitoring rather than by a customer.

A strong answer resists proposing a new algorithm. That is the move the question is testing for.

Why This Helps You

On the job: the most common failure in deployed ML is not a bad model, it is an unmonitored one. Teams ship at 94%, celebrate, and discover eighteen months later that nobody defined a re-training trigger. Being the person who asks "what fires loop 2, and who sees it?" before launch is worth more than a point of accuracy.

In interviews: "your model's performance is degrading in production — walk me through it" is a standard question, and reaching immediately for a different algorithm is the answer that ends the conversation. The four beats above are what an interviewer is checking for.

On the exam: Task 1.3 questions are symptom-shaped, and every distractor is a real activity performed at the wrong point in the lifecycle. Locating the stage before choosing the fix is what keeps you from picking a competent answer to the wrong problem.

Chapter Checklist

• I can sequence the pipeline and say what each stage produces
• I can separate EDA from pre-processing, pre-processing from feature engineering, training from tuning, and evaluation from monitoring
• I can draw both feedback loops and name what triggers each
• I can give the three-part evidence signature of data drift
• I can distinguish drift from overfitting by time signature
• I can attach Bedrock, Amazon Q, Amazon Quick, Kiro and SageMaker AI to their pipeline stages
• I can recognise a real service offered for the wrong stage
• I can choose a model source from speed versus control, and say what custom training requires
• I can choose managed or self-hosted, and say why a preference is not a constraint
• I can name all seven MLOps concepts and say what MLOps adds beyond "the model works"
• I can choose accuracy, precision, recall, F1 or a business metric from the cost of being wrong
• I can explain why accuracy is misleading on imbalanced data, with a worked number

After the Chapter

1. Complete student/project.md — parts 11, 12, 13 and 14 of the AI/ML Decision Sheet you began in Chapter 01. Bring the same sheet. These parts complete its Domain 1 coverage.
2. Take student/quiz.md closed-book, then review the reasoning for every question you guessed, including the ones you got right.
3. Open the official v1.1 exam guide's Domain 1 page and confirm you can attach a concept from this chapter to each of the six bullets under Task 1.3. Read the final bullet carefully and check it against any other study material you own — if a third-party source disagrees with the published guide, the published guide wins.
4. Next chapter: Chapter 04 — Foundation Models, Tokens, Embeddings, and the FM Lifecycle (Domain 2, 24%). Domain 2 is where v1.1 changed most, and the vocabulary load is higher than anything in Domain 1: tokens, chunking, embeddings, vectors, transformer-based LLMs, multi-modal and diffusion models, the GenAI use-case families, and the FM lifecycle from data selection through feedback.