AIF-C01 / AWS Certified AI Practitioner1 / 2 · Guide 1.1
65 questions · 90 minutes · 50 scored + 15 unscored · Pass 700/1000 · No penalty for guessing

1.1 / AI & ML foundations · 20% domain

AI → ML → DLAI is the broad field; ML learns from data; deep learning uses multilayer neural networks.
SupervisedKnown labels; classification predicts categories; regression predicts numbers.
UnsupervisedNo target labels; clustering discovers groups; anomaly detection spots unusual behavior.
ReinforcementActions → environment → rewards; optimize cumulative reward.
Training / inferenceLearn model parameters / apply the trained model to new inputs.
DataTabular = rows/columns; time series = ordered observations; text, images and audio often unstructured.

1.1 / Choose an inference mode

Real-timeSustained requests; predictable low latency; hosted endpoint.
ServerlessIntermittent synchronous calls; no instance management; tolerate cold starts.
AsynchronousQueue large/long requests; retrieve results later from S3; can scale to zero.
BatchWhole dataset available upfront; offline results; no persistent endpoint.
Deciding detailAsync and serverless both reduce idle costs; output timing and payload distinguish them.

1.2 / Managed service selection

Transcribe / PollySpeech → text / text → speech.
TranslateText between languages; custom terminology for selected terms.
ComprehendEntities, sentiment, key phrases and supported PII in text.
TextractScanned documents → text, forms, tables.
RekognitionImages/video → labels, faces, text and moderation.
LexConversational interface; intents and slots.
PersonalizePersonal recommendations from interaction data.
Rules vs MLExact published rules → deterministic logic; learned predictions → ML; open-ended generation → FM.

1.3 / ML lifecycle & tools

LifecycleBusiness goal → data → preparation → training → evaluation → deployment → monitoring; iterate.
Ground TruthTraining labels and human labeling workforce.
Data WranglerVisual data preparation and quality insights.
Feature StoreReusable features for training and inference.
Pipelines / RegistryWorkflow orchestration / model versions and approvals.
Model MonitorProduction data/model quality; supported bias and attribution drift.
JumpStartDiscover, deploy and customize supported pretrained models.
Split dataTrain learns; validation selects/tunes; holdout test gives final assessment; avoid leakage.

2.1 / GenAI foundations · 24% domain

Foundation modelBroad pretraining; adaptable to multiple downstream tasks.
TokensText pieces; input and output token counts affect cost and context limits.
EmbeddingsNumeric representations for semantic similarity and retrieval.
MultimodalSupported combinations of text, images, audio or video vary by model.
HallucinationPlausible unsupported content; grounding reduces risk, never guarantees truth.
Agentic AIGoal-directed planning and tool use with observations and iteration.
Short/long memoryWithin-session context / remembered information across sessions.

2.1 / AgentCore & agent building

RuntimeManaged hosting and isolated agent sessions.
MemoryShort-term events and long-term extracted context.
GatewayExpose supported APIs/Lambda as tools through MCP.
IdentityAgent and user credentials for service access.
PolicyDeterministic authorization of tool calls.
ObservabilityTraces and operational metrics.
Code / BrowserIsolated code execution / browser interaction.
Strands / A2AAgent SDK / protocol for agent-to-agent communication.

2.2 / Value, limitations & selection

Fit the taskModality, latency, context window, quality, cost, license, privacy and Region requirements.
LimitationsHallucinations, bias, cutoff, variable output and limited interpretability.
Business valueTask completion, customer satisfaction, cost per interaction and conversion rate.
Human reviewUse for high-impact or low-confidence outputs; define escalation and accountability.
Small vs largeSmaller can reduce cost/latency; measure quality on representative tasks before choosing.

2.3 / AWS GenAI technologies

BedrockManaged access to FMs and application building capabilities.
SageMaker AICustom ML training, deployment and lifecycle control.
NovaAWS model family; match model capabilities to modality and task.
Kiro / TransformAgentic development / modernization of existing workloads.
Amazon QuickAI-assisted business analysis and insights.
On-demandVariable workload; usage-based inference without provisioned capacity commitment.
ProvisionedReserved throughput for suitable steady workloads; model support matters.
Batch / cachingDelay-tolerant bulk inference / reuse eligible repeated prompt prefixes.

EXAM TRAPS / First-pass checks

Service namesChoose by required capability, not the newest or most powerful name.
No guaranteeLow temperature, RAG and guardrails do not guarantee correct answers.
Shared dutyManaged infrastructure does not remove customer access, data and application responsibilities.
Recent changesLearn guide objectives; avoid making availability notices or dates the deciding fact.
Reported emphasisResponsible AI, service selection and governance lead research signals; anecdotes are not official weights.
AIF-C01 / AWS Certified AI Practitioner2 / 2 · Guide 1.1
65 questions · 90 minutes · 50 scored + 15 unscored · Pass 700/1000 · No penalty for guessing

3.1 / FM applications · 28% domain

RAG ingestionPrepare/chunk documents → embed → index/store → sync changes.
RAG inferenceQuery → retrieve relevant context → augment prompt → generate with citations.
Vector searchSemantic similarity; hybrid combines semantic and keyword retrieval.
Metadata filtersRestrict retrieval by document attributes before or during search.
RerankingReorder retrieved candidates for relevance; not model fine-tuning.
Access controlsProtect both source retrieval and output; citations do not grant access permission.
AgentsUse tools for actions; knowledge retrieval alone is not autonomous execution.

3.2 / Prompt engineering

Prompt partsInstruction, context, input, output format; delimit trusted instructions from untrusted data.
Zero / one / fewNo examples / one example / several examples in context.
Chain of thoughtStructured reasoning technique; do not confuse prompt instructions with retraining.
ChainingSequence prompts and pass intermediate results.
TemperatureLower usually more predictable; higher more variable.
Top K / top PTop K candidates / cumulative probability nucleus.
Max tokens / stopOutput length limit / sequences that end generation.
InjectionDirect user instruction or indirect retrieved content tries to override intended behavior.

3.3 / Customization choices

PromptingFast task instructions; no weight updates.
RAGCurrent/private factual context at inference; no weight updates from retrieval.
Fine-tuningAdapt behavior with training examples; changes model weights.
Continued pretrainAdditional domain corpus; adapt domain language and patterns.
DistillationLarge teacher → smaller student; trade quality against cost/latency.
LoRAParameter-efficient adaptation; train a smaller set of parameters.
RLHF / DPOPreference alignment from human or preference data.
Data qualityRepresentative, permitted, clean data; separate train/validation/test and evaluate regressions.

3.4 / Evaluation metrics

PrecisionTP / (TP + FP); minimize false alarms.
RecallTP / (TP + FN); minimize missed positives.
F1 / accuracyBalance precision + recall / overall correct fraction; imbalance can mislead.
MAE / RMSEAbsolute numeric error / stronger penalty for large errors.
BLEU / ROUGETranslation-oriented n-gram precision / summary-reference overlap.
BERTScoreContextual embedding similarity; not factuality proof.
PerplexityNext-token predictive uncertainty; compare under compatible evaluation conditions.
Human / LLM judgeHuman expertise / scalable rubric scoring; calibrate and check judge bias.

3.4 / Evaluate the entire application

Context relevanceAre retrieved passages relevant to the query?
Context coverageDo passages contain the required reference information?
FaithfulnessAre answer claims supported by retrieved passages?
CorrectnessIs the answer accurate? A faithful answer can repeat an incorrect source.
CompletenessDoes it address all parts of the request?
AgentsTool selection, argument accuracy, task/goal success, harmfulness.
BusinessMeasure cost, latency, satisfaction and completed user tasks alongside model quality.

4.1 / Responsible AI · 14% domain

FairnessAvoid harmful group disparities; evaluate representative subgroups.
ExplainabilityUnderstand why the system produced an outcome.
Privacy & securityProtect personal information and system/data access.
SafetyReduce harmful outputs and consequences.
ControllabilityMonitor, guide and intervene in system behavior.
Veracity / robustnessTruthfulness and reliable behavior under variation or adversity.
GovernanceOwnership, policies, review and accountability across the lifecycle.
TransparencyDisclose intended use, limitations and AI involvement.
Clarify / A2IBias/feature attribution / human review of supported predictions.

4.1 / Guardrails distinctions

Content filtersHarmful content categories and supported prompt attacks.
Denied topicsBlock topic categories, not merely literal words.
Word filtersBlock configured terms and supported profanity.
Sensitive infoDetect/block/mask supported PII and configured patterns.
GroundingCheck answer grounding and relevance to supplied context.
Automated ReasoningValidate against formal policy rules; findings need application handling.
Layer defensesGuardrails complement IAM, tool authorization, retrieval permissions and evaluation.

4.2 / Explainability & documentation

InterpretabilityUnderstand internal workings directly.
Local / globalExplain one prediction / aggregate model behavior.
SHAPFeature attributions; association is not proof of causation.
Model CardsYour model: intended use, risk, training, evaluation.
AI Service CardsAWS-published service uses, limitations and responsible design.
TradeoffsExplainability vs privacy; transparency vs security/IP; model complexity vs interpretability.
Human-centeredPresent useful explanations to the actual decision-maker; keep escalation paths.

5.1 & 5.2 / Security & governance · 14%

IAM / KMSLeast-privilege resource access / encryption key control.
PrivateLinkPrivate service connectivity; separate from encryption and authorization.
Macie / InspectorSensitive data in S3 / supported software vulnerabilities.
CloudTrailWho called which AWS API and when.
CloudWatchOperational logs, metrics and alarms.
Config / ArtifactResource compliance rules / AWS compliance reports.
PseudonymizeRe-identification mapping remains; protect separately.
Anonymize / DPIrreversible de-identification / calibrated privacy noise.
GovernanceData lineage, approvals, retention, residency, access and incident response.
ISO 42001 / NISTAI management-system standard / voluntary AI risk framework.
Security scopesConsumer app → enterprise app → pretrained → fine-tuned → self-trained; increasing customer ownership.
FINAL TRAP700 scaled ≠ 70% raw. Review all unanswered items; no guessing penalty.