AWS Learning Material Certification Cheatsheets
AIF-C01 Domain 2 24% domain

Task 2.2: Understand the capabilities and limitations of generative AI for solving business problems

Fundamentals of GenAI · 3,719 words · source: Certified-AI-Practitioner-AIF-C01/domain-2/task-2-2-cheat-sheet.md

Domain 2: Fundamentals of GenAI

Task Statement 2.2: Understand the capabilities and limitations of generative AI for solving business problems

Domain 2 is 24% of scored exam content. Task 2.2 focuses on matching GenAI to business problems, recognizing where it creates value, understanding its risks and limitations, choosing the right model for the job, and measuring whether the solution is successful.

Use this sheet for scenario questions. The exam often asks whether GenAI is appropriate, what limitation explains a bad outcome, which factor should drive model selection, or which metric proves business value.

Exam Objective Map

You should be able to:

AWS exam objective What to know for the exam
Describe advantages of GenAI Adaptability, responsiveness, simplicity of interaction, content generation, personalization, summarization, productivity, faster ideation
Describe disadvantages of GenAI Hallucinations, inaccurate or biased responses, nondeterminism, lack of explainability, privacy/security risks, cost, latency, compliance concerns
Identify factors in model selection Modality, capability, performance, context length, accuracy, cost, latency, region availability, customization needs, responsible AI requirements
Determine business value metrics ROI, revenue, conversion rate, average revenue per user, customer lifetime value, efficiency, productivity, quality, customer satisfaction

The Big Picture

GenAI is strongest when a business problem involves language, knowledge, creativity, summarization, search, content transformation, or conversational workflows.

It is weaker when the business problem requires exact deterministic logic, guaranteed factual correctness without verification, fully explainable decision paths, real-time control without latency tolerance, or regulated decisions without strong governance.

Exam shortcut:

GenAI is useful for flexible content and knowledge tasks.
GenAI is risky for exact, factual, regulated, or deterministic decisions.
Use evaluation, grounding, guardrails, and human review to reduce risk.
Measure business value with outcome metrics, not only model metrics.

GenAI Capability Chooser

Business need GenAI capability AWS anchor
Summarize long documents, tickets, calls, or reports Text summarization and extraction Amazon Bedrock, Amazon Q, Amazon Transcribe + Bedrock
Answer questions from company knowledge Retrieval Augmented Generation (RAG) and semantic search Bedrock Knowledge Bases, Amazon Q Business
Draft emails, product descriptions, marketing copy, or documentation Text generation and rewriting Amazon Bedrock
Help developers write, review, or explain code Code generation and developer assistance Amazon Q Developer
Build a conversational assistant Natural language interaction with memory, tools, or workflows Amazon Bedrock Agents, Amazon Q, Amazon Lex
Generate images or creative variations Image generation or editing Bedrock image models, SageMaker AI
Classify, route, or extract from text using natural language Text understanding with prompting or FMs Amazon Bedrock, Amazon Comprehend
Search by meaning instead of keywords Embeddings and vector similarity Bedrock embeddings, OpenSearch vector search, Aurora pgvector
Automate multistep knowledge work Agentic workflow with tool use Bedrock Agents, AWS Lambda tools/APIs

Exam trap: if the requirement is simple, stable, deterministic, and rule-based, a traditional application rule or managed AI service can be better than a GenAI solution.

Advantages of GenAI

Advantage What it means Exam signal phrase
Adaptability One foundation model can support many tasks through prompting, RAG, or fine-tuning. "Use the same model for summarization, Q&A, and drafting"
Responsiveness Can generate near-real-time answers or content for interactive users. "Chat with customers or employees"
Natural language interface Users can express tasks in normal language instead of fixed forms or commands. "Ask questions in plain English"
Content generation Creates new text, code, images, audio, or synthetic examples. "Draft", "generate", "rewrite", "brainstorm"
Personalization Can adapt responses to user context, preferences, tone, role, or history. "Tailor response to each customer"
Summarization Compresses large content into key points, actions, or executive summaries. "Summarize support calls"
Semantic understanding Uses meaning rather than exact keyword matching. "Find related policies even if wording differs"
Productivity Reduces manual drafting, searching, triage, and repetitive knowledge-work effort. "Reduce time spent by analysts"
Scalability Can serve many users through managed model APIs and application integration. "Make expertise available across the organization"

Memory hook:

GenAI helps with content, conversation, context, and creativity.

Disadvantages and Limitations

Limitation Exam-ready meaning Mitigation direction
Hallucination Model can produce plausible but false or unsupported information. Ground with RAG, cite sources, evaluate, add human review
Inaccuracy Output can be wrong, incomplete, outdated, or misaligned with the task. Use tests, validation, structured outputs, confidence thresholds
Nondeterminism Same or similar prompts can produce different outputs. Lower temperature, constrain format, add validation, avoid deterministic-only use cases
Lack of explainability Model reasoning may not be transparent or fully auditable. Use explainability where possible, keep logs, require review for high-impact decisions
Bias and fairness risk Outputs can reflect bias from training data, prompts, or retrieved content. Evaluate across user groups, improve data, use responsible AI controls
Privacy risk Sensitive data in prompts, logs, training data, or retrieved context can be exposed. Data classification, access controls, encryption, redaction, private connectivity
Security risk Prompt injection, data exfiltration, unsafe tool use, and malicious inputs can affect behavior. Guardrails, least privilege tools, input/output filtering, monitoring
IP and copyright risk Generated content may raise ownership, attribution, or licensing concerns. Legal review, content policies, source controls
Cost Token usage, model size, output length, and throughput affect cost. Choose right-sized model, optimize prompts, cache, summarize, monitor usage
Latency Larger models and long contexts can slow responses. Use smaller models, shorter prompts, streaming, async patterns
Context-window limits Models can only process a limited amount of text/input per request. Chunking, retrieval, summarization, context selection
Domain knowledge gaps A model may not know private, recent, or specialized information. RAG, knowledge bases, fine-tuning, data updates

Exam shortcut:

Hallucination = sounds right but may be false.
Nondeterminism = output can vary.
Explainability = hard to know why the model responded that way.
Grounding/RAG = add trusted context at inference time.
Guardrails = apply safety and policy controls.

When GenAI Is a Good Fit

Good-fit signal Why
The task uses natural language, images, audio, or other unstructured data FMs are strong at unstructured data understanding and generation.
There are many valid answer styles GenAI can produce flexible outputs.
Human review is acceptable for high-impact output Review reduces risk from errors or hallucinations.
The goal is productivity, drafting, summarization, or search These map directly to common GenAI strengths.
The solution benefits from conversational interaction Users can ask follow-up questions and refine intent.
The knowledge source can be grounded RAG can provide private or current information.
The organization can evaluate and monitor quality GenAI needs ongoing measurement and governance.

When GenAI Is a Poor Fit

Poor-fit signal Better direction
Exact arithmetic, deterministic logic, or strict rule execution is required Use application code, rules engines, databases, or calculators.
Every answer must be guaranteed correct without verification Use authoritative systems, validation, and human approval.
The decision is high-impact and must be fully explainable Use governed ML/statistical models, explainability, and review workflows.
The data is highly sensitive and cannot be sent to the chosen model/service Use private deployment patterns, redaction, or different architecture.
Latency must be extremely low and predictable Use smaller models, cached answers, traditional ML, or precomputed outputs.
The task is already solved well by a cheaper managed AI service Use the purpose-built service.
The process requires only simple keyword search Use traditional search rather than embeddings/GenAI.

Exam trap: GenAI can support business decisions, but for regulated or high-impact decisions it usually needs governance, validation, auditability, and human oversight.

Model Selection Factors

Choose the model that meets the business requirement with the lowest acceptable complexity, risk, latency, and cost.

Factor What to ask Exam signal
Modality Does the task require text, image, audio, video, embeddings, or multimodal input/output? "Analyze images", "generate text", "semantic search"
Capability Can the model perform the task well: reasoning, summarization, coding, extraction, multilingual, image generation? "Need strong code generation"
Accuracy / quality Does it meet quality targets on representative examples? "Highest answer correctness"
Context window Can the prompt, history, retrieved chunks, and expected output fit? "Long documents"
Latency Can it respond within the user or system SLA? "Interactive chat under 2 seconds"
Cost Are input/output tokens, image generation, throughput, and customization costs acceptable? "Minimize cost at scale"
Throughput and quotas Can it handle expected request volume? "Thousands of requests per minute"
Region availability Is the model available in the required AWS Region? "Data must stay in Region"
Customization Does the use case need prompting, RAG, fine-tuning, continued pre-training, or a custom model? "Company-specific style or domain terms"
Integration Does it support required APIs, tools, agents, streaming, structured output, or knowledge bases? "Need tool use and workflow actions"
Security and compliance Does it meet privacy, encryption, access control, retention, and compliance requirements? "Sensitive customer records"
Responsible AI controls Can harmful, unsafe, biased, or off-policy outputs be detected and blocked? "Prevent toxic or unsafe answers"

Model-size tradeoff:

Option Strength Tradeoff
Larger / more capable model Better reasoning, instruction following, and complex task handling Higher cost, higher latency
Smaller / faster model Lower cost and latency, good for simple repetitive tasks Lower performance on complex reasoning
Embedding model Best for semantic search, similarity, clustering, and retrieval Does not generate final natural-language answers by itself
Image or multimodal model Best for visual input/output tasks Not ideal for text-only workloads if text model is sufficient

Exam shortcut:

Pick by modality first.
Then validate quality.
Then optimize cost, latency, region, compliance, and integration.

Adaptation Strategy for Business Problems

Need Best first approach Why
Improve answer format, tone, or instructions Prompt engineering Cheapest and fastest
Give private or current facts RAG / knowledge base Adds information without changing model weights
Improve semantic search Embeddings + vector store Retrieves by meaning
Teach consistent domain style or task behavior Fine-tuning Changes model behavior with examples
Build from massive proprietary corpus Continued pre-training or custom training Highest cost and complexity
Enforce safety or topic policy Guardrails and application controls Reduces harmful or off-policy outputs
Execute actions Agent with tool/API permissions Lets model plan and call approved tools

Exam traps:

  • Use RAG when the model needs private, current, or source-grounded knowledge.
  • Use fine-tuning when the model needs better behavior, style, or task pattern.
  • Fine-tuning is not the default fix for missing knowledge.
  • Prompt engineering and RAG do not change model weights.

Evaluation: Model Metrics vs Business Metrics

Do not confuse model quality metrics with business value metrics.

Metric type Measures Examples
Model quality Whether the model output is good enough Accuracy, relevance, completeness, groundedness, faithfulness, toxicity, robustness
System performance Whether the application works operationally Latency, throughput, error rate, uptime, cost per request
Business value Whether the solution improves the business outcome ROI, efficiency, conversion rate, revenue, average revenue per user, customer lifetime value
User experience Whether users trust and adopt it CSAT, NPS, adoption rate, task completion, deflection rate
Risk and governance Whether the solution stays within policy Hallucination rate, unsafe-output rate, policy violations, human override rate

Amazon Bedrock evaluation methods can include automatic evaluation and human evaluation, depending on the use case. For exam purposes, remember that GenAI solutions should be evaluated on representative prompts and expected business outcomes, not only on generic benchmarks.

Business Value Metrics

Metric What it tells you Example exam wording
ROI Whether financial benefit exceeds cost. "Did the assistant produce more value than it costs?"
Efficiency / productivity Time, effort, or cost saved. "Reduce manual review time by 40%"
Conversion rate Percentage of users who complete a desired action. "More shoppers complete checkout after using recommendations"
Average revenue per user (ARPU) Revenue generated per user over a period. "Increase monthly revenue per active user"
Customer lifetime value (CLV) Expected long-term value of a customer relationship. "Improve retention through personalized support"
Customer satisfaction (CSAT) User satisfaction with the experience. "Support users rate answers higher"
Deflection rate Share of support cases resolved without human agent escalation. "Virtual assistant handles common questions"
Average handle time (AHT) Time required to resolve a customer interaction. "Summaries help agents close cases faster"
First contact resolution Issue resolved in one interaction. "Agent has enough context to answer correctly"
Cost per task / request Operational cost for each completed unit of work. "Token usage makes each answer too expensive"

Business value formula:

Value = better outcomes - total cost - added risk

For the exam, the "best" metric depends on the goal:

Goal Best metric family
Save employee time Efficiency, productivity, cost reduction
Improve sales Conversion rate, revenue, ARPU
Improve customer support CSAT, deflection rate, AHT, first contact resolution
Improve retention CLV, churn rate
Control operating cost Cost per request, token usage, infrastructure cost
Reduce risk Policy violations, hallucination rate, human override rate

Common Scenario Patterns

Scenario Best answer
"The chatbot invents a refund policy." Hallucination; use RAG/grounding, source citations, guardrails, and review.
"The company wants answers from internal HR documents." RAG or Amazon Q Business/Bedrock Knowledge Bases.
"The output must always follow an exact JSON schema." Prompt constraints plus validation; do not trust free-form output alone.
"A legal decision must be fully auditable and explainable." GenAI alone is risky; use governance, review, and explainable decision process.
"The assistant is too expensive at high volume." Use smaller model, shorter context, output limits, caching, batching, or task routing.
"The response is too slow for chat." Reduce context, use lower-latency model, stream output, or precompute common answers.
"The model lacks company-specific facts." Use RAG, knowledge base, or controlled data source.
"The model tone is wrong but facts are available." Improve prompt or fine-tune with examples if prompting is insufficient.
"The system needs image input plus text questions." Use a multimodal model.
"The system only needs semantic document retrieval." Use embeddings and vector search; generation may be optional.

Risk Controls for GenAI Business Systems

Risk Control
False answers Grounding, citations, retrieval filters, answer abstention, human review
Unsafe content Guardrails, content filters, denied topics, output moderation
Prompt injection Treat retrieved/user content as untrusted, separate instructions from data, tool permission limits
Data leakage IAM, encryption, network controls, redaction, least privilege, logging controls
Bad tool actions Narrow tool scopes, approval workflows, validation, idempotency, audit logs
Bias or unfairness Representative evaluation, bias testing, human review, monitoring
Cost overruns Token budgets, output limits, model routing, quotas, usage alerts
Quality drift Continuous monitoring, feedback loops, regression test prompt sets

Exam shortcut:

High-risk use case = stronger governance.
Sensitive data = stronger access control.
External users = stronger guardrails.
Tool use = least privilege and validation.

AWS Service Anchors for Task 2.2

Need AWS service or feature
Access foundation models through an API Amazon Bedrock
Compare and evaluate foundation models Amazon Bedrock model evaluation
Build RAG over enterprise data Amazon Bedrock Knowledge Bases, Amazon Q Business
Add safety controls to GenAI applications Guardrails for Amazon Bedrock
Build agents that call tools/APIs Amazon Bedrock Agents
Build, train, tune, or deploy custom ML models Amazon SageMaker AI
Detect bias in ML workflows Amazon SageMaker Clarify
Purpose-built NLP Amazon Comprehend
Speech-to-text or text-to-speech Amazon Transcribe, Amazon Polly
Translation Amazon Translate
Conversational voice/text bot Amazon Lex

Exam trap: Bedrock is the main AWS service anchor for foundation models and GenAI apps, but purpose-built AI services can be better for narrow tasks like transcription, translation, text-to-speech, or entity/sentiment analysis.

Fast Comparison Drills

Prompt Correct association
"Generate a personalized email for every customer segment." GenAI advantage: personalization and content generation
"Summarize thousands of product reviews into key themes." GenAI advantage: summarization
"The model produces a confident answer that is not in the source documents." Hallucination / poor grounding
"The same prompt produces slightly different outputs." Nondeterminism
"The company cannot explain exactly why the output was generated." Lack of explainability
"The task uses images and text together." Multimodal model selection
"The task requires semantic search over documents." Embedding model + vector store
"The task requires strict exact calculations." Use deterministic code/tools, not GenAI alone
"The main goal is to reduce manual processing time." Efficiency/productivity metric
"The main goal is to increase purchases." Conversion rate/revenue metric
"The main goal is to increase long-term customer value." Customer lifetime value
"The model must stay within company safety policy." Guardrails and output controls

One-Page Memorization Version

  • GenAI is useful for flexible content, conversation, summarization, semantic search, personalization, and knowledge-work productivity.
  • GenAI is risky when answers must be deterministic, fully explainable, guaranteed correct, or used for high-impact regulated decisions without oversight.
  • Hallucinations are plausible false outputs.
  • Nondeterminism means outputs can vary.
  • Lack of explainability means the reasoning path is hard to audit.
  • RAG grounds answers in trusted data without changing model weights.
  • Fine-tuning changes behavior or style with examples.
  • Choose models by modality, capability, quality, context length, latency, cost, region, compliance, customization, and integration needs.
  • Bigger models are not automatically better; they can cost more and respond more slowly.
  • Business value metrics include ROI, efficiency, conversion rate, revenue, ARPU, CLV, CSAT, deflection rate, and cost per task.
  • Model metrics prove output quality; business metrics prove business value.
  • High-risk GenAI requires evaluation, monitoring, guardrails, access control, and often human review.

Mini Practice Questions

  1. A company wants a chatbot to answer employee HR questions from internal policy documents. The model sometimes invents policy details. What should the company add?

    • Answer: Use RAG/knowledge grounding from approved HR documents, add source citations, evaluate answers, and use guardrails or human escalation for sensitive topics.

  2. A retail company uses GenAI to write personalized product descriptions. Which metric best measures whether it increases sales?

    • Answer: Conversion rate, revenue, or average revenue per user depending on the stated goal.

  3. A finance workflow requires exact calculations and a fully auditable formula. Should GenAI generate the final answer directly?

    • Answer: No. Use deterministic calculation logic or tools; GenAI can help explain or draft text around verified results.

  4. A support assistant is accurate but too expensive at scale. What should be considered?

    • Answer: Use a smaller or lower-cost model, reduce prompt/context size, limit output tokens, cache common answers, route simple tasks to cheaper models, and monitor token usage.

  5. A solution must process both product images and customer text questions. What model-selection factor matters first?

    • Answer: Modality; choose a multimodal model that supports image and text input.

  6. A model gives different but acceptable versions of a marketing paragraph for the same prompt. Which limitation does this demonstrate?

    • Answer: Nondeterminism. This may be acceptable for creative drafting but not for deterministic decisions.

Source Anchors