Skip to content

Skill 1.1.1: Extract data from data sources

I notice there's a mismatch: the skill item provided is Skill 2.1.3: Compare and select appropriate ML models, generative AI (GenAI) models, algorithms, and solution templates to meet business needs, but the header says "Skill 1.1.1: Extract data from data sources."

I'll align with the actual skill item content you provided (2.1.3) and generate questions accordingly.

Skill 2.1.3: Compare and select appropriate ML models, generative AI (GenAI) models, algorithms, and solution templates to meet business needs

Question 1 - Foundation Model Selection vs. Custom Training

A financial services company needs to build an automated system that classifies customer emails into 12 internal categories (e.g., "loan inquiry," "fraud report," "account closure"). The company has approximately 50,000 labeled historical emails in Amazon S3. The ML team must decide between fine-tuning a foundation model from Amazon Bedrock or training a custom model using Amazon SageMaker AI. The system must achieve high classification accuracy, support explainable predictions for compliance audits, and minimize ongoing inference costs.

Which approach BEST satisfies all three requirements? (Select ONE)

  • A) Fine-tune an Amazon Bedrock foundation model on the 50,000 labeled emails using supervised fine-tuning.
  • B) Train a custom XGBoost classifier in Amazon SageMaker AI using TF-IDF text features, configured with SageMaker Clarify for feature attribution.
  • C) Use Amazon Bedrock with zero-shot prompting and few-shot examples in the prompt (no fine-tuning).
  • D) Use Amazon Comprehend's custom classification API to train a model on the labeled emails.
  • E) Train a multi-layer transformer from scratch using SageMaker distributed training on a large GPU cluster.
Click to reveal Answer & Explanation

Correct Answer: D

Correct Options Analysis

  • D is Correct: Amazon Comprehend custom classification is a managed AI service specifically designed for text categorization. It trains a custom classifier on labeled data, provides confidence scores, and requires no infrastructure management—offering the best balance of accuracy (trained on proprietary labels), explainability (confidence scores and word-level contributions), and low operational/inference cost (pay-per-use with no GPU hosting).

Distractor Analysis

  • A is Incorrect: Fine-tuning a foundation model on 50,000 labeled emails would achieve high accuracy but incurs significantly higher inference costs than Amazon Comprehend custom classification. Bedrock fine-tuned model inference is priced per token and requires provisioned throughput for consistent latency—making it far more expensive for high-volume text classification.
  • B is Incorrect: Training a custom XGBoost model is viable but introduces unnecessary operational overhead compared to Amazon Comprehend. The team would need to manage data preprocessing, feature engineering, hyperparameter tuning, model hosting, and endpoint scaling—work that Comprehend fully manages. SageMaker Clarify adds value for explainability, but Comprehend's built-in explanations suffice for this use case.
  • C is Incorrect: Zero-shot or few-shot prompting without fine-tuning generally underperforms a model trained on 50,000 labeled examples for a 12-class specialized classification task. Additionally, per-token foundation model inference costs would be substantially higher than Amazon Comprehend's classification API.
  • E is Incorrect: Training a transformer from scratch requires enormous amounts of data (millions of examples), GPU clusters, and months of research/engineering time. This is completely disproportionate to the task and violates the cost constraint.

Key Takeaway: When an AWS managed AI service already solves the task (text classification, entity extraction, OCR), it typically beats both custom ML models and foundation models on combined accuracy, explainability, and cost.

Question 2 - Prompt Engineering vs. Fine-Tuning

A healthcare startup is building an application that generates patient-facing summaries of clinical notes using Amazon Bedrock with a Llama 3 model. The summaries must use plain language and avoid medical jargon. Initial testing shows that the model produces overly clinical summaries despite a clear instruction in the prompt. The team has a dataset of 2,000 expert-written plain-language summaries paired with clinical notes.

What is the MOST cost-effective first step to close this quality gap? (Select ONE)

  • A) Immediately initiate supervised fine-tuning on the 2,000 paired examples using Amazon Bedrock.
  • B) Add 3–5 few-shot examples of clinical notes paired with plain-language summaries directly into the prompt and re-evaluate.
  • C) Switch to a larger, more expensive foundation model available in Amazon Bedrock.
  • D) Train a custom seq2seq model from scratch on the 2,000 examples.
  • E) Use Amazon Comprehend Medical to rewrite the summaries before presenting them to patients.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: The standard adaptation hierarchy is prompt engineering first, fine-tuning second. Adding few-shot examples (in-context learning) directly into the prompt immediately changes model behavior with zero training overhead, zero infrastructure cost, and no model versioning. With 2,000 high-quality examples available, selecting 3–5 representative ones for in-context learning is the most cost-effective first step.

Distractor Analysis

  • A is Incorrect: Supervised fine-tuning incurs significant overhead: data formatting (JSONL), training compute costs, model versioning, and evaluation. Per the AWS best practice hierarchy, fine-tuning should only be attempted after prompt engineering (including few-shot examples) has demonstrably failed to close the quality gap.
  • C is Incorrect: Switching to a larger model increases per-inference costs and may not address the stylistic gap. Larger models don't automatically produce plainer language; explicit examples and instructions in the prompt are more effective and cheaper.
  • D is Incorrect: Training a sequence-to-sequence model from scratch on only 2,000 examples is unrealistic—custom generative models typically require millions of examples and massive GPU resources. This option also creates unnecessary operational overhead.
  • E is Incorrect: Amazon Comprehend Medical extracts medical entities (conditions, medications, dosages) from text; it does not rewrite or simplify text for patient-facing audiences. This is a misapplication of the service.

Key Takeaway: Follow the "Prompt First, Fine-Tune Second" order: task-specific instructions + few-shot examples usually close stylistic/formatting gaps instantly at zero training cost.

Question 3 - RAG vs. Fine-Tuning for Knowledge Freshness

A legal technology company is building an assistant that answers questions about a client's corporate policies. The policy documents change multiple times per week, and the assistant must always cite the specific document and section behind each answer. The legal team cannot accept answers that cite outdated policies. The application uses Amazon Bedrock Knowledge Bases.

Which architecture BEST satisfies the citation and freshness requirements? (Select ONE)

  • A) Fine-tune a foundation model on the latest policy documents so the model internalizes the most recent knowledge.
  • B) Use Amazon Bedrock Knowledge Bases to retrieve relevant policy passages at request time and include source metadata in the prompt for citation.
  • C) Use Amazon Transcribe to convert all policy documents to audio and store them in S3 for fast retrieval.
  • D) Encode all policy documents into a single large prompt and submit it with every question using the model's full context window.
  • E) Train a custom document classification model to route questions to the appropriate policy document before querying a database.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: Retrieval-Augmented Generation (RAG) via Amazon Bedrock Knowledge Bases fetches fresh, relevant passages from a vector index at query time. Because documents are synced to the vector store as they change, answers reflect the newest policy versions. Each retrieved chunk carries source metadata (document title, section, page number), enabling the model to cite evidence in every response.

Distractor Analysis

  • A is Incorrect: Fine-tuning embeds knowledge into model weights. When policies change multiple times per week, the team would need continuous retraining—hours or days each time—to keep the model current. Fine-tuned models also cannot reliably cite the exact source document behind a specific answer, violating the citation requirement.
  • C is Incorrect: Amazon Transcribe converts speech to text (or creates text from audio); it is completely unrelated to document retrieval or citation generation. Converting documents to audio and storing them in S3 provides no retrieval or Q&A capability.
  • D is Incorrect: Stuffing all policy documents into the prompt on every request is expensive (token costs escalate linearly with document size), may exceed the model's context window as documents grow, and does not inherently provide clean citation prompts. It is a non-scaling workaround, not an appropriate architecture.
  • E is Incorrect: Training a custom routing classifier adds unnecessary complexity. Amazon Bedrock Knowledge Bases already manages semantic retrieval across documents without a custom routing model, which would require separate training, hosting, and maintenance.

Key Takeaway: RAG is the correct pattern when content changes frequently and citations are required; fine-tuning is better when behavior (style/terminology) is the gap, not knowledge.

Question 4 - Interpretability Requirements in Model Selection

A credit union needs to deploy a model that predicts whether to approve or reject loan applications. The model must meet regulatory compliance (the lending team must be able to explain exactly why each application was approved or denied), exhibit low inference latency (decisions within 50 ms), and minimize operational overhead. The team has 5 years of historical loan data with ~200 features.

Which TWO options satisfy the regulatory and latency requirements? (Select TWO)

  • A) Train an XGBoost model in Amazon SageMaker AI and use SageMaker Clarify to generate SHAP values for each prediction.
  • B) Use Amazon Bedrock with a large language model and include the applicant's details in the prompt.
  • C) Train a deep neural network with 10 hidden layers in SageMaker AI for maximum predictive accuracy.
  • D) Deploy a trained XGBoost model to a SageMaker real-time endpoint with auto-scaling configured.
  • E) Use Amazon Rekognition to analyze applicant ID photos for fraud detection signals.
Click to reveal Answer & Explanation

Correct Answers: A and D

Correct Options Analysis

  • A is Correct: XGBoost is a tree-based gradient boosting algorithm that provides excellent predictive performance on tabular data while remaining interpretable through feature importance scores and SHAP values. SageMaker Clarify provides ready-to-use SHAP explanations, satisfying the regulatory compliance requirement.
  • D is Correct: Deploying the trained XGBoost model to a SageMaker real-time endpoint with auto-scaling ensures sub-50 ms inference latency for individual predictions while minimizing operational overhead through managed infrastructure.

Distractor Analysis

  • B is Incorrect: Foundation models are not designed for binary classification on tabular loan data with 200 numerical features. They introduce high per-token inference costs, latency overhead, and—critically—cannot provide the precise feature-level explanations regulators require for lending decisions.
  • C is Incorrect: Deep neural networks are "black box" models. While SageMaker Clarify can generate approximate explanations, the interpretability is weaker than tree-based feature attribution, and the regulatory requirement for exact "why this applicant was denied" explanations is harder to satisfy. The added complexity also increases operational overhead.
  • E is Incorrect: Amazon Rekognition analyzes images and videos for content; it is not designed to evaluate loan applications or provide financial decision explainability. Using it for fraud signal extraction is a misapplication of the service.

Key Takeaway: When regulatory interpretability is mandatory, favor tree-based models (XGBoost, LightGBM) with SHAP feature attribution via SageMaker Clarify over deep learning or foundation models.

Question 5 - Managed AI Service vs. Custom Build

An insurance company receives thousands of scanned health insurance claim forms daily. These forms are semi-structured: layout varies by provider, but each contains standard fields like patient name, date of service, diagnosis codes, and billed amount. The company needs to extract these fields automatically, minimize development time, and avoid ongoing model maintenance.

Which approach delivers a working pipeline with the LEAST development effort? (Select ONE)

  • A) Train a custom computer vision model in Amazon SageMaker AI to detect form fields and apply OCR.
  • B) Use Amazon Textract's AnalyzeDocument API with the FORMS feature to extract key-value pairs and table data from the scanned forms.
  • C) Fine-tune a multimodal foundation model in Amazon Bedrock on past claim forms to perform key-value extraction.
  • D) Build an Apache Spark pipeline with AWS Glue to process each form manually and pattern-match the fields.
  • E) Use Amazon Comprehend's entity recognition to extract medical fields from the form images.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: Amazon Textract's AnalyzeDocument API with the FORMS feature is purpose-built for extracting key-value pairs, form fields, and structured table data from scanned documents. It is a fully managed pre-trained service requiring zero model training, zero infrastructure management, and no ongoing maintenance—delivering a working pipeline within hours, not weeks.

Distractor Analysis

  • A is Incorrect: Training a custom computer vision model requires bounding-box annotation workflows, OCR detector selection, specialized GPU training, and hosting infrastructure. This is a multi-week/multi-month effort with ongoing model maintenance—massive overkill when Amazon Textract already solves this exact problem.
  • C is Incorrect: Fine-tuning a multimodal foundation model is resource-intensive (dataset preparation, GPU provisioning, model versioning) and unnecessary when a managed service like Textract performs standard form extraction at lower cost and higher reliability.
  • D is Incorrect: Building a custom Spark pipeline in AWS Glue to pattern-match fields manually is brittle and high-maintenance. Every new form layout requires manual rules; this approach reinvents OCR and document understanding, which Textract already provides.
  • E is Incorrect: Amazon Comprehend performs NLP on raw text—it cannot process images or extract structured fields from scanned documents. The entity recognition API requires text as input, not image files.

Key Takeaway: When the task is a "solved problem" in AWS (document extraction, image analysis, speech-to-text), choose the managed AI service—it's the fastest and cheapest path with zero maintenance.

Question 6 - Foundation Model Selection Criteria

A media company is building a real-time chat summarization feature for customer support conversations. The summaries must be generated within 2 seconds and the application expects 10,000 summarization requests per day. The team has narrowed options to two Amazon Bedrock foundation models:

  • Model A: Larger, ~70B parameters, higher-quality summaries (4.5/5 average rating), slower inference (~4 seconds per request)
  • Model B: Smaller, ~7B parameters, good-quality summaries (4.⅖ average rating), faster inference (~1 second per request)

The company must also minimize cost per request. Which model should the team select, and why? (Select ONE)

  • A) Model A, because higher quality summaries justify the higher latency and cost.
  • B) Model B, because it meets the latency SLA at lower cost per request while maintaining acceptable quality.
  • C) Model A, but configure provisioned throughput to force lower latency.
  • D) Both models in parallel (A/B test), routing 50% of traffic to each.
  • E) Neither model—train a custom summarization model using SageMaker AI.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: The selection must optimize against the stated constraints: ≤2 seconds latency and minimum cost per request. Model B (1 second latency) meets the SLA, while Model A (4 seconds) fails it. Model B's lower parameter count also translates to lower per-token pricing in Bedrock. The quality differential (4.2 vs. 4.5) is not large enough to justify violating a hard latency requirement.

Distractor Analysis

  • A is Incorrect: Model A violates the 2-second latency requirement regardless of quality. In production systems, hard SLAs are non-negotiable; a 4-second response for a real-time chat summarization feature creates a poor user experience.
  • C is Incorrect: Provisioned throughput in Amazon Bedrock improves consistency and can reduce latency by eliminating cold starts, but it does not typically overcome the fundamental inference time of a 70B parameter model when a 1-second model exists. Provisioned throughput also significantly increases cost—directly contradicting the cost minimization requirement.
  • D is Incorrect: A/B testing is reasonable during evaluation but does not satisfy the production requirement. Routing 50% of traffic to Model A means half of all users experience violations of the latency SLA. This is an evaluation tactic, not a production decision.
  • E is Incorrect: Training a custom summarization model from scratch is entirely unnecessary. Foundation models in Bedrock already excel at summarization, and building a custom model would take months, require massive datasets, and create permanent maintenance overhead.

Key Takeaway: Foundation model selection is constraint-driven: identify the hard SLA (latency, cost, quality), filter models that fail the SLA, then optimize the remaining axes.

Question 7 - Fine-Tuning Strategy Selection

An e-commerce company wants to adapt a foundation model in Amazon Bedrock to generate product descriptions in a specific brand voice (playful, informal, emoji-friendly). The team has 500 high-quality example descriptions showing the exact writing style. The current model with prompt instructions produces descriptions that are "too formal and corporate."

Which fine-tuning strategy is MOST appropriate for this scenario? (Select ONE)

  • A) Reinforcement fine-tuning (RFT) with a reward function that scores brand-voice adherence.
  • B) Supervised fine-tuning (SFT) on the 500 labeled input-output pairs that demonstrate the target voice.
  • C) Model distillation from a larger model to a smaller one using the 500 examples as prompts.
  • D) Domain-adaptation fine-tuning (continued pre-training) on the company's entire product catalog.
  • E) No fine-tuning—simply increase the model's temperature setting to make outputs more creative.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: Supervised fine-tuning (SFT) is the correct strategy when the gap is stylistic/formatting and you have a labeled dataset of prompt-response pairs that demonstrate the target behavior. The 500 high-quality example descriptions serve as the ground truth for updating model weights to produce consistently playful, informal, emoji-friendly text. SFT is available in Amazon Bedrock for supported foundation models.

Distractor Analysis

  • A is Incorrect: Reinforcement fine-tuning (RFT) is best when the desired behavior can be scored but not easily demonstrated with labeled examples (e.g., "summaries should balance brevity with informativeness"). With 500 high-quality examples showing exactly what the output should look like, SFT is simpler, more direct, and more cost-effective than designing reward functions and running RLHF pipelines.
  • C is Incorrect: Model distillation compresses a larger teacher model into a smaller student to reduce latency/cost while preserving capability. It does not inherently teach a specific brand voice; distillation uses prompts to transfer general behavior, not to fine-tune a specific style on 500 labeled examples.
  • D is Incorrect: Domain-adaptation fine-tuning (continued pre-training) is used to inject broad domain knowledge (e.g., legal terminology, medical vocabulary) into a model before task-specific fine-tuning. It is not the right tool for adjusting writing style, and it requires substantially more data than 500 examples.
  • E is Incorrect: Increasing temperature makes outputs more random/diverse, not more aligned with a specific brand voice. High temperature often produces inconsistent or off-topic outputs rather than consistently playful descriptions. Temperature tuning is a prompt-level adjustment that has already been tried without success.

Key Takeaway: Supervised fine-tuning (SFT) is the correct path when you have high-quality labeled examples of the target behavior; reinforcement fine-tuning (RFT) is for cases where you can score outputs but not easily label them.

Question 8 - Model Selection with Cost and Training Time Constraints

A logistics company needs to predict package delivery delays using 5 years of historical shipment data (10 million records, 80 features). The model must be deployed within 2 weeks, achieve high predictive accuracy, and the team has limited ML expertise and budget. They must choose between a SageMaker built-in algorithm or a custom deep learning solution.

Which approach BEST satisfies the timeline, budget, and accuracy requirements? (Select ONE)

  • A) Train a custom transformer-based model using SageMaker distributed training on a p4d.24xlarge GPU cluster.
  • B) Use Amazon SageMaker AI's built-in XGBoost algorithm with automatic hyperparameter tuning (HyperparameterTuner).
  • C) Use Amazon Bedrock to fit a foundation model on the tabular delivery data in the prompt.
  • D) Build a K-means clustering model to segment shipments into delay/no-delay groups.
  • E) Train a custom LSTM recurrent neural network for time-series delay prediction.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: SageMaker built-in XGBoost is purpose-built for tabular classification problems with millions of rows and hundreds of features. It trains quickly (hours, not days), achieves state-of-the-art accuracy on tabular data, and requires minimal ML expertise to use. The SageMaker HyperparameterTuner automates hyperparameter search, further reducing the expertise barrier. This fits the 2-week deployment timeline and budget constraints perfectly.

Distractor Analysis

  • A is Incorrect: Training a custom transformer on tabular data is nontraditional and extremely expensive. GPU clusters (p4d instances) cost thousands of dollars per hour, and training a transformer on 10M records would likely exceed the 2-week timeline and budget. This is a "bleeding edge" approach with no accuracy advantage over XGBoost for tabular data.
  • C is Incorrect: Foundation models are not designed for precise tabular numerical prediction. Stuffing millions of rows of tabular data into a prompt is impossible (context window limits), costly (per-token pricing), and produces unreliable numerical forecasts. This is a misapplication of foundation model technology.
  • D is Incorrect: K-means is an unsupervised clustering algorithm for discovering natural groupings, not for supervised prediction of a known target variable (delay/no-delay). Using clustering for classification would yield poor accuracy and would not leverage the historical label data.
  • E is Incorrect: LSTMs are designed for sequential data (time-series, text), not for tabular data with 80 handcrafted features. Training custom LSTM models requires significant ML expertise, hyperparameter tuning, and typically delivers no accuracy improvement over gradient boosted trees on tabular problems.

Key Takeaway: For tabular data with structured features, gradient-boosted trees (XGBoost) almost always outperform deep learning while training faster and at a fraction of the cost.

Question 9 - RAG Architecture Selection

A pharmaceutical company wants to build a research assistant that answers scientists' questions using a corpus of 50,000 internal research papers. The answers must be grounded in citations, support complex multi-hop reasoning (e.g., "Which compounds studied in Paper A were also tested in Paper B?"), and handle documents up to 200 pages long. The system must use Amazon Bedrock Knowledge Bases.

Which RAG architecture pattern BEST aligns with these requirements? (Select ONE)

  • A) Basic semantic search RAG with standard chunking (fixed 500-token chunks).
  • B) Parent-document retrieval with metadata filtering enabled in the Knowledge Base.
  • C) Hierarchical chunking (document-level summary + section-level chunks) in the Knowledge Base, with a reranking step.
  • D) A custom solution that retrieves all 50,000 paper embeddings simultaneously on every query.
  • E) A purely generative approach—no retrieval—where the model answers from its training data.
Click to reveal Answer & Explanation

Correct Answer: C

Correct Options Analysis

  • C is Correct: Hierarchical chunking stores both document-level summaries and section-level chunks in the vector index, enabling initial retrieval at the document level followed by targeted section retrieval. Adding a reranking step (available in Bedrock Knowledge Bases) improves precision for multi-hop reasoning by re-scoring retrieved passages using a more powerful cross-encoder model. This supports 200-page documents—which would overflow simple chunking pipelines—by first navigating to the right document, then to the right section.

Distractor Analysis

  • A is Incorrect: Fixed 500-token chunking fragments 200-page documents into hundreds of context-isolated pieces. This approach cannot support multi-hop reasoning across documents (connecting compounds in Paper A to tests in Paper B) and often retrieves irrelevant chunks when documents are large. Reranking and document-level context are missing.
  • B is Incorrect: Parent-document retrieval helps by retrieving parent context around matching chunks, but for 200-page documents, the "parent" scope is still unclear. Metadata filtering supports retrieval by paper title, author, or year—not by relationships between compounds across papers. This doesn't fully address multi-hop reasoning.
  • D is Incorrect: Retrieving all 50,000 embeddings simultaneously on every query is computationally infeasible and would exceed context window limits. RAG relies on top-k nearest neighbor search; this option essentially abandons the retrieval mechanism.
  • E is Incorrect: A purely generative approach without retrieval cannot cite specific internal research papers. The model's training data doesn't include proprietary internal research, and it hallucinates heavily when asked about specific company documents. This option also violates the citation requirement.

Key Takeaway: Multi-hop reasoning across large corpora requires hierarchical retrieval (document-level navigation before section-level retrieval) and ideally a reranking step for precision.

Question 10 - GenAI Model Performance/Cost Tradeoff

A startup is building a customer support chatbot using Amazon Bedrock. The chatbot handles two distinct tasks:

  • Task 1: Classifying incoming messages into 8 intent categories (e.g., "billing," "technical support," "return request").
  • Task 2: Generating empathetic, context-aware responses to customer messages.

The startup has limited engineering resources and must minimize total monthly costs while maintaining acceptable quality for both tasks.

Which architecture BEST optimizes cost and quality? (Select ONE)

  • A) Use a single large foundation model for both tasks (the same Bedrock model classifies intent and generates responses).
  • B) Use Amazon Comprehend's custom classification API for Task 1 and a foundation model in Amazon Bedrock for Task 2.
  • C) Use a foundation model in Amazon Bedrock for Task 1 and Amazon Comprehend for Task 2.
  • D) Train a custom intent classification model in SageMaker for Task 1 and a custom response generation model in SageMaker for Task 2.
  • E) Use Amazon Rekognition to analyze message images and classify intent, then Bedrock for response generation.
Click to reveal Answer & Explanation

Correct Answer: B

Correct Options Analysis

  • B is Correct: This architecture splits the workload appropriately: (1) Amazon Comprehend custom classification handles the intent routing task at a fraction of foundation model costs (pay-per-request pricing for text classification vs. per-token pricing for generative inference). (2) Amazon Bedrock handles response generation, which inherently requires generative capability that Comprehend cannot provide. This optimizes cost by using a specialized, cheaper service for the classification task and a foundation model only for the generative task.

Distractor Analysis

  • A is Incorrect: Using a single large foundation model for both tasks means every intent classification call consumes expensive generative tokens. Classification is a simple structured-output task that Amazon Comprehend handles at a much lower cost. This approach violates the cost minimization requirement.
  • C is Incorrect: Amazon Comprehend cannot generate empathetic, context-aware responses—it performs NLP analysis like entity detection, sentiment analysis, and custom classification. Reversing the task assignment is a misapplication of services.
  • D is Incorrect: Training custom models for both tasks maximizes engineering effort and ongoing maintenance. A custom response generation model from scratch is realistically infeasible for a startup with limited resources (requires millions of examples, GPU training infrastructure, and months of development time).
  • E is Incorrect: Amazon Rekognition analyzes images and videos, not text messages. It cannot classify customer message intents. This is a misapplication of the service.

Key Takeaway: Split mixed traffic between managed AI services (cheap, task-specific) and foundation models (expensive, generative) to minimize total cost—use each service for what it is designed to do.