CertifyDeckCargando contenido…
The official AWS Certified AI Business Strategist syllabus, organized by domain. Check off each point as you review it.
Inicia sesión para guardar tu progreso entre sesiones.
Describe core AI concepts and define terminology
During training, an algorithm learns the relationship between inputs and outputs from historical data, and the model stores what it learned in its parameters; at inference, the trained model receives new data and returns a prediction with a degree of confidence, not a certainty. The lifecycle starts with a measurable business goal and continues with data, development, deployment and monitoring.
AI is the umbrella term; ML is one of its branches, which learns patterns from data without explicit instructions, and not all AI is ML (rule-based systems exist too). Deep learning uses layered neural networks, and generative AI, an advanced form of deep learning, creates new content with pre-trained foundation models, although ML still performs better on many tasks.
Structured data fits a schema of rows and columns and is easy to query; unstructured data (audio, video, documents, sensor streams) takes more storage and needs preprocessing and ML to analyze. Semi-structured data blends both, such as videos with tags. The data type guides the model: basic ML for tables and deep learning for unstructured content.
Structured vs Unstructured Data - AWS ↗What is an AI Model? - AWS ↗
Incorrect, biased or incomplete data produces inaccurate predictions, both during training and when the model is used in production. Accuracy depends on reliable labels (ground truth) and on cleansing duplicates, formats and missing values; data preparation is often the longest phase. Data governance sets standards for accuracy, completeness and consistency and counters data silos.
What is Data Cleansing? - AWS ↗What is Data Governance? - AWS ↗
Supervised learning needs historical data labeled with the outcome, and unsupervised learning discovers groups in unlabeled data. An overfit model is accurate on its training data but does not generalize; separate validation data detects it. If historical data is not representative or carries bias, the model perpetuates it, and in production drift must be monitored and the model retrained.
What is Machine Learning? - AWS ↗What is Overfitting? - AWS ↗
Global standards provide a common vocabulary and framework that foster trust and interoperability. ISO/IEC 23053 describes AI systems that use machine learning, and ISO/IEC 42001 sets requirements for an AI management system (AIMS) that an independent third party can certify. A provider's certification does not certify its customers by association, although it eases their path.
ISO/IEC 42001 FAQs - AWS ↗AWS achieves ISO/IEC 42001:2023 accredited certification ↗
Identify and select appropriate AI solution types
Not every problem needs AI: when the result can be produced with rules, calculations or fixed steps, rule-based automation is cheaper, explainable and auditable. AI pays off when rules depend on too many overlapping factors, keep changing (as in fraud) or the input is unstructured text. The existing rule-based solution is the baseline to compare accuracy, response time and total cost; combining both approaches is often best.
MLCOST02-BP01 Identify if machine learning is the right solution - Machine Learning Lens ↗Orchestration models: From rule-based to AI-native - AWS Prescriptive Guidance ↗
An AI agent receives a goal from people and decides on its own which actions to take: it plans in steps, uses external tools and APIs, keeps context in memory and coordinates with other agents. A single agent suits a well-defined process; multi-agent systems use an orchestrator or supervisor that assigns work to specialists. Open protocols (MCP, A2A) reduce vendor dependency, and risky actions can be gated by human-in-the-loop approval.
A model trained on historical data loses performance when input data changes (data drift), when the relationship between inputs and outcomes changes (concept drift) or when bias appears over time (bias drift). That is why a baseline is set before production, monitoring is continuous with alerts on thresholds tied to business KPIs, and retraining happens only when performance falls below those thresholds, not on a calendar.
MLPERF06-BP04 Monitor, detect, and handle model performance degradation - Machine Learning Lens ↗MLSUS06-BP02 Retrain only when necessary - Machine Learning Lens ↗
Banning generative AI or adopting it slowly pushes employees to unmanaged consumer tools (shadow AI). The answer is twofold: provide approved tooling and gain visibility through usage and policy-violation monitoring. A published list of approved, restricted or under-evaluation tools, with a responsive review process, makes the governed path more attractive; new AI features in already-approved SaaS also need review.
Shadow generative AI - Navigating the security landscape of generative AI ↗Governance scope - AWS Prescriptive Guidance (Governing and architecting the diversity of agentic AI at scale) ↗
Apply GenAI concepts and techniques
Response quality depends on prompt quality. A good prompt combines instruction, context, examples and input text; it clearly defines the expected response and its format, and balances simplicity and detail. Examples (few-shot) steer style, step-by-step reasoning helps with complex problems, and the instruction works best at the end. It is iterative: test variants and measure them against a baseline.
What is Prompt Engineering? ↗Prompt engineering concepts - Amazon Bedrock ↗
Models process text as tokens (words, word parts or punctuation) and each model has a context limit. Very long documents, full conversation histories or entire tool catalogs can exceed that limit, raise the cost of every request and increase latency. Mitigations are splitting documents into chunks, summarizing older turns, including only relevant context and leaving a buffer for edge cases.
AGENTPERF03-BP02 Optimize context window utilization and prompt management - Agentic AI Lens ↗Tips for managing model token limits - Generative AI Application Builder on AWS ↗
Retrieval Augmented Generation (RAG) makes the model consult an authoritative knowledge base before answering: it brings in current documents without retraining, cites the source and reduces hallucinations, but it summarizes whole documents poorly. Fine-tuning adapts the model to a style or a new task, takes hours to days and usually needs specialist skills. The two techniques can be combined.
Comparing Retrieval Augmented Generation and fine-tuning ↗What is RAG (Retrieval-Augmented Generation)? ↗