2.1 AI, ML, Deep Learning, and Core Terminology

The Vocabulary Stack

Artificial intelligence is the broad idea of computer systems performing tasks that usually require human judgment, perception, language understanding, planning, or decision support. In a business conversation, AI might mean a chatbot, a document classifier, a fraud score, a recommendation list, or a visual inspection workflow. The term is useful, but it is too broad for choosing an AWS service or approving a project.

Machine learning is a narrower approach inside AI. Instead of coding every rule by hand, a model learns patterns from data and then applies those patterns to new inputs. A team might train a model on historical support tickets to classify future tickets, or use historical transactions to estimate fraud risk. The business question is whether the past data is reliable enough to guide future decisions.

Deep learning is a family of ML methods that uses neural networks with many layers. Deep learning is common in image recognition, speech processing, natural language processing, translation, and foundation models. A practitioner does not need to calculate model weights for this exam, but should understand that deep learning usually needs more data, compute, monitoring, and cost awareness than a simple rules engine.

The model is not the whole solution. A useful AI workflow also needs input data, identity and access controls, logging, cost controls, user experience, fallback handling, and a way to measure whether the output helps. For AWS, that may involve IAM, S3, CloudWatch, CloudTrail, KMS, and business analytics alongside AI services. A non-builder still needs to ask who owns each piece.

Managed AI services hide much of the ML complexity. Amazon Rekognition can analyze images and video, Amazon Textract can extract text and structure from documents, Amazon Transcribe can convert speech to text, Amazon Translate can translate text, Amazon Polly can synthesize speech, and Amazon Comprehend can extract language insights. These are strong choices when the task matches the service boundary and the business does not need to own model training.

Foundation model services are different. Amazon Bedrock provides access to foundation models for generative AI tasks such as summarization, drafting, question answering, embeddings, and agent workflows. Amazon Q provides role and application oriented generative AI experiences. These services are useful when language or multimodal reasoning is central, but they introduce hallucination, prompt injection, content safety, and data governance questions.

Custom ML is a better fit when the organization has a specific prediction problem, sufficient relevant data, a measurable objective, and a need for control over model behavior. Amazon SageMaker AI supports the builder path for training, deployment, and ML lifecycle work. Amazon SageMaker Canvas supports no-code or low-code business analysis patterns. The practitioner should know the difference without pretending every team must build models.

A practical approval checklist starts with the outcome. What decision will change if the output exists? What data proves the model can work? What harm happens if the output is wrong? What service gives the simplest reliable path? What human review is needed? What will be logged, monitored, and retired? These questions matter more than whether the solution sounds advanced.

• Use a managed AI service when the task is common and well bounded, such as extracting text, translating language, or recognizing labels in images.
• Use Amazon Bedrock or Amazon Q when the core need is generative language, search assistance, summarization, or guided action over enterprise content.
• Use SageMaker AI when the organization needs a custom model lifecycle and has the skill, data, budget, and governance to operate it.
• Use rules, reports, or workflow automation when the outcome must be deterministic or the data is too weak for learning.

Scenario judgment: a compliance team wants every invoice over a fixed threshold to require approval. That is not an ML problem; the threshold is known and deterministic. A Lambda function, application rule, or workflow step is easier to explain, test, and audit. If the same team wants to identify unusual invoice patterns that humans have not explicitly defined, ML may be appropriate after data readiness and risk review.

Another scenario: a product manager asks for AI because competitors advertise AI features. The better first step is to define the user pain. If customers spend too long searching internal documentation, Amazon Q or a Bedrock retrieval workflow might fit. If users need exact account balances, a deterministic database query is the right core. If support agents need suggested categories, classification may help but should be measured against human baseline performance.