7.1. Terms A–Z

A/B testing: Randomly assigning cases or users to an AI-assisted process or the current process and comparing outcomes, to attribute impact to AI. See 3.2.2.

Accountable owner: The single named person, usually a business leader, answerable for an AI system's outcomes, risks, and decisions. See 4.2.1, 5.3.2.

Agent-to-agent communication (A2A): AI agents exchanging tasks and results with one another; the Agent2Agent protocol standardizes this across vendors and teams. See 2.2.2.

AI agent: An AI system that pursues a goal by planning steps, using tools and business systems, observing results, and acting or escalating. Contrast with chatbot. See 2.2.2.

AI center of excellence (CoE): A team that concentrates AI expertise, standards, reusable assets, and enablement to help the organization scale AI. Operating models include centralized, hub-and-spoke, and federated. See 5.4.3.

AI impact assessment: A structured review at use-case intake describing purpose, stakeholders, data, potential harms, risk tier, and mitigations. See 4.1.3.

AI inventory: A register of an organization's AI systems with purpose, owner, risk tier, and status. See 4.2.1.

AI literacy: The knowledge and skills people need to use, evaluate, and govern AI appropriately in their roles. See 5.3.5.

AI Service Cards: AWS documentation describing the intended uses, limitations, and responsible AI design choices of AWS AI services and models. See 4.1.1.

Algorithm: The learning method that produces a model from data. See 2.1.1.

Amazon Bedrock: AWS's managed service for building generative AI applications on foundation models, with Knowledge Bases, Guardrails, Agents, and AgentCore. Mainly consumption-based pricing. See 1.3.1.

Amazon Bedrock AgentCore: Managed building blocks for running AI agents securely (runtime, memory, gateway for tools, identity, observability). See 2.2.2.

Amazon Bedrock Guardrails: Configurable safeguards for generative AI, including content filters, denied topics, sensitive information filters, contextual grounding checks, and Automated Reasoning checks. See 4.1.4.

Amazon Bedrock Knowledge Bases: Managed retrieval augmented generation that connects foundation models to company content. See 2.3.3.

Amazon Quick: An AI-powered workspace for business users that evolved from Amazon QuickSight, including Quick Sight (BI), Quick Research, Quick Flows, Quick Automate, and Quick Index. Seat-based pricing. See 1.3.1.

Amazon SageMaker AI: AWS's service for building, training, deploying, and monitoring custom ML models. Instance-based pricing. See 1.3.1.

Amazon SageMaker Clarify: A bias-detection and explainability capability for custom ML. Closed to new customers; existing customers can keep using it, and AWS directs new teams to standard bias metrics, SHAP, and Amazon Bedrock Evaluations. See 4.3.2.

Artificial intelligence (AI): The broad field of machines performing tasks that normally require human intelligence. See 2.1.2.

Augmentation: AI recommends or assists and a person decides. Contrast with automation. See 1.1.1.

Automated Reasoning checks: A Bedrock Guardrails capability that validates outputs against defined rules or policies. See 4.1.4.

AWS Cloud Adoption Framework (AWS CAF): AWS guidance organizing the capabilities needed for cloud and AI adoption into six perspectives: Business, People, Governance, Platform, Security, Operations. CAF-AI is the AI-specific version. See 1.3.2.

AWS Cost Explorer: A tool for analyzing actual AWS spending and trends after deployment. See 3.2.5.

AWS Marketplace: A catalog for buying third-party software, AI models, and data with billing through an AWS account. See 1.3.1, 3.1.2.

AWS Pricing Calculator: A tool for estimating AWS costs before deployment. See 3.2.5.

Baseline: A measurement of current performance captured before AI is implemented, used to attribute later changes. See 3.2.2.

Batch inference: Processing many requests asynchronously at a lower price than on-demand, for work that does not need real-time answers. See 3.2.5.

Bias: Systematic differences in how an AI system treats comparable people or cases; can enter at any lifecycle stage. See 4.3.2.

Bias drift: Change in a system's fairness over time as populations, behavior, or data sources shift. See 4.3.2.

Build-buy-partner: The sourcing decision for AI capabilities, weighing budget, timelines, capabilities, vendor proposals, and regulatory requirements. See 3.1.2.

Champions network: Respected people across business units who model AI use, help peers, and relay feedback. See 5.3.1.

Concept drift: A change in the relationship between inputs and correct outputs, so learned patterns no longer hold. See 2.2.3.

Consumption-based pricing: Paying per unit used, such as tokens or requests. See 3.2.5.

Context window: The maximum number of tokens a model can consider at once, including instructions, history, retrieved content, and output. See 2.3.2.

Contextual grounding check: A safeguard that flags responses not supported by the provided source material. See 4.1.4, 4.3.4.

Controllability: The responsible AI dimension concerned with mechanisms to monitor and steer AI behavior. See 4.1.1.

Data drift: A change in the input data a model receives compared with its training data. See 2.2.3.

Data flywheel: A loop in which usage generates proprietary data that improves the model, which improves the product and attracts more usage. See 3.3.3.

Data owner / data steward: The owner is accountable for a data domain's use and quality; the steward maintains its definitions and quality rules day to day. See 5.2.2.

Data silo: Data held separately by one team or system and not accessible or consistent across the organization. See 5.2.1.

Deep learning: Machine learning using multi-layer neural networks, effective on images, audio, and text. See 2.1.2.

Envision, experiment, launch, scale: The iterative transformation phases named in the AIB-C01 exam guide. See 5.4.1.

EU AI Act: European regulation that classifies AI systems by risk (unacceptable, high, limited, minimal) with obligations scaled to each tier. See 4.2.2, 4.2.4.

Explainability: The ability to understand and evaluate why an AI system produced an output. See 4.1.1.

Fairness: Considering the impacts of an AI system on different groups of stakeholders. See 4.1.1.

Few-shot prompting: Including examples of the desired output in a prompt. Contrast with zero-shot. See 2.3.1.

Fine-tuning: Further training a foundation model on curated examples to change its behavior, style, or task performance. See 2.3.3.

Foundation model: A large model pretrained on broad data that can be adapted to many tasks. See 2.1.2.

Generative AI: AI that creates new content such as text, images, or code. See 2.1.2.

Governance by design: Integrating responsible AI and governance from use-case intake and planning through the full lifecycle. See 4.1.3.

Guardrails: Policies and mechanisms that constrain AI inputs and outputs to prevent harm. See 4.1.4.

Hallucination: Fluent but false or unsupported content produced by a generative model. See 4.3.4.

Hub-and-spoke: A CoE model in which a central hub sets standards and shares assets while embedded teams build in business units. See 5.4.3.

Human-in-the-loop / human-on-the-loop: A person approves each decision before it takes effect, or monitors and can override an AI that acts. See 4.1.4.

Inference: Using a trained model to produce predictions or outputs on new input. See 2.1.1.

Instance-based pricing: Paying for provisioned compute capacity by time. See 3.2.5.

IP indemnification: A vendor commitment to defend customers against certain third-party intellectual property claims related to outputs, subject to terms. See 4.3.3.

ISO/IEC 22989: International standard for AI concepts and terminology. See 2.1.4.

ISO/IEC 23053: International framework describing AI systems that use machine learning, their components and functions. See 2.1.4.

ISO/IEC 42001: International, certifiable standard for an AI management system governing how an organization manages AI. See 2.1.4.

KPI (key performance indicator): A metric that reflects the outcome an initiative exists to change; may be tangible or intangible. See 3.2.1.

Leading indicator / lagging indicator: An early signal that predicts success (adoption, usage) versus a result measured after the fact (ROI, revenue). See 3.2.4.

Least privilege: Granting users, applications, and agents only the access they need. See 4.2.3.

Machine learning (ML): A subset of AI in which systems learn patterns from data rather than following written rules. See 2.1.2.

Maturity model: A staged description of how repeatably an organization turns AI into value, from exploring to transforming. See 5.1.2.

Model: The learned pattern produced by training an algorithm on data. See 2.1.1.

Model Context Protocol (MCP): An open standard for connecting AI agents and models to tools and data sources. See 2.2.2.

Model drift: Decline in model performance over time as the world changes. See 2.2.3, 4.3.4.

NIST AI Risk Management Framework (AI RMF): Voluntary US framework organizing AI risk management into Govern, Map, Measure, and Manage. See 4.2.4.

Pilot purgatory: Running many pilots that never reach production because operationalizing capabilities are missing. See 5.1.2.

Prompt engineering: Designing prompts (role, task, context, constraints, examples, format) to produce useful outputs. See 2.3.1.

Proof of concept (POC): A small, time-boxed test of whether an approach can work. See 3.1.3, 5.3.5.

Provisioned Throughput: Amazon Bedrock capacity purchased by the hour, optionally with a commitment, for steady high-volume workloads. See 3.2.5.

Retrieval Augmented Generation (RAG): Retrieving relevant company content at question time and supplying it to a model so answers are grounded and citable. See 2.3.3.

Risk classification: Sorting AI use cases into tiers by potential harm and applying proportionate controls. See 4.2.4.

ROI (return on investment): (Total benefits − total costs) ÷ total costs, using a comprehensive view of both. See 3.2.3.

Rule-based automation: Software that executes explicitly written logic; deterministic and auditable. See 2.2.1.

Seat-based pricing: Paying per user, typically per month. See 3.2.5.

Shadow AI: Use of AI tools without the organization's knowledge or approval. See 2.2.4.

Shared responsibility model: AWS is responsible for security of the cloud; the customer is responsible for security and appropriate use in the cloud. See 1.3.2, 4.2.3.

Structured / unstructured data: Data in defined rows and fields versus data without a predefined schema, such as documents, emails, and images. See 2.1.3.

Token: A small chunk of text that models read and generate; the unit of consumption pricing and context limits. See 2.3.2.

Training: The process of producing a model by having an algorithm learn from historical data. See 2.1.1.

Transparency: Enabling stakeholders to make informed choices about engaging with an AI system, including knowing that AI is involved. See 4.1.1.

Veracity and robustness: Achieving correct outputs, even with unexpected or adversarial inputs. See 4.1.1, 4.3.4.

Working backwards: Starting from the desired customer or business outcome and reasoning back to the capability and technology required. See 1.2.1.

Zero-shot prompting: Prompting with instructions only and no examples. See 2.3.1.