Enterprise AI Glossary

Coordinating multiple AI agents to work together on complex workflows, each specializing in specific tasks.

AI systems that can autonomously plan, execute, and adapt multi-step tasks with minimal human intervention. Unlike chatbots that respond to individual prompts, agentic AI breaks down goals into subtasks, uses tools (APIs, databases, code execution), and iterates based on results.

Autonomous processes where AI agents perceive, plan, and execute multi-step tasks without human intervention, escalating only exceptions.

Systematic testing of AI systems for discriminatory outcomes across protected characteristics such as race, gender, age, and disability. Increasingly mandated by regulation: NYC Local Law 144 requires annual independent audits for automated employment decision tools.

A systematic approach to AI governance: enforcing human-in-the-loop gates, logging every decision with immutable audit trails, scoring risks in real time, and blocking unauthorized operations.

The process of splitting documents into smaller segments for indexing in a RAG system. Chunking strategy (fixed-size, semantic, recursive) directly impacts retrieval quality because it determines what context the LLM receives when generating answers.

A safety mechanism in production AI systems that automatically stops agent operations when anomalies are detected, such as error rate spikes, unusual output patterns, or confidence score drops below threshold. Prevents cascading failures in autonomous systems.

A numeric score (typically 0-100) assigned to an AI system's output indicating the system's certainty in its response. Used to route decisions: high-confidence outputs are auto-delivered, medium-confidence triggers review, low-confidence escalates to a human.

Maximum amount of text (in tokens) an LLM can process in a single request.

The principle that data is subject to the laws and governance structures of the nation or jurisdiction where it is collected or processed. For AI, this means AI models processing data must comply with local data protection laws and the data must remain within approved jurisdictions.

Numerical representation of text that captures semantic meaning, used for similarity search and RAG.

Data pipeline process for extracting data from sources, transforming it, and loading it into target systems.

Comprehensive AI regulation framework from the European Union categorizing AI systems by risk level.

The ability to understand and articulate why an AI system made a specific decision.

The process of adapting open-weights models (like Llama 3) on proprietary enterprise data to increase accuracy and reduce hallucinations without API leakage.

Process of identifying organizational bottlenecks and time-consuming activities that could be automated with AI.

EU data privacy law requiring explicit consent, right to deletion, and data protection for personal information.

When an AI model generates false or fabricated information that sounds plausible but is factually incorrect.

Workflow design where humans review and approve high-stakes AI decisions before execution.

A retrieval approach that combines dense vector search (semantic similarity) with sparse keyword search (exact term matching) for better retrieval quality. Improves Mean Reciprocal Rank by approximately 18.5% compared to vector-only or keyword-only search.

An international standard for AI Management Systems published by the International Organization for Standardization. Provides a structured approach to ethical AI, transparency, and trust through a management system framework. Complements NIST AI RMF by adding operational governance structure.

AI models trained on vast text data to understand and generate human language (GPT-4, Claude, Llama).

A centralized proxy layer that sits between applications and language models, providing policy enforcement, usage monitoring, cost management, model routing, and security controls. Often implements an OpenAI-compatible API so applications can switch between cloud and on-premise models without code changes.

A parameter-efficient fine-tuning technique that adapts pre-trained language models by training only a small set of additional parameters rather than modifying all model weights. Achieves approximately 95% of full fine-tuning performance at roughly 10% of the compute cost.

The practice of applying DevOps principles to machine learning systems: automating model training, testing, deployment, monitoring, and retraining. Covers the full lifecycle from data preparation through production operations.

Degradation of AI model performance over time as real-world data patterns change from training data.

An architecture where multiple specialized AI agents collaborate on complex tasks, each responsible for a specific function. An orchestrator coordinates their interactions, resolves conflicts, and enforces governance rules across the system.

A voluntary, rights-preserving framework published by the National Institute of Standards and Technology for managing AI risk. Operates through four functions: Govern (establish policies), Map (document systems), Measure (assess performance and risk), and Manage (implement controls). Released January 2023 with a Generative AI Profile (AI 600-1) added in July 2024.

Personal data such as SSN, email, phone numbers, medical records that must be protected under privacy regulations.

Crafting precise instructions and examples to guide AI models toward desired outputs.

Architecture pattern where LLMs query a knowledge base before generating responses, combining retrieval and generation for more accurate, grounded answers.

A secondary scoring step in RAG retrieval where a specialized model re-evaluates initial search results to improve precision before passing context to the LLM. Uses cross-encoder models that compare the query and each retrieved chunk more carefully than the initial vector search.

Automatic evaluation of an AI decision on a 0-100 scale. High scores trigger human review.

Unauthorized use of AI tools by employees through personal accounts on free services like ChatGPT, Gemini, or Claude. Creates uncontrolled data exposure because employees input company data into systems the organization does not monitor or govern.

Testing approach where AI systems run in parallel with human operations, making recommendations but not decisions, to validate accuracy.

AI systems trained, deployed, and operated on infrastructure you control, within jurisdictions you choose, independent of third-party API providers. Unlike SaaS AI where every request sends data to external servers, sovereign AI keeps data, models, and processing on your servers or private cloud.

Architecture where all LLMs, embeddings, and decision logs remain within your corporate firewall or VPC. Zero API calls to external cloud providers (no OpenAI, Azure, or public inference). All model weights stay under your control, whether deployed on-premise or within a private cloud.

Unit of text processed by LLMs (roughly 4 characters). Models have token limits and pricing is per-token.

LLM capability to invoke external APIs, databases, or functions as part of task execution.

Ranking AI opportunities from highest-impact to lowest, considering both business value and technical complexity.

Specialized database (like Pinecone, Weaviate) that stores embeddings and enables semantic search for RAG pipelines.

A deployment pattern where data flows through AI systems for processing but is never persisted outside the organization's infrastructure. No training data, input queries, or generated outputs are stored by external providers.