Enterprise AI Glossary

The degree to which an AI system maintains correct behavior when faced with deliberately crafted adversarial inputs. Includes resistance to prompt injection, jailbreaks, and data poisoning. Evaluated through red teaming and adversarial test sets. Critical for AI systems in security-sensitive or public-facing roles.

The coordination of multiple AI agents toward a shared goal. Includes task delegation, result aggregation, conflict resolution between agents, error handling, and managing task dependencies. Frameworks include CrewAI, LangGraph, and AutoGen.

AI systems capable of autonomous goal-directed behavior: planning multi-step tasks, selecting and using tools, adapting strategies based on feedback, and taking actions in digital or physical environments. Distinguished from chatbots by their ability to act, not just respond.

An automation in which an AI agent makes some of the decisions, such as classifying an email, choosing which data to fetch or deciding the next step, while fixed workflow steps handle the predictable parts. Usually built in n8n, Make or custom code, with approval steps for high-stakes actions.

The policies, processes, accountability structures, and controls organizations establish to manage AI systems throughout their lifecycle. Covers model approval workflows, bias monitoring, incident response, vendor due diligence, and board-level reporting. Mature AI governance programs map to NIST AI RMF and ISO 42001.

A comprehensive catalog of all AI systems deployed across an organization: purpose, data sources, model versions, risk classification, owners, and compliance status. Required by EU AI Act for high-risk systems. Enables governance, audit readiness, and decommissioning of shadow AI.

An AI deployment with no network connection to external systems. All model weights, data, and compute are isolated within a physical or logical boundary. Required for classified environments, ITAR compliance, maximum data sovereignty, and some HIPAA deployments.

The field of research focused on ensuring AI systems behave according to human values and intentions. Covers value alignment (doing what humans want), goal alignment (pursuing intended objectives), and behavioral alignment (acting safely under distribution shift). Central challenge as AI capabilities scale.

An open-source application for building private document assistants. It ingests your files into workspaces and answers questions with retrieval-augmented generation, using local models (for example via Ollama) or a chosen cloud model.

Connecting two or more systems through their APIs so data moves between them automatically, for example creating a CRM deal when a form is submitted. Good API integrations include authentication, retries, error alerts and logging, and are documented so they can be maintained.

A neural network component that dynamically weights the importance of different input tokens when generating each output token. Enables LLMs to capture long-range dependencies. Self-attention (each token attends to all others) is the core of transformer architecture. Scaled dot-product attention is the most common form.

Microsoft's open-source framework for building multi-agent conversational AI systems. Enables agents to converse with each other to solve tasks. Supports human-in-the-loop integration, code execution, and tool use. Particularly strong for coding-heavy agent workflows and research tasks.

Systematic testing of AI model outputs for unfair disparities across demographic groups (race, gender, age, disability). Includes representation bias in training data, measurement bias in labels, and aggregation bias in predictions. Required by EU AI Act for high-risk AI systems. Tools: Fairlearn, AI Fairness 360, Holistic AI.

A prompting technique instructing the model to reason step-by-step before giving an answer. Significantly improves accuracy on arithmetic, logical reasoning, and multi-step problems. Zero-shot CoT adds "Let's think step by step." Few-shot CoT provides worked examples. Base for reasoning models like o1/o3.

The method used to segment documents into pieces for embedding and retrieval in RAG pipelines. Common strategies: fixed-size (simple), sentence-boundary (clean splits), hierarchical (parent/child), and semantic (meaning-based). Strategy selection significantly affects retrieval precision.

An Anthropic alignment technique where a model is trained to critique and revise its own outputs against a set of principles ("constitution"). Reduces reliance on human labelers for safety. The Claude model family uses Constitutional AI as a core safety method.

The maximum text an LLM can process in a single inference call, measured in tokens. Modern models range from 4K (older) to 1M+ tokens (Gemini 1.5 Pro). Larger windows allow full-document processing but increase cost and latency. Not all tokens are weighted equally — "lost in the middle" degradation is common.

An open-source Python framework for building multi-agent AI systems. Defines agent roles, tools, tasks, and execution flows. Widely adopted for enterprise workflow automation. Supports hierarchical, sequential, and collaborative agent structures.

The design of how a CRM holds data and moves work: pipelines and stages, required fields, record ownership, handoff rules between teams, and connections to other systems. Sound CRM architecture keeps automations, AI and operational reporting reliable as a business adds people, products and pipelines.

A unified data architecture combining the low-cost storage of a data lake with the structure and query performance of a data warehouse. Supports structured and unstructured data. Platforms: Databricks Delta Lake, Apache Iceberg, AWS Lake Formation. Increasingly used as the data foundation for enterprise RAG pipelines.

The principle that data is subject to the laws and governance of the nation where it is collected or processed. Drives requirements for on-premise or region-specific AI deployments. Critical in regulated industries: healthcare (HIPAA), finance (FINRA), and EU data subjects (GDPR).

Retrieval using vector embeddings where semantically similar content is matched even when exact keywords differ. Contrasted with sparse retrieval (BM25/TF-IDF). Most modern RAG systems use hybrid retrieval combining both approaches.

A Stanford framework for programming LLM pipelines using declarative signatures and automated optimization instead of manual prompt engineering. Optimizes prompts and few-shot examples programmatically to maximize metrics. Valuable for complex multi-step reasoning pipelines where manual prompting is brittle.

AI inference running on local devices (laptops, phones, industrial controllers) rather than cloud servers. Enabled by quantization and efficient models (Phi-3, Gemma 2B). Advantages: zero latency, offline operation, privacy. Tradeoffs: limited model capability and context window compared to cloud.

A numerical vector representation of text, image, or other data that captures semantic meaning in a high-dimensional space. Similar content produces nearby vectors, enabling semantic search. Generated by embedding models (OpenAI, Cohere, sentence-transformers). Fundamental to RAG systems and semantic caching.

Abilities that appear unexpectedly in large models but not in smaller versions — not explicitly trained for. Examples: arithmetic, multi-step reasoning, code generation. Difficult to predict in advance. Raises challenges for capability evaluation and safety testing before deployment.

The European Union's comprehensive AI regulation establishing a risk-based framework. Classifies AI systems into prohibited (social scoring, real-time biometrics in public), high-risk (hiring, credit, medical, critical infrastructure), limited-risk (chatbots), and minimal-risk categories. Fully applicable to high-risk systems from August 2026.

An in-context learning technique where a small number of input-output examples are included in the prompt to guide model behavior. No model weights are updated. Effective for format adherence, task specification, and tone calibration. Contrasted with zero-shot (no examples) and fine-tuning (weight updates).

Additional training of a pre-trained model on domain-specific data to improve performance on targeted tasks. Types include full fine-tuning (all weights), LoRA (low-rank adaptation), and RLHF (reinforcement learning from human feedback). Effective but requires curated datasets and GPU compute.

A large AI model trained on broad data that can be adapted to many downstream tasks. Examples: GPT-4o, Claude 3.5, Llama 3, Gemini 1.5, Mistral Large. Distinguished from task-specific models by their generalist capabilities and emergent behavior at scale.

A structured form of tool calling where the LLM outputs a JSON-formatted request to invoke a predefined function. Supported natively by GPT-4, Claude, and Gemini. Enables reliable structured output and API integration. Underlying mechanism for most agent tool use.

The EU General Data Protection Regulation. Applies to all AI systems processing personal data of EU residents. Key AI implications: Article 22 restricts fully automated decision-making; DPIAs required for high-risk processing; data minimization constrains AI training data; individuals have rights to explanation of automated decisions.

A RAG architecture using knowledge graphs to represent relationships between entities. Outperforms vector RAG on multi-hop queries (e.g., "find all contracts involving company X where liability exceeds $1M and renewal is within 90 days"). Best suited for highly relational knowledge domains: legal, regulatory, financial.

The practice of connecting LLM outputs to verifiable external sources — documents, databases, or APIs — to reduce hallucination and improve factual accuracy. RAG is the primary grounding technique. Grounded responses can cite sources, enabling human verification of AI-generated claims.

Programmatic constraints on AI system inputs and outputs. Input guardrails sanitize prompts to prevent injection attacks and policy violations. Output guardrails validate, filter, or flag generated content before delivery. Libraries: Guardrails AI, Llama Guard, custom regex/classifier chains.

When an LLM generates confident-sounding but factually incorrect information not supported by its training data or retrieved context. A fundamental limitation of generative AI. Primary mitigation: RAG (grounding responses in retrieved documents), output validation, and human review for consequential decisions.

The Health Insurance Portability and Accountability Act. Applies to all AI systems processing Protected Health Information (PHI). Requires BAAs with AI vendors, audit logging, data minimization in training, and proper de-identification meeting Expert Determination or Safe Harbor standards.

A design pattern where humans review, approve, or override AI decisions at defined checkpoints. Required by EU AI Act for high-risk AI categories. Critical for consequential decisions in healthcare, lending, hiring, and legal contexts. Balances automation efficiency with accountability.

A retrieval architecture combining sparse (keyword/BM25) and dense (vector embedding) retrieval. Often outperforms either method alone — sparse retrieval handles exact entity matches while dense handles semantic similarity. RRF (Reciprocal Rank Fusion) is a common hybrid merging strategy.

The process of running a trained model to generate outputs from new inputs. Distinct from training. Production inference must optimize for latency, throughput, and cost. Key variables: batch size, context length, model precision, and hardware (GPU/CPU/TPU).

AI-powered extraction of structured data from unstructured documents: PDFs, scanned forms, emails, contracts. Combines OCR, NLP classification, and ML extraction. Replaces manual data entry. Key metrics: extraction accuracy, straight-through processing rate, exception rate.

The first international standard for AI Management Systems (AIMS), published by ISO in 2023. Provides a framework for responsible AI development, deployment, and operation. Covers AI policy, risk assessment, impact evaluation, and continual improvement. Increasingly referenced in enterprise procurement and government AI contracts alongside NIST AI RMF.

Techniques used to bypass LLM safety alignments and generate prohibited content. Common methods: role-playing scenarios, token smuggling, multi-turn manipulation, competing objectives. Model developers continuously patch jailbreaks; defense requires layered guardrails beyond model-level alignment alone.

A structured representation of information as entities (nodes) and relationships (edges). Enables multi-hop reasoning and complex relationship queries. Used in GraphRAG to improve retrieval precision. Tools: Neo4j, Weaviate Knowledge Graphs, Microsoft GraphRAG library.

Key-Value cache storing computed attention states for prior tokens to avoid recomputation during autoregressive generation. Critical for inference efficiency: eliminates O(n²) recomputation in long sequences. PagedAttention (vLLM) manages KV-cache in non-contiguous memory pages for GPU efficiency.

A Python/JavaScript framework for building LLM applications. Provides abstractions for chains, agents, tools, and memory. Broadly adopted but often replaced with lighter frameworks in high-performance production environments due to abstraction overhead.

A LangChain extension for building stateful, graph-based agent workflows. Models agent logic as directed graphs where nodes are actions and edges are transitions. Enables cycles, conditional branching, and persistent state — essential for complex multi-step agentic systems requiring loop control.

Time from request to first token (TTFT) or complete response. LLM latency drivers: model size, context length, server load, and network round-trips. Typical targets: <500ms TTFT for interactive use cases, <2s for document processing. Mitigation: smaller models, quantization, caching, streaming.

A data framework for building LLM-powered applications with external knowledge sources. Specializes in ingestion, indexing, and retrieval pipelines. Strong support for advanced RAG patterns: hierarchical indexing, hybrid retrieval, sub-question decomposition.

A deep learning model trained on large text corpora, capable of generating, summarizing, classifying, and reasoning about natural language. Foundation of modern conversational AI, document processing, and agentic systems. Key parameters: model size (parameters), context window, training data, and alignment method.

A parameter-efficient fine-tuning technique that inserts trainable low-rank matrices into frozen model layers. Updates only 1–5% of total parameters, dramatically reducing GPU memory and training time. Multiple LoRA adapters can be swapped onto the same base model for different tasks.

A visual automation platform (formerly Integromat) for building multi-step "scenarios" that connect apps, transform data and branch on conditions. Well suited to complex visual workflows; production scenarios should include error handlers, logging and alerts.

An architecture where only a subset of model parameters ("experts") activates for each token. Enables very large total parameter counts while keeping per-token compute constant. Examples: GPT-4 (rumored), Mixtral 8x7B, Grok-1. Improves efficiency at scale but complicates distributed serving.

Machine Learning Operations — practices and tooling for productionizing, monitoring, and maintaining ML models. Covers CI/CD for models, feature stores, experiment tracking (MLflow, W&B), data versioning (DVC), model registries, and drift monitoring. Enterprise MLOps platforms: SageMaker, Vertex AI, Databricks MLflow.

A standardized documentation artifact accompanying AI models describing: intended use, training data, evaluation metrics, performance by demographic subgroup, known limitations, and ethical considerations. Introduced by Google. Required by EU AI Act for high-risk models. Promotes transparency in AI supply chains.

An open protocol by Anthropic enabling standardized communication between AI models and external tools, data sources, and services. Emerging standard for tool integration in agentic systems. Enables AI to query databases, search the web, read files, and call APIs through a common interface.

The policies and processes controlling how ML models are developed, validated, deployed, monitored, and retired. Includes: version control, staging environments, approval gates, performance thresholds, drift detection, and audit trails. Regulated industries (banking, insurance) require formal model governance under SR 11-7 and similar guidance.

The infrastructure and software layer that exposes trained models via APIs for inference requests. Concerns: batching, load balancing, autoscaling, model versioning, A/B testing, and monitoring. Production stacks: vLLM, TGI, Triton Inference Server, BentoML, Ray Serve.

An AI architecture comprising multiple specialized agents that collaborate to complete complex tasks. Architectures: hierarchical (manager assigns to specialists), collaborative (peer agents share context), or competitive (agents check each other's work). Enables parallelism and specialization beyond single-agent capability.

AI systems that process and generate multiple modalities of data: text, images, audio, video, and code. Examples: GPT-4o (text + vision + audio), Gemini 1.5 Pro, Claude 3.5 Sonnet (text + vision). Enables document understanding with images, visual QA, and voice-driven interfaces.

A workflow automation platform that can be self-hosted or run in the cloud. Its node-based workflows support custom code, AI agent nodes and high volumes, and self-hosting keeps regulated data inside your own environment.

The NIST Artificial Intelligence Risk Management Framework. A voluntary framework organizing AI risk management across four functions: Govern (policies), Map (context and harms), Measure (analysis), and Manage (response and monitoring). Increasingly referenced in enterprise procurement and government contracts.

The ability to understand an AI system's internal state from its external outputs: logs, traces, and metrics. LLM-specific observability includes: prompt/response logging, token usage, latency percentiles, retrieval quality, and error rates. Platforms: LangSmith, Arize, Helicone, Langfuse. Required for production reliability and compliance audit trails.

Technology converting images of text (scanned documents, photos) into machine-readable text. Modern AI-based OCR (Tesseract, AWS Textract, Azure Form Recognizer) handles handwriting, tables, and complex layouts. Foundation layer of Intelligent Document Processing pipelines.

An open-source tool for running quantized LLMs locally on consumer hardware (CPU + GPU). Manages model downloads, GGUF format inference via llama.cpp, and exposes an OpenAI-compatible REST API. Ideal for developer environments, prototyping, and edge deployments. Not designed for high-concurrency production.

An LLM deployed entirely within an organization's own infrastructure — no data leaves the environment. Common models: Llama 3, Mistral, Qwen, Phi-3. Serving frameworks: vLLM, Ollama, TGI. Required for data sovereignty, HIPAA, ITAR, and latency-sensitive applications.

A self-hosted, browser-based chat interface for local and private language models, commonly paired with Ollama. Supports user accounts, role-based access and document chat, giving staff a ChatGPT-style experience without data leaving your environment.

A specification standard for describing REST APIs in a machine-readable format (JSON/YAML). Used to auto-generate LLM tool schemas, enabling agents to call external APIs without custom integration code. Supported natively by OpenAI function calling and many agent frameworks.

The middleware component in an AI system responsible for routing requests, managing agent lifecycles, sequencing tool calls, handling retries, and aggregating outputs. Abstracts execution complexity from application logic. Critical for reliability and observability in production agentic deployments.

A memory management technique used in vLLM that organizes KV-cache memory in non-contiguous pages (like OS virtual memory). Eliminates wasted reserved memory, enabling 2–4x higher GPU throughput for concurrent inference. Core innovation behind vLLM's production adoption.

Any data that can identify an individual: names, emails, SSNs, IP addresses, biometrics. AI systems handling PII must comply with GDPR, CCPA, or HIPAA depending on jurisdiction and data type. PII must be minimized in training data, masked in prompts, and protected in RAG document stores.

A sales-focused CRM organised around visual pipelines of deals and activities. It supports multiple pipelines, custom fields, webhooks and a full API, which makes it well suited to automation with n8n, Make or Zapier, prefilled client forms and AI-assisted deal workflows.

The practice of designing input prompts to optimize LLM outputs. Includes system prompts, few-shot examples, chain-of-thought instructions, output format specification, and persona definition. Effective prompt engineering can dramatically improve accuracy and consistency without fine-tuning.

An attack where malicious instructions in user input or retrieved documents override an AI system's intended behavior. Direct injection: user writes "ignore previous instructions." Indirect injection: attacker embeds instructions in a document the AI will retrieve. Primary security threat for LLM-based applications.

Quantized LoRA — fine-tuning technique combining 4-bit quantization of the base model with LoRA adapters. Reduces fine-tuning memory requirements by ~75%, enabling fine-tuning of 65B+ parameter models on a single consumer GPU. Democratized domain-specific fine-tuning for open-source models.

Reducing the numerical precision of model weights to decrease memory requirements and increase inference speed. FP32 → FP16 → INT8 → INT4. INT4 quantization reduces memory by ~75% with modest accuracy loss. Methods: GPTQ, AWQ, GGUF (for CPU inference). Enables running large models on fewer GPUs.

An architecture combining information retrieval with LLM generation. At query time, relevant documents are retrieved from a knowledge base and injected into the LLM context. Grounds responses in current, domain-specific facts and dramatically reduces hallucination rates. Now the standard approach for enterprise knowledge applications.

An agent architecture pattern alternating between Reasoning (thinking about what to do) and Acting (invoking a tool). Each cycle: Thought → Action → Observation → Thought. Improves agent reliability by making reasoning explicit and enabling step-level error detection and correction.

LLM variants trained to perform extended step-by-step reasoning before answering. Examples: OpenAI o1/o3, Claude 3.5 with extended thinking. Significantly better on complex logic, math, and multi-step planning. Trade-off: slower inference, higher cost, and "thinking" tokens are consumed before the answer.

Structured adversarial testing of AI systems by dedicated teams attempting to elicit harmful, biased, or incorrect outputs. Mimics real attacker behavior. Required by EU AI Act for high-risk AI and by US Executive Order 14110 for frontier models. Critical for identifying blind spots before deployment.

A second-stage scoring step in RAG pipelines that re-evaluates retrieved candidates using a more sophisticated cross-encoder model. Improves relevance beyond first-stage retrieval. Common models: Cohere Rerank, BGE Reranker. Adds latency but significantly improves answer quality for complex queries.

A training approach where human preferences between model outputs are used as a reward signal to fine-tune model behavior. Three steps: supervised fine-tuning, reward model training, PPO optimization. Used to align GPT-4, Claude, and Llama 2 chat variants for helpfulness and safety.

Software that automates rule-based, repetitive digital tasks by simulating human interaction with UIs and applications. Traditional RPA requires structured inputs and is brittle if UI changes. Modern "agentic RPA" adds AI for exception handling, unstructured data, and adaptive decision-making. Tools: UiPath, Automation Anywhere, Power Automate.

Empirical relationships between model performance and scale factors: parameters, training compute, and training data. Chinchilla scaling laws (2022) showed models are often over-parameterized relative to training data. Guides decisions on model size vs. training budget trade-offs.

Caching LLM responses based on semantic similarity of queries, not exact string match. Reduces API calls when users ask semantically equivalent questions ("What is your return policy?" vs. "Can I return an item?"). Reduces cost and latency for high-volume, repetitive query patterns.

Microsoft's open-source SDK for integrating LLMs into .NET, Python, and Java applications. Supports plugin-based skill registration, planner agents, and memory connectors. Designed for enterprise application integration with Azure OpenAI and other providers.

A deployment strategy where a new AI model runs in parallel with an existing system, processing real requests but not serving responses to users. Used to compare outputs, catch regressions, build confidence before cutover, and gather ground-truth labels without production risk.

A voluntary auditing standard from AICPA covering security, availability, processing integrity, confidentiality, and privacy. Increasingly required by enterprise buyers for AI vendors. A Type II report verifies that controls operated effectively over a 6–12 month period, which is more rigorous than a Type I point-in-time assessment.

Traditional keyword-based retrieval using TF-IDF or BM25 scoring. Efficient and interpretable; excels on exact term matches and entity names. Struggles with synonyms and paraphrasing. Still valuable in hybrid RAG systems paired with dense retrieval for comprehensive coverage.

Instructions provided to an LLM before user input to set context, persona, output format, and constraints. Invisible to end users but controls model behavior throughout the session. Prompt injection attacks attempt to override system prompt instructions via user input or retrieved content.

A sampling parameter controlling output randomness. Temperature 0 produces deterministic, highest-probability outputs. Temperature 1 preserves the model's original distribution. Values above 1 increase diversity but reduce coherence. Production systems typically use 0–0.3 for factual tasks, 0.7–1.0 for creative tasks.

The number of tokens or requests an LLM serving system can process per unit time (tokens/sec or requests/sec). Key production metric. vLLM's PagedAttention improves throughput by 2–4x over naive serving. Throughput and latency often trade off; batch processing maximizes throughput at the cost of per-request latency.

The basic unit of text in LLM processing. Roughly 0.75 words in English, though tokenization varies by model and language. LLM pricing and context limits are measured in tokens. Efficient prompt design reduces token consumption. Non-English text often requires more tokens per word than English.

An LLM capability to invoke external functions or services during inference. The model generates a structured request (function name + parameters), the host executes the tool, and returns results to the model. Enables agents to search the web, query databases, execute code, call APIs, and read files.

The neural network architecture introduced in "Attention Is All You Need" (2017) that underlies virtually all modern LLMs. Uses self-attention to process tokens in parallel (unlike RNNs). Encoder models (BERT) suit classification; decoder models (GPT, Llama) suit generation; encoder-decoder (T5) suit translation and summarization.

A database optimized for storing and querying high-dimensional embedding vectors. Supports approximate nearest neighbor (ANN) search for semantic retrieval. Examples: Weaviate, Pinecone, Qdrant, Milvus, pgvector. Core infrastructure component for all RAG systems and semantic search applications.

A model that combines image encoding with language generation, enabling visual question answering, image captioning, chart interpretation, and document extraction from images. Examples: LLaVA, PaliGemma, Phi-3 Vision. Valuable for processing scanned documents, medical imaging, and manufacturing quality control.

An open-source, high-throughput LLM serving engine. Uses PagedAttention for efficient GPU memory management, achieving 2–4x higher throughput than naive inference. Supports OpenAI-compatible API, multiple model backends, and streaming. The dominant open-source choice for production on-premise LLM serving.

A platform for designing, testing and deploying conversational AI agents for chat and voice. Often used with voice platforms such as Vapi or Retell, and connected to CRMs and knowledge bases so agents can complete tasks, not just answer questions.

The use of software to execute predefined sequences of business processes with minimal human intervention. AI-enhanced workflow automation adds decision points where models handle exceptions, route documents, or draft responses. Platforms: n8n, Zapier, Make, Microsoft Power Automate, custom agentic stacks.

A cloud automation service that links thousands of apps through trigger-and-action "Zaps". The fastest option for simple, low-volume tasks; complex or high-volume workflows are often better suited to Make, n8n or custom code.

Using a model for a task without providing any labeled examples in the prompt, relying solely on instructions. Foundation models with sufficient scale generalize well zero-shot. Effective for well-described tasks; few-shot generally improves performance on edge cases and formatting.