AI Agents
Autonomous software entities that perceive context, reason about goals, and take actions through tools and APIs. Unlike static scripts, agents plan multi-step work, adapt to new inputs, and operate under explicit policy boundaries.
#The Enterprise AI Glossary.
Precise definitions for the vocabulary of agentic enterprise AI — from multi-agent orchestration to EU AI Act post-market monitoring. Written for decision-makers, not engineers.
Agentic AI
Multi-Agent Orchestration
The coordinated operation of multiple specialized agents toward a shared outcome. An orchestrator decomposes work, routes tasks, merges results, and removes duplicates — turning a swarm of models into one coherent system.
#Agentic Process Automation
The automation of cross-system business processes by agents that hand work between ERP, CRM, and document systems autonomously — with human sign-off at defined gates instead of humans driving every step.
#Human-in-the-Loop (HITL)
An operating pattern in which AI proposes and humans dispose: agents draft, optimize, and execute within bounds, while designated decisions require explicit human approval before they take effect.
#Circuit Breakers (AI)
Deterministic safeguards that arrest an agent the moment it deviates from policy bounds, enters a reasoning loop, or attempts an unauthorized action — before cloud credits drain or faulty records are written.
#Models & Inference
Large Language Models (LLMs)
Foundation models trained on broad text corpora that generate, summarize, and reason over language. In enterprise settings their value is bounded by grounding: what they can retrieve and verify, not what they memorized.
#Hallucination
Confident, fluent output that is factually wrong or unsupported. In production systems hallucination is an operational defect — mitigated by retrieval grounding, deterministic verification, and source-cited answers.
#Model Drift
The degradation of model behavior over time: as user behavior shifts and foundation models receive silent backend updates, outputs drift from their validated baseline. Drift requires continuous measurement, not one-off testing.
#Semantic Caching
The reuse of answers to semantically equivalent questions. By matching intent rather than exact text, semantic caches cut redundant model calls — typically 40–60% of token spend — without changing the user experience.
#Dynamic Model Tiering
Routing each query to the smallest model that can handle it: simple tasks run on small local models, while only genuinely complex reasoning invokes high-cost frontier models.
#Inference
The act of running a trained model to produce an output. Inference — not training — is where enterprise AI budgets concentrate, and where optimization, caching, and routing deliver their savings.
#Data & Grounding
Retrieval-Augmented Generation (RAG)
A pattern in which the model retrieves relevant enterprise documents before answering, grounding responses in your data rather than generic training. Enterprise-grade RAG adds access control, citations, and index freshness.
#Knowledge Graph
A structured map of entities and relationships — who owns what, which contract depends on which supplier. Graphs let agents reason across connected facts that keyword search cannot traverse.
#Embeddings
Numerical vector representations of text that capture semantic meaning, enabling search by intent. Embedding quality decays as language and data evolve — an index that is never refreshed quietly degrades every answer.
#Enterprise Ontology
A model-driven semantic map of operational reality — assets, batches, work orders, contracts — that unifies disparate systems into one coherent graph. The ontology is what turns systems of record into systems of intelligence.
#Prompt Injection
An attack in which adversarial text inside retrieved content or user input overrides an agent's instructions. Mitigation combines input sanitation, least-privilege tool access, and output verification.
#Operations & Compliance
AI Site Reliability Engineering (AI SRE)
The application of SRE discipline — monitoring, incident response, capacity planning, blameless post-mortems — to production AI. AI SRE watches model health, cost, and behavior, not just servers.
#LLMOps
The operational practice of running language models in production: versioning, evaluation, observability, cost management, and safe rollout of prompt and model changes.
#Shadow AI
Unsanctioned AI use inside an organization — employees pasting client data into consumer chatbots, teams wiring unvetted APIs into workflows. Shadow AI is simultaneously a compliance exposure and a signal of unmet demand.
#EU AI Act
The European Union's AI regulation, imposing risk-based obligations on providers and deployers — including post-market monitoring, audit logging, AI literacy (Article 4), and strict duties for high-risk systems such as HR and credit scoring.
#ISO/IEC 42001
The international management-system standard for AI. Certification demonstrates governed, auditable AI operations — increasingly a procurement requirement in European enterprise.
#Post-Market Monitoring
The EU AI Act requirement to continuously observe deployed AI systems after go-live: tracking performance, incidents, and drift, and logging evidence that the system behaves as validated.
#AI Literacy (Article 4)
The legal obligation for organizations deploying AI to ensure personnel possess sufficient AI competence for their roles — verifiable through testing and certification, not attendance at a webinar.
#Data Sovereignty
The principle that data remains subject to the laws and jurisdiction where it is held. On European bare metal, enterprise data stays beyond the reach of extraterritorial discovery mandates such as the US CLOUD Act.
#Air-Gapped Deployment
An architecture physically isolated from public networks — the highest assurance tier for sovereign AI, used where data must never leave a controlled perimeter.
#