Enterprise AI Glossary
Definitions for deployment, governance, and production AI — written for regulated enterprises.
Agentic AI
Agentic AI systems plan multi-step workflows, use tools, and execute actions within guardrails — unlike single-turn chatbots.
AI Accelerator
An AI accelerator is a structured Derisk360 programme that moves one use case from discovery to governed production go-live in weeks with embedded FDEs.
AI Deployment
AI deployment is the end-to-end discipline of moving models and agents from pilot to governed production — context, evaluation, guardrails, VPC go-live, and 24/7 operations.
AI Governance
AI governance is the combination of policies, technical controls, audit mechanisms, and operating models that keep enterprise AI compliant and accountable in production.
AI Ops
AI Ops is continuous monitoring, evaluation, tuning, and incident response for production AI systems after go-live — not launch-and-leave.
AI Risk
AI risk spans model, operational, regulatory, and reputational failure modes — deployment risk (context, governance, ops) dominates in regulated enterprises.
Context Engineering
Context engineering unifies fragmented enterprise data, tools, and knowledge so production agents act on accurate, governed information — via MCP, graphs, and qualified retrieval.
Forward Deployed Engineer
A Forward Deployed Engineer (FDE) embeds inside your organisation to implement governed production AI — context, agents, evaluation, and go-live — with outcome accountability.
Forward Deployed Evaluation Engineer
A Forward Deployed Evaluation Engineer (FDEE) builds eval harnesses, runs red teams, and monitors production AI quality, safety, and compliance continuously.
Guardrails
Guardrails are policy and technical controls — policy engines, tool restrictions, HITL — that constrain agent behaviour before and during execution.
Hallucination
Hallucination is when AI generates confident but incorrect outputs — mitigated in production by grounding, knowledge graphs, MCP context, and continuous eval.
Knowledge Graph
A knowledge graph structures entities and relationships so agents retrieve explainable, governed context — essential for coverage, policy, and compliance questions.
Large Language Model
A large language model (LLM) is a neural network trained on text to generate and reason over language — one component in a governed enterprise agent stack.
Model Context Protocol
Model Context Protocol (MCP) standardises how AI agents connect to enterprise tools, databases, and APIs with governed, reusable connectors.
Multi-Agent System
A multi-agent system coordinates specialised agents — document, policy, escalation — to complete complex enterprise workflows with shared audit state.
Pilot to Production
Pilot to production is the transition from proof-of-concept to governed, operational AI with SLAs, monitoring, and regulatory audit trails.
Retrieval-Augmented Generation
Retrieval-Augmented Generation (RAG) grounds LLM responses by retrieving relevant documents at query time — a baseline for enterprise context before graphs and MCP.
Red Teaming
Red teaming systematically probes AI for safety, security, and compliance failures — documented for model risk before production go-live.
Responsible AI
Responsible AI ensures fairness, transparency, accountability, and human oversight in enterprise deployments — engineered into agents, not policy PDFs alone.
Vector Database
A vector database stores embeddings for semantic search and retrieval in RAG pipelines — one layer in governed enterprise context architecture.
Agent Orchestration
Agent orchestration routes tasks, manages shared state, and coordinates tools across multiple agents under supervisor policy.
AI Evaluation
AI evaluation measures quality, safety, and business outcomes before and after deployment — continuous, not a one-time QA pass.
AI Readiness
AI readiness assesses whether data access, governance, evaluation, and operations can support production AI — scored before accelerator budget commits.
Audit Trail
An audit trail logs agent decisions, tool calls, and data access for regulatory review and incident investigation.
Copilot
A copilot is an AI assistant embedded in workflows to augment human decisions with governed, explainable suggestions — not unsupervised automation.
Deployment Risk
Deployment risk is the probability an AI initiative fails to reach or sustain production value — the core problem Derisk360 derisks.
Embedding
An embedding is a numerical vector representation of text or data used for semantic similarity search in RAG and retrieval pipelines.
Enterprise AI
Enterprise AI applies governed AI systems to core business processes at scale — with SLAs, audit trails, and operational ownership in large organisations.
Eval Harness
An eval harness is automated infrastructure that scores AI outputs against business, safety, and policy criteria on every release and in production.
Fine-Tuning
Fine-tuning adapts a base model to domain-specific language using curated training data — used selectively when grounding and context are insufficient.
Generative AI
Generative AI creates text, code, or content from prompts using large models — the foundation for enterprise copilots and agents when governed.
Grounding
Grounding ties model outputs to verified sources — documents, graphs, MCP data — to reduce hallucination and improve auditor trust.
Human-in-the-Loop
Human-in-the-loop (HITL) keeps people in review or approval steps for high-stakes agent actions in regulated workflows.
Inference
Inference is running a trained model to produce predictions or generations in production — with cost, latency, and scaling implications.
Model Drift
Model drift is degradation in AI performance as data distributions, regulations, or usage patterns change after go-live.
Observability
Observability provides metrics, logs, and traces to understand production AI behaviour — quality scores, latency, cost, and policy violations.
Policy Engine
A policy engine enforces business and regulatory rules on agent actions before execution — programmatic governance, not manual review alone.
Prompt Engineering
Prompt engineering designs inputs, instructions, and tool schemas to elicit reliable model behaviour in production workflows.
Semantic Search
Semantic search finds content by meaning using embeddings — powering RAG retrieval when combined with governance and freshness controls.
Tool Use
Tool use enables agents to call APIs, databases, and enterprise systems via MCP to complete tasks — with scoped permissions and audit.
VPC Deployment
VPC deployment runs AI workloads inside the client's private cloud for data sovereignty, network isolation, and regulatory alignment.
Workflow Automation
Workflow automation replaces manual steps with governed agent-driven processes — orchestrating humans, systems, and AI in audited sequences.
Zero Trust AI
Zero trust AI assumes no implicit trust — every agent action is authenticated, authorised, logged, and evaluated against policy.
AI Total Cost of Ownership
AI total cost of ownership includes models, infrastructure, integration, evaluation, operations, and risk — often 3–5× API spend alone.
Change Management
Change management prepares people and processes for AI-driven workflow transformation — training, escalation paths, and role clarity.
Data Lineage
Data lineage tracks where data originated and how it flows into AI context and outputs — required for audit and model risk.
Explainability
Explainability makes AI decisions interpretable for auditors, regulators, and business users — citations, reasoning traces, and structured summaries.
Model Card
A model card documents purpose, limitations, evaluation results, and intended use of an AI model — standard artefact for model risk submission.
Operating Model
An AI operating model defines roles, runbooks, escalation, and ownership for production AI operations after go-live.
Production AI
Production AI runs in live business processes with SLAs, continuous monitoring, governance, and measurable business outcomes — not demo environments.
Risk Tiering
Risk tiering classifies AI use cases by impact to apply proportional controls — auto-execute low risk, human approval for high impact.
Shadow AI
Shadow AI is unsanctioned use of consumer AI tools by employees — creating data leakage and compliance gaps without enterprise guardrails.
Synthetic Data
Synthetic data is artificially generated data for training or testing when real data is scarce or restricted — used with governance for eval, not production deception.
Token
A token is a unit of text processed by language models — affecting inference cost, latency, and context window limits at production scale.
Use Case Prioritisation
Use case prioritisation ranks AI opportunities by value, feasibility, and deployment risk — selecting one production path instead of pilot sprawl.
Value Realisation
Value realisation measures whether deployed AI delivers promised business outcomes — workflow metrics, not demo accuracy alone.
Ready for an AI implementation partner?
Book a discovery call and we'll map your highest-value use case — and exactly how we get it into production.