AI Glossary: Agentic AI from A to Z.

The key terms around AI agents, language models and the productive use of AI — explained concisely and continuously expanded. For further detail, many of our role profiles link to the relevant application scenarios.

Agentic AI

Agentic AI refers to AI systems that pursue goals independently across multiple steps: they plan tasks, call tools and systems, evaluate interim results and adapt their approach. The difference from classic generative AI lies in acting — Agentic AI does not merely answer, it gets things done.

AI Agent

An AI agent is a software system with a language model at its control core that carries out multi-step tasks: it breaks a goal down into steps, uses tools such as databases, APIs or applications, and escalates to humans when uncertain. Production agents require guardrails, audit logging and clearly limited permissions.

LLM (Large Language Model)

A large language model is an AI model trained on vast amounts of text that understands and generates language. LLMs form the foundation of modern AI applications — from chat assistants and RAG systems through to agents. Well-known model families come from OpenAI, Anthropic, Google and Mistral, among others.

RAG (Retrieval-Augmented Generation)

RAG is an architectural pattern in which a language model retrieves relevant content from enterprise sources before answering — such as documents, wikis or databases — and bases its response on it. Benefits: up-to-date information, source citations and no model adaptation; the enterprise data does not alter the model.

Prompt Engineering

Prompt engineering is the systematic design and testing of the instructions used to steer language models. Professional prompt engineering works with evaluation datasets and versioning — the goal is repeatable quality, not the occasional lucky hit.

Fine-Tuning

Fine-tuning is the further training of a pre-trained AI model on your own example data in order to adjust behaviour, style or domain knowledge. It only becomes worthwhile once prompting and RAG are no longer sufficient — the maintenance effort is higher, as every model update may require retraining.

Hallucination

A hallucination is a statement by an AI model that is factually wrong but linguistically convincing. Hallucinations cannot be eliminated entirely, but they can be substantially reduced — through RAG with source grounding, controlled answer ranges and evaluation. In business processes, critical AI statements fundamentally belong under human review.

Vector Database

A vector database stores content as numerical representations of meaning (embeddings) and thereby finds semantically similar items — even without matching words. It is the typical search infrastructure behind RAG systems; examples include Qdrant, Weaviate and pgvector.

Embedding

An embedding is the conversion of text (or images) into a numerical vector that captures the meaning: similar content lies close together in vector space. Embeddings make semantic search possible and are a fundamental building block of RAG architectures.

MCP (Model Context Protocol)

The Model Context Protocol is an open standard through which AI applications and agents access external tools and data sources in a uniform way. Instead of building a separate integration for each connection, systems speak a common interface — which significantly lowers integration costs.

Multi-Agent System

A multi-agent system has several specialised AI agents work together on a task — for instance research, drafting and quality review as separate roles with defined handovers. The added value lies in specialisation and mutual control; the price is greater architectural complexity.

Guardrails

Guardrails are technical protective mechanisms that constrain the behaviour of AI systems: permitted topics and actions, permission boundaries, input and output filters, approval points for critical steps. For agents with system access, guardrails are not an option but a prerequisite.

Human-in-the-Loop

Human-in-the-loop refers to process designs in which humans review or approve AI results at defined points — for example before sending a customer response or executing a booking. It is the most important pattern for combining AI automation with accountability.

MLOps / LLMOps

MLOps applies DevOps discipline to machine learning: versioning of data and models, automated pipelines, monitoring for quality degradation (drift). LLMOps applies the same discipline to language-model applications — with prompt versioning, evaluation pipelines and cost monitoring.

EU AI Act

The EU AI Act is the European AI regulation. It categorises AI systems into risk classes and ties graduated obligations to them — from transparency through documentation and human oversight to requirements for employees’ AI literacy. The obligations apply regardless of company size and take effect in stages.

SAP Joule

Joule is SAP’s AI copilot across the SAP application landscape — from S/4HANA through SuccessFactors to Ariba. Users ask questions and trigger processes in natural language; Joule Agents extend this with agentic automation of multi-step business processes, and custom extensions are created via Joule Studio.

Salesforce Agentforce

Agentforce is Salesforce’s platform for AI agents in service, sales and marketing: agents answer customer enquiries or qualify leads independently, work on the CRM data, and are governed through topic boundaries, guardrails and escalation rules.

ServiceNow Now Assist

Now Assist is the generative AI layer of the ServiceNow platform: it summarises tickets, suggests responses and generates knowledge articles in ITSM, HR and Customer Service. ServiceNow AI Agents extend this with the independent handling of standard cases.

Microsoft 365 Copilot

Microsoft 365 Copilot is the AI assistant in Word, Excel, Outlook, Teams and the rest, which accesses enterprise content within the Microsoft tenant. Before rollout, permissions and data classification should be reviewed — Copilot makes visible what users have access to, including unintentionally. Custom agents are created via Copilot Studio.

Prompt Injection

Prompt injection is an attack in which manipulated inputs or doctored content (e.g. an email processed by an agent) lead an AI system to perform unwanted actions. It is the central security risk of agents with system access and is mitigated through permission boundaries, content isolation and approval points.

Token

Tokens are the computational units of language models — word fragments into which text is broken down. They determine costs and context limits: billing is per processed token, and each model can only consider a limited amount of tokens at once (the “context window”).