AI glossary: 80+ key terms for SMEs and fiduciary firms

Glossary from A to Z

This glossary explains the most important AI terms relevant to SMEs, fiduciary firms and compliance work in May 2026. Each entry has one or two sentences of definition. Where a dedicated knowledge page exists, the note "see slug" points to the deeper page.

A

Agent. An AI system that executes multiple steps autonomously, uses external tools and evaluates its own intermediate results. A plain LLM without tools is not an agent. See was-ist-ai-agent.

AGI (Artificial General Intelligence). Hypothetical AI with human-equivalent general intelligence. As of May 2026 not achieved; a marketing term in press releases, not a product.

Alignment. Research field and training technique that binds model behaviour to human values and instructions. Methods: RLHF, DPO, Constitutional AI.

API (Application Programming Interface). Programming interface. Every LLM vendor (OpenAI, Anthropic, Mistral) provides a REST API. See integration-rest-api-graphql.

Attention. Mechanism in transformer models that decides per token which other tokens in the context matter. Core component of all modern LLMs. See was-ist-attention-mechanismus.

Audit-Trail. Complete record of every AI request and response including model, version, prompt, source and time. Required for audit-grade use in fiduciary and legal work. See ai-audit-trail-design.

AVV (Auftragsverarbeitungsvertrag). GDPR contract between controller and processor. In Switzerland the equivalent is the processing contract under revFADP. See drittlandtransfer-tia.

B

Backup-3-2-1. Standard rule: three copies, two different media, one copy offsite. Also applies to vector indexes and AI configurations. See backup-strategien-3-2-1.

BERT (Bidirectional Encoder Representations from Transformers). Google model from 2018, predecessor of modern LLMs. Rarely used directly in May 2026, but a basis of many embedding models.

BGE (BAAI General Embedding). Family of open-source embedding models from China. BGE-M3 is multilingual and strong on German. See tool-bge-m3.

Bias. Systematic distortion in training data or model answers. Relevant in Switzerland for anti-discrimination (e.g. HR and credit decisions). See bias-fairness-audits.

BPE (Byte-Pair Encoding). Algorithm to split text into tokens. Standard with GPT, Claude and Mistral. See was-ist-token.

C

Chain-of-Thought (CoT). Prompt technique that asks the model to write out its intermediate steps. Substantially improves accuracy on logic tasks.

Chunking. Splitting long documents into pieces (typically 300-800 tokens) for RAG indexing. Bad chunking yields bad answers. See chunking-strategien-rag.

Cloud-Act. US law from 2018 giving US authorities access to data held by US firms in the cloud, even when servers are in the EU or Switzerland. Core argument for Swiss cloud hosting. See drittlandtransfer-tia.

Closed-Weight. Model whose weights are not public (GPT-4, Claude, Gemini). Opposite: open-weight. See trend-open-weight-vs-closed.

Constitutional AI. Anthropic method of aligning models not with human feedback but with an explicit list of principles. The basis of Claude.

Context Window. Maximum number of tokens a model processes per request. May 2026: GPT-4o 128k, the current top Claude model 200k, Gemini 2 Pro 2M. See was-ist-context-window.

CoPilot. Microsoft brand name for AI assistants in Microsoft 365 and GitHub. Not to be confused with the generic "co-pilot" concept. See integration-microsoft-365-graph-api.

CrowdSec. Open-source intrusion prevention. We put it in front of every production AI service as a barrier against bots and brute-force. See firewall-und-crowdsec.

D

Distillation. Training technique in which a large model trains a smaller model on a specific task. Common in May 2026 to make local models efficient.

DPA (Data Processing Agreement). English term for AVV. Mandatory for any use of OpenAI, Anthropic or Google with personal data. See dsgvo-und-llms.

DPO (Direct Preference Optimization). Training method as an alternative to RLHF, simpler and more stable. Standard for alignment training in May 2026.

Drift. Gradual change in model answers over time, e.g. due to changing inputs or model versions. Requires regression tests. See regression-testing-llm.

DSGVO (General Data Protection Regulation). EU regulation from 2018. Applies to Swiss firms when EU citizens are involved. See dsgvo-und-llms.

E

Edge-AI. AI models that run directly on end devices (phone, laptop, microcontroller) without cloud. See trend-edge-ai-und-on-device.

Embedding. Numerical representation of a text as a vector. Similar texts land close together in vector space. Basis of RAG. See embeddings-und-vektoren.

Encoder. Part of a transformer model that converts input into an internal representation. BERT is encoder-only; GPT is decoder-only.

EU AI Act. EU regulation from 2024, staggered entry into force 2025-2027. Classifies AI applications by risk tier. Applies to Swiss firms with EU-market exposure. See eu-ai-act-2026.

Eval. Test suite that measures a model on defined tasks. Mandatory before any production use. See eval-frameworks-für-llms.

Extended Thinking. Anthropic feature since 2025: Claude writes an extended thinking pass before its answer. Similar to OpenAI o-series. See was-ist-reasoning-modell.

F

Faithfulness. RAG metric measuring whether the answer actually comes from the retrieved sources and is not invented. See ki-qualität-kpis-rag.

Fine-Tuning. Retraining an existing model on your own data. In May 2026 often replaced by RAG; worthwhile for domain-specific vocabulary. See was-ist-fine-tuning-vs-rag.

FINMA. Swiss financial market supervisory authority. Has issued an AI circular binding on banks and insurers. See finma-ki-rundschreiben.

FlashAttention. Algorithm that makes attention memory-efficient. Built into all modern LLM inference engines in May 2026.

Function-Calling. Ability of a model to call external functions with structured arguments. Predecessor of MCP. See was-ist-tool-use-function-calling.

G

GGUF. File format for quantized open-weight models, standard with llama.cpp and Ollama. See tool-llama-cpp.

GPU (Graphics Processing Unit). Hardware accelerator for AI training and inference. May 2026: NVIDIA H100/H200 standard for cloud, RTX 5090 for workstations. See gpu-kosten-rechner.

Grounding. Procedures to bind model answers to verifiable external data. RAG is the most common form. See structured-extraction-grounding.

Guardrails. Rule layer in front of and behind the LLM that filters harmful input and blocks wrong output. See red-teaming-für-ki.

H

Hallucination. Model invents plausible but wrong content. Reducible with RAG, not eliminable. See halluzinationen-begrenzen.

HNSW (Hierarchical Navigable Small World). Standard index algorithm for vector databases (Qdrant, Weaviate, pgvector). Fast and stable in May 2026. See was-ist-vektor-index.

HyDE (Hypothetical Document Embeddings). Retrieval technique: the model generates a hypothetical answer, which is then embedded and searched. Improves recall on short questions.

I

Inference. Application of a trained model to new inputs. Contrast with training. Inference produces running cloud cost.

Instruction-Tuning. Retraining a base model so it follows instructions. Turns an auto-completer into a usable assistant.

IVF (Inverted File Index). Older vector-index algorithm, largely replaced by HNSW. Faiss default.

J

Jailbreak. Input that gets the model to bypass its safety guardrails. Largely patched in May 2026, but input filtering remains mandatory.

JSON-Mode. Model mode that guarantees structured JSON outputs. Important for tool use and data extraction. See output-formatierung-json-modus.

K

Kontext-Fenster. German term for context window. See was-ist-context-window.

K-Nearest-Neighbor (kNN). Search algorithm that finds the k most similar vectors. Core of every RAG search.

L

Latent Space. Mathematical space in which the model encodes semantic meaning. Embeddings are points in this space.

LLM (Large Language Model). Large language model. Standard for text AI in May 2026, from GPT-4o and the current top Claude model to Llama 4 and Mistral Large 2. See wie-funktioniert-ein-llm.

Long-Context. Models with context window from 100k tokens upward. Standard with all premium vendors in May 2026.

LoRA (Low-Rank Adaptation). Efficient fine-tuning method that trains only a small part of the weights. Standard for SME budgets.

Lost-in-the-Middle. Phenomenon that models often miss information in the middle of a long context. Argument for precise retrieval rather than huge prompts.

M

MCP (Model Context Protocol). Anthropic standard from Nov 2024 for tool use. Adopted by OpenAI, Google and Microsoft as of May 2026. See was-ist-mcp.

Mistral. French LLM vendor, EU-sovereign. Models Mistral Large 2 and Pixtral. See mistral-eu-evaluation.

MoE (Mixture of Experts). Architecture where only a part of the model parameters are active per token. Mistral Mixtral and GPT-4 use MoE. See was-ist-mixture-of-experts.

Multi-Head Attention. Multiple parallel attention calculations per layer. Standard component of transformer models.

Multimodal. Model handles more than text, e.g. images, audio, video. GPT-4o, the current top Claude model and Gemini 2 are multimodal. See was-ist-multimodal-ki.

N

NDA (Non-Disclosure Agreement). Standard confidentiality contract. Required with any AI vendor handling client data.

revDSG (Swiss revised FADP). Swiss data protection act in force since September 2023. Mandatory basis for any AI use with personal data. See ndsg-revfadp-ki.

NER (Named Entity Recognition). Detection of persons, firms, places and dates in text. Often used as a pre-processing step before RAG in May 2026.

Neuron. Mathematical unit of a neural network, essentially a weighted sum plus activation function. Term simplified; biological analogy ends here.

O

OCR (Optical Character Recognition). Conversion of an image or scan into text. Prerequisite for RAG with paper receipts. See ocr-für-belege-und-verträge.

OIDC (OpenID Connect). Auth standard built on OAuth. Standard for AI tools with SSO in May 2026. See integration-sso-saml-oidc.

OpenAI. US LLM vendor, maker of the GPT family and Whisper. Market leader by share in May 2026. See openai-gpt-evaluation.

Open-Weight. Model whose weights can be publicly downloaded. Examples: Llama, Mistral, Apertus. See trend-open-weight-vs-closed.

OWASP LLM Top 10. List of the ten most common LLM security risks, maintained by the OWASP foundation. Version 2.0 in May 2026. Mandatory reading before production use.

P

PCA (Principal Component Analysis). Dimensionality reduction technique. Rarely directly applied to LLMs, sometimes used in embedding analysis.

Pretraining. First training phase of an LLM on a huge text corpus. Very expensive (millions of CHF), seldom undertaken by SMEs themselves. See was-ist-pretraining.

Prompt. Input to an LLM. Covers system prompt (role instruction) and user prompt (concrete query). See prompt-engineering-grundlagen.

Prompt-Injection. Attack in which manipulated inputs get the model to take unintended actions. Still an open issue in May 2026, input filtering and RBAC remain mandatory. See red-teaming-für-ki.

Pseudonymisation. Replacing real names with tokens before a cloud LLM call. Standard for sensitive fiduciary data in May 2026. See anonymisierung-pseudonymisierung.

Q

QLoRA. Quantized variant of LoRA, enables fine-tuning of large models on a workstation GPU.

Quantization. Reducing the precision of model weights (e.g. from 16-bit to 4-bit) to save memory. Standard for local deployment. See was-ist-quantisierung.

Query-Expansion. Technique that extends the original query with synonyms or paraphrases. Improves retrieval recall. See query-expansion-und-rewriting.

R

RAG (Retrieval-Augmented Generation). Architecture that couples LLM answers to your own documents. Standard pattern for company AI in May 2026. See retrieval-augmented-generation.

RAG-Evaluation. Test method for RAG systems: faithfulness, context-recall, answer-relevance. See ki-qualität-kpis-rag.

Rate-Limit. Limit of API calls per second. Strict at all cloud vendors in May 2026; protection against cost blow-up and bots. See firewall-und-crowdsec.

RBAC (Role-Based Access Control). Role-based permission management. Standard for any multi-user AI system. See rbac-und-rechtemanagement.

Reinforcement Learning. Learning paradigm with a reward signal. Basis of RLHF.

Reranker. Model that re-sorts an initial retrieval hit list. Cohere Rerank 3, BGE reranker. See vergleich-rerankers.

S

SAML (Security Assertion Markup Language). Older SSO standard, still common in corporates. See integration-sso-saml-oidc.

SBERT (Sentence-BERT). Embedding model family, predecessor of modern multilingual embeddings.

Self-Attention. Attention variant within the same sequence. Core mechanism of transformer models.

SLM (Small Language Model). Models under roughly 10 billion parameters. Growing strongly in May 2026 (Phi-4, Mistral Small 3, Llama 3.2). Runs locally.

SSO (Single Sign-On). Single login for multiple applications. Mandatory in larger SMEs. See integration-sso-saml-oidc.

Streaming. Model output is sent token by token rather than in one block. Improves perceived latency.

System-Prompt. Instruction to the LLM defining its role, style and limits. Sent before every user query. See was-ist-system-prompt.

T

TIA (Transfer-Impact-Assessment). Check whether a data transfer to a third country (e.g. the USA) is admissible. Mandatory under revFADP. See drittlandtransfer-tia.

TLS (Transport Layer Security). Standard encryption for web traffic. Mandatory for every AI API. See ssl-tls-zertifikate-letsencrypt.

Token. Smallest processing unit of an LLM. A German word equals about 1.3-2 tokens. Model prices bill per token. See was-ist-token.

Tokenizer. Component that splits text into tokens. Each vendor has its own.

Tool-Use. Ability of a model to call external functions. Standard with all premium models in May 2026. See was-ist-tool-use-function-calling.

Transformer. Neural network architecture from 2017 (Vaswani et al.). Basis of all modern LLMs. See was-ist-transformer-architektur.

U

Use-Case. Concrete application of AI in business operations (e.g. dunning, client triage). See industry pages such as ki-für-treuhandbüros.

UTF-8. Standard text encoding. Important on data imports: wrong encoding destroys umlauts and thus retrieval quality.

V

Vector-Embedding. See embedding. The term emphasises the vector nature. See embeddings-und-vektoren.

Vector database. Database that stores and efficiently queries embeddings. May 2026: Qdrant, Weaviate, Milvus, pgvector. See vector-databases-vergleich.

Voice-Cloning. Synthesis of a voice from a few minutes of sample. Compliance-critical: Switzerland and the EU require disclosure. See voice-agent-telefon.

W

WAF (Web Application Firewall). Filter in front of web apps against common attacks. Mandatory in front of any public AI endpoint. See cloudflare-dns-waf.

Webhook. HTTP callback triggered on an event. Standard pattern for AI workflows. See integration-webhooks-ereignisbasiert.

Whisper. OpenAI model for speech-to-text. Still the reference for Swiss German in May 2026 thanks to multilingual training. See tool-whisper.

Workflow. Sequence of automated steps, often orchestrated in tools like n8n or Make. See n8n-workflow-automation.

X

XAI (Explainable AI). Research and practice field of making model decisions understandable. The EU AI Act demands transparency for high-risk systems.

Z

Zero-Shot. Model solves a task with no examples in the prompt. Opposite: few-shot.

Z-Score. Statistical standardisation. Sometimes used as a normalisation step when comparing embeddings.

Why a glossary matters

AI projects often fail on language rather than on technology. When management, IT and external vendors mean different things by "agent", "fine-tuning" or "hallucination", requirements drift, costs rise and the project loses trust. A shared glossary is the cheapest risk reduction in any project.

For Swiss SMEs there is a second point: many terms are Anglo-Saxon and not consistently translated in practice. "Compliance" can mean revFADP, GDPR, FINMA, AMLA or sector-specific supervisory law depending on context. Anyone who does not distinguish buys consulting twice or buys the wrong consulting.

The glossary therefore serves three roles. Management uses it to compare vendor proposals. IT uses it to document architecture decisions. Compliance uses it to name obligations clearly toward supervisors and clients.

How to use the glossary

The glossary is sorted alphabetically. Each entry has one or two sentences of core definition and, where available, a "see slug" pointer to the deeper knowledge page. Slugs are the URL parts of the matching page (for example /wissen/was-ist-mcp).

Three practical uses.

As a clarification tool at project kickoff. Before kicking off your AI project, send the glossary to everyone involved. Collect terms about which there is uncertainty. Resolve them in the first hour of the kickoff. In practice this saves 4-8 hours of misunderstandings later in the project.

As a comparison aid for vendor proposals. When three vendors promise "RAG", "Agentic AI" and "compliance-ready", check with the glossary whether they mean the same thing. A vendor promising "Agentic AI" but actually building only an LLM without tools is not delivering what you buy.

As onboarding material for new staff. As of May 2026 every SME role expects baseline AI knowledge. An internal glossary (this one as a starting point) materially shortens onboarding time.

Roll out the glossary in 5 steps

1. 01Copy the glossary into your internal knowledge base (Confluence, Notion, SharePoint) and extend it with firm-specific terms.
2. 02Distribute to all involved before project kickoff; collect and define unclear terms in a first clarification round.
3. 03In vendor proposals, mark the central terms and compare with the glossary definition. Resolve discrepancies in writing.
4. 04In training, include a 60-minute briefing with the glossary as a mandatory part for everyone working with AI tools.
5. 05In audits and compliance reports, use the glossary vocabulary consistently rather than inventing new terms for each report.

When to use the glossary

The glossary belongs on the table in four situations.

First, in an initial audit of a new vendor or project. Which terms does the vendor use? Do the definitions match those captured here? If not, clarification is mandatory before the contract is signed.

Second, in requirements workshops. "We need a chatbot" means, depending on the person, an FAQ search, a calendar bot, a client agent with tools or a voice phone system. Clarify terms first, then write requirements.

Third, in audits and reviews. Whoever explains to an audit firm or to FINMA which AI is running in the business should use the terms cleanly and consistently. The glossary supplies the vocabulary list.

Fourth, in training. Staff working with AI tools need the basic terms. The glossary is teaching material for a 60-minute briefing.

When the glossary is not enough

The glossary is entry-level vocabulary. It does not replace three things.

First, it does not replace the deeper knowledge pages. Whoever wants to build an MCP server needs the was-ist-mcp page, not the single sentence in the glossary. Anyone implementing RAG needs retrieval-augmented-generation, chunking-strategien-rag and embeddings-und-vektoren. The glossary definitions are trail markers, not directions.

Second, it does not replace legal advice. Terms like "processor", "third-country transfer" or "professional secrecy" appear here in two sentences; their application to a concrete case needs legal counsel or a lawyer memorandum.

Third, it does not replace model comparison. Whoever wants to know whether the current top Claude model or GPT-4o is better for their use-case needs to run an eval with their own data (see eval-frameworks-für-llms). The glossary provides only the vocabulary.

The glossary is maintained continuously. As of May 2026 it contains over 80 terms. We welcome suggestions for additions by email; every entry is added with a citation and date.

Trade-offs

Related topics

Sources

1. EDÖB – Leitfäden zur Bearbeitung von Personendaten mit KI · 2026-05
2. EU AI Act – konsolidierter Text und Stand der Umsetzung 2026 · 2026-05
3. OWASP – Top 10 for Large Language Model Applications v2.0 · 2026-05
4. Model Context Protocol – Specification and SDKs · 2026-05
5. NIST AI Risk Management Framework 1.1 · 2026-04