Glossary: AI and data terms explained simply

An AI agent is an AI system that does more than generate text. It independently plans and executes multi-step tasks, for example retrieving data, checking it and entering it into a downstream system. In the mid market, agents take over recurring process steps under clearly defined rules called guardrails.

Learn more: AI Agents & Delivery

An API (Application Programming Interface) is a defined interface through which two systems exchange data in a structured way. It allows ERP, MES, CRM and analytics tools to work together automatically and in real time, instead of relying on manual exports.

Bias refers to a systematic distortion in data or AI models that leads to one sided or unfair results. It is usually caused by imbalanced training data. In industry, bias can quietly distort quality or forecasting models and requires testing procedures to keep it in check.

Business Intelligence covers the methods and tools used to collect, prepare and analyze company data in dashboards and reports. The goal is reliable, fast decisions based on current, verified metrics.

Learn more: Reliable Data Foundation & BI

A sovereign cloud is a cloud environment whose data is verifiably processed within the EU and under European law (GDPR). For the mid market, it enables scalable AI without giving up control over sensitive data.

The context window is the maximum amount of text a language model can process at once, measured in tokens. It determines how much document content or conversation history the model can take into account at the same time.

A CRM system (Customer Relationship Management) manages all customer information: offers, contracts, service history. It is a central data source whenever sales and service metrics need to feed into shared reporting.

Data governance is the set of rules that defines who may use, change and protect which data, and how. It covers roles, responsibilities, quality standards and documentation, and forms the basis for audit ready and AI ready data.

Learn more: Reliable Data Foundation & BI

A data lake is a central store for large volumes of raw data in any format, both structured and unstructured. It allows machine, sensor and business data to be collected before being prepared for analysis or AI.

Data maturity describes how well a company collects, consolidates, governs and uses its data. It ranges from an early stage with scattered isolated solutions to a scale ready organization. The maturity level determines which AI use cases are realistic.

Learn more: Start the Data-Readiness Check

A data silo is an isolated set of data that is only accessible to one department or system. Silos prevent a shared overview, lead to contradictory figures and are one of the most common obstacles to AI projects.

Data sovereignty means a company retains full control over where its data is stored, who can access it and how it is processed, at all times. This is especially relevant for cloud and AI services when it comes to securing compliance and trade secrets.

A data warehouse is a central database where cleaned and structured data from various sources is brought together for analysis. It is the classic foundation for reliable BI reporting.

Embeddings are numerical representations of text, images or other content as vectors of numbers. They make it possible to find content that is similar in meaning and form the basis for semantic search and RAG systems.

An ERP system (Enterprise Resource Planning) manages core business processes such as purchasing, production, inventory and finance. As the leading data source, it is usually the starting point when building a single point of truth.

ETL stands for Extract, Transform, Load, the process of extracting data from source systems, cleaning and converting it, and loading it into a target system such as a data warehouse. ETL pipelines ensure that analyses are based on consistent data.

Feature engineering is the process of turning raw data into meaningful features that a machine learning model can learn from. Good features are often more decisive for model quality than the algorithm itself.

Fine tuning is the process of retraining a pretrained AI model on your own domain specific data. This allows a general language model to be adapted to a company terminology and tasks, as an alternative or complement to RAG.

The General Data Protection Regulation (GDPR) is the EU wide framework for handling personal data. It requires transparency, purpose limitation and security, which every AI and data solution must meet from the outset.

Generative AI refers to models that create new content: text, images, code or summaries. In business use, it can support tasks such as drafting reports and proposals or researching large document collections.

Guardrails are technical and organizational boundaries that limit the behavior of AI systems: permitted data sources, validation of outputs and escalation rules. They make AI agents controllable and safe to operate.

A hallucination is a factually incorrect but convincingly worded answer from an AI model. Countermeasures include grounding responses in verified sources (RAG), validating outputs and using guardrails. This is essential before any business critical use.

ISO 27001 is an international standard for information security management systems (ISMS). Certification demonstrates structured security processes and is often a requirement in supply chains and when handling sensitive data.

A KPI (Key Performance Indicator) is a performance metric that makes progress toward a goal measurable, for example cycle time, scrap rate or on time delivery. Clear KPIs are the prerequisite for meaningful dashboards and AI use cases.

A large language model is an AI model trained on huge amounts of text that understands and generates language. LLMs are the foundation of modern assistants and AI agents used for summarizing, answering and structuring information.

Machine learning is a branch of AI in which systems learn patterns from data instead of being explicitly programmed. In industry, it underlies things like forecasts, anomaly detection and quality classification.

An MES (Manufacturing Execution System) controls and records production in real time, from machine data to order progress. As a data source, it supplies the production data needed for industrial AI and shop floor dashboards.

MLOps brings together the practices needed to reliably deploy, monitor and update AI models. It applies principles from software development, such as automation, versioning and monitoring, to the lifecycle of models.

Model drift describes the decline in model quality over time as real world data changes. Continuous monitoring detects drift early so a model can be retrained in time, which is essential for long term operation.

Predictive analytics uses historical data and statistical or ML models to forecast future events, such as demand, maintenance needs or failures. It shifts decisions from reactive to forward looking.

A prompt is the input instruction given to an AI model. Prompt engineering is the craft of phrasing these instructions so the model delivers precise, reliable results, an important lever for quality without changing the model.

RAG combines a language model with a search across verified company documents. Relevant content is retrieved and fed into the answer, so AI responds based on your own current sources, with far fewer hallucinations.

Learn more: Article: What is RAG?

Reporting is the regular, structured preparation of metrics for decision makers. Automated reporting replaces manual spreadsheet work with reports that are always current and consistent, built on a reliable data foundation.

ROI (Return on Investment) measures the ratio of benefit to resources invested. In AI and data projects, a clear ROI definition helps prioritize the most valuable use case first and makes investments easier to justify.

Shadow IT refers to tools and data flows used without the knowledge of the IT department, such as local spreadsheets or private cloud services. It puts data quality, security and compliance at risk and is a typical symptom of missing governance.

A single point of truth is a central, consolidated data foundation that all systems and analyses refer to. It removes contradictory figures and is the foundation for reliable reporting and AI.

Learn more: Case: Single Point of Truth

Structured data is organized in clear tables and fields, such as ERP records. Unstructured data is not, such as emails, PDFs and images. Modern AI can use both, provided they are made accessible and analyzable.

TISAX (Trusted Information Security Assessment Exchange) is an assessment and exchange standard for information security in the automotive industry. Suppliers often need it as proof in order to win contracts.

A token is the smallest unit of processing for a language model, roughly a piece of a word. Tokens determine context size and cost, since inputs and outputs are measured and billed in tokens.

A use case is a concrete application with a clear goal, benefit and measurable result. In an AI context, a clean use case definition determines whether a project creates real economic value.

Learn more: AI & Data Strategy

A vector database stores embeddings and finds content that is similar in meaning in a fraction of a second. It is the backbone of RAG systems and semantic search across large document collections.

Vendor lock in is a strong dependency on a single provider that makes switching expensive or difficult. Open standards, portable data and interchangeable models reduce this risk, an important factor when choosing AI platforms.