The most important terms around artificial intelligence, digital transformation and the energy transition -- clearly explained.
25 terms from A to Z
AI systems that autonomously pursue goals, make decisions and execute multi-step tasks without constant human guidance. Unlike traditional AI applications that respond to individual queries, agentic systems actively plan and dynamically adapt their approach. In enterprise practice, this means AI agents that take over entire workflows, not just individual steps.
An autonomous software system that perceives its environment, makes decisions and takes actions to achieve specific goals. AI agents differ from simple chatbots through their ability to use tools, create plans and independently execute multi-step tasks. In enterprises, they are increasingly deployed for research, data analysis and process automation.
Hypothetical AI systems that can understand and solve any intellectual task that a human can accomplish -- including reasoning, learning and creative problem-solving across domains. AGI does not exist today; current systems like LLMs are specialised (narrow) AI. The debate around AGI is central to long-term AI strategies and regulation.
The systematic design and optimisation of the context available to an AI model during processing. This includes system prompts, example data, document selection and instruction design. Context Engineering goes beyond Prompt Engineering because it shapes the entire information architecture around the model -- not just the individual input.
AI agents deployed as full team members with defined roles and responsibilities within organisations. The innobu Digital Colleagues framework gives AI agents concrete job titles -- Architect, Strategist, Analyst -- and integrates them into existing governance structures. The approach solves the most common problem in AI projects: the technology works, but nobody agrees on what it should actually do.
A virtual representation of a physical object, process or system that is synchronised with real-time data. In the energy sector, digital twins are used to model grid infrastructure, simulate loads and plan maintenance proactively. The value comes from connecting sensor technology, data platforms and decision support.
The holistic description of an organisation's structures -- business processes, IT systems, data flows and technology infrastructure. Enterprise Architecture creates transparency about how individual systems interact and provides the decision basis for investments in technology and transformation. Without EA, every large programme becomes a blind flight.
The European AI regulation (Artificial Intelligence Act) that establishes a risk-based regulatory framework for the deployment of AI in the EU since 2024. Systems are classified into risk categories -- from minimal to unacceptable. High-risk systems are subject to strict requirements for transparency, data quality and human oversight. For enterprises, the EU AI Act means concrete compliance obligations that must be implemented now.
AI systems that can create new content such as text, images, code or music, rather than just analysing or classifying existing data. The best-known examples include ChatGPT, Claude and Midjourney. Generative AI is typically based on large pre-trained models (foundation models) and has fundamentally changed the discussion around AI in enterprise contexts since 2023.
The phenomenon where AI models generate plausible-sounding but factually incorrect or entirely fabricated information. Hallucinations are not a bug but a systemic characteristic of generative models that operate on probabilities. For enterprise use, this means every AI output needs verification mechanisms. RAG, grounding and human oversight are the most important countermeasures.
Large neural networks trained on extensive text data that can understand and generate natural language. LLMs such as GPT-4, Claude and Gemini form the foundation of the current generative AI wave. They work with tokens, use the Transformer architecture and are optimised for specific applications through fine-tuning and Reinforcement Learning from Human Feedback (RLHF).
A subfield of artificial intelligence where algorithms learn from data and recognise patterns without being explicitly programmed. The three main categories are supervised learning (with labelled data), unsupervised learning (pattern recognition in unstructured data) and reinforcement learning (learning through reward). Machine learning is the foundation of virtually all modern AI applications.
An open standard that enables AI models to access external data sources and tools in a structured way. MCP was initiated by Anthropic and works like a universal adapter: instead of building a custom integration for each data source, MCP provides a standardised interface. For enterprises, this reduces integration effort and improves interoperability between AI systems.
An architecture in which multiple AI agents collaborate to solve complex tasks. Each agent has specialised capabilities and a defined role; coordination happens via protocols or an orchestration agent. Multi-agent systems enable the decomposition of complex problems into manageable sub-tasks and are the foundation for scalable AI operating models.
A subfield of AI concerned with the processing, understanding and generation of human language. NLP encompasses tasks such as text classification, sentiment analysis, named entity recognition, machine translation and question answering. Modern LLMs have revolutionised NLP -- tasks that required specialised models just a few years ago are now solved by generic language models as a side effect.
The art and science of formulating input instructions (prompts) for AI models to achieve optimal results. This includes techniques such as chain-of-thought prompting, role instructions, format specifications and few-shot examples. Good prompt engineering can drastically improve the quality of AI outputs without needing to modify the model itself.
An attack vector where malicious instructions are injected into the input of an AI model to manipulate its behaviour. In indirect prompt injection, instructions are hidden in data sources that the model processes -- such as websites or documents. Prompt injection is one of the biggest security challenges for AI systems in enterprise deployment.
An architecture that enriches AI models with external knowledge by retrieving relevant documents before generating a response. Instead of relying solely on training knowledge, the system searches a knowledge base and inserts relevant information into the context. RAG reduces hallucinations, keeps answers current and enables the use of proprietary enterprise data without retraining the model.
A framework for developing and deploying AI systems with consideration for ethical principles such as fairness, transparency, privacy and non-discrimination. Responsible AI goes beyond compliance: it requires proactive measures such as bias testing, explainability of decisions and clear accountability structures. In the EU context, it is increasingly becoming mandatory through the AI Act.
The central communication unit of an intelligent metering system (iMSys) that enables secure data exchange between electricity meters, market participants and grid operators. Smart Meter Gateways must be BSI-certified and form the technical foundation for the German smart meter rollout. They are the prerequisite for time-variable tariffs, Section 14a control and the integration of decentralised energy generation.
Section 14a of the German Energy Industry Act (Energiewirtschaftsgesetz) regulates the grid-oriented control of controllable consumption devices such as heat pumps, wallboxes and battery storage. Grid operators can temporarily dim power to 4.2 kW during bottlenecks, but must grant reduced grid fees in return. Implementation requires Smart Meter Gateways, new control boxes and significant adaptations to IT systems and processes at grid operators.
The smallest processing unit in language models -- a word, word fragment or special character into which text is broken down. An English word typically consists of 1 to 2 tokens. Tokens determine both the input and output length of a model and are the basis for cost calculation when using APIs. A model's context length (e.g. 200,000 tokens) determines how much information can be processed simultaneously.
A neural network architecture introduced in 2017 that is based on the self-attention mechanism and forms the foundation of virtually all modern language models. Transformers can capture relationships between distant words in a text and can be trained in parallel. GPT, Claude, Gemini and Llama are all based on this architecture. The name of the paper: "Attention Is All You Need".
A software development approach where developers describe their intentions in natural language and AI generates the code. The term was coined in 2025 by Andrej Karpathy. Vibe Coding drastically lowers the entry barrier for software development and enables domain experts without deep programming skills to build functional prototypes and applications. Quality assurance remains the critical human task.
Techniques where AI models can solve tasks without (zero-shot) or with only a few (few-shot) examples. In zero-shot, the model receives only a task description; in few-shot, a few example pairs (input/output) are provided. This capability is a key advantage of large language models over traditional machine learning, which requires large labelled datasets.