- Agent
- Generative AI that plans and takes actions in other systems to reach a goal.
- Agent washing
- Relabeling assistants, chatbots or automation tools as agents without real agentic capabilities.
- AI Act
- Regulation (EU) 2024/1689, the EU law on AI, amended by Regulation (EU) 2026/1744.
- AI literacy
- Understanding AI well enough to judge its use, its value and its risks.
- Baseline
- Today’s figures: the current value of the number an AI project should move, agreed with finance before work starts.
- Bias
- Systematic unfairness in outcomes, often learned from historical data.
- Closed model
- A model offered only as a service by its provider.
- Context
- The documents, data and conversation a model can see when it answers.
- Deepfake
- AI-generated audio, image or video that imitates a real person.
- Deployer
- Under the AI Act, an organization that uses an AI system under its own authority.
- EBIT
- Earnings before interest and taxes, a common measure of operating profit.
- Evaluation
- Systematic testing of an AI system’s output quality.
- Excessive agency
- An agent with more autonomy, permissions or access than its task needs.
- Fallback
- A tested alternative, another model or the manual process, for when an agent or model is unavailable.
- Feasibility
- How achievable an AI use case is: the state of the data, how much the process must change, the skills available and the risk involved.
- Fine-tuning
- Extra training of a model on specific examples.
- Gateway
- One control point between your applications and AI models, which lets you switch models.
- Generative AI
- AI that creates new text, images, code or sound.
- Grounding
- Connecting a model to trusted sources so that it answers from them. A common technique is retrieval-augmented generation (RAG).
- Hallucination
- Confident but false output from an AI model.
- High risk
- Under the AI Act, uses such as recruitment or credit scoring that carry strict obligations.
- Hub and spoke
- A way to organize AI in which a central team sets standards, platforms and controls, and business units build and use AI.
- Human in the loop
- A design in which a person reviews or approves AI output before it takes effect.
- Human-agent ratio
- The balance between work done by agents and by people, set by where human contact and judgment matter.
- Large language model
- A generative AI model trained on large amounts of text to predict the next word.
- Least privilege
- Giving an agent only the tools, data and permissions its task needs.
- Machine learning
- AI that learns patterns from examples instead of following written rules.
- Observed exposure
- Anthropic’s measure of how much of an occupation’s work AI covers in real use.
- Open-weight model
- A model whose trained internal settings (weights) can be downloaded and run in your own environment or with a hosting partner.
- Operating model
- The way a company organizes its AI work, for example central, by business unit, or hub and spoke.
- Prompt
- The instruction or question given to a model.
- Prompt injection
- Input that changes a model’s behavior in unintended ways; indirect when hidden in content the model reads.
- Provider
- Under the AI Act, an organization that develops an AI system and puts it on the market under its own name.
- Shadow AI
- AI used in an organization outside any approved policy.
- Stop rule
- A condition, agreed in advance, under which a project is stopped.
- Time horizon
- The length of tasks an AI agent can complete at a given success rate, in human working time.
- Token
- A whole word or a fragment of a word; the unit in which models read, write and charge.
- Transformer
- The neural network design, published in 2017, on which today’s language models build.
- Value lever
- One of the four ways AI creates value: efficiency, growth, quality or risk.
