Glossary
AI, without the jargon
Thirty-three terms you will meet in vendor pitches and project plans, each in a sentence or two.
- Agentic AI
- The approach of building systems from AI agents that act with some autonomy, under limits and review set by people. Read more →
- AI agent
- Software that plans steps, uses tools such as your CRM or email, checks its own result and finishes a task, rather than only answering a question. Read more →
- Air-gapped
- A system with no connection to outside networks. Used where data must never leave the premises. Read more →
- Barge-in
- The ability of a voice agent to stop talking the moment the caller starts, as a person would. Read more →
- Computer-use agent
- An agent that operates software through the screen, clicking and typing, for systems that have no API. Read more →
- Context engineering
- Deciding what information a model receives for each task: the right documents, tools and instructions, no more and no less. Read more →
- Distillation
- Training a small model to copy the behaviour of a large one on a specific task, so it can run faster and cheaper. Read more →
- Document AI
- Software that reads documents such as invoices, claims and contracts, pulls out the data and checks it before filing. Read more →
- Embedding
- A list of numbers that captures the meaning of a piece of text or an image, so similar things can be found by search. Read more →
- Evaluation (evals)
- A repeatable test of an AI system on real examples with known good answers, run on every change. Read more →
- Fine-tuning
- Training an existing model further on your own examples so it learns your format, vocabulary and judgement. Read more →
- Golden set
- A fixed set of real cases with agreed correct answers, used to measure whether an AI system is improving. Read more →
- Guardrails
- Checks around a model that block unsafe, off-topic or policy-breaking input and output. Read more →
- Hallucination
- A confident answer from a model that is not supported by facts or sources. Read more →
- Human in the loop
- A design in which a person reviews or approves certain AI actions before they take effect. Read more →
- Inference
- Running a trained model to produce an answer. Inference engineering makes that fast and cheap. Read more →
- LLM
- Large language model: a model trained on vast text that can read, write and reason in language. Read more →
- LLMOps
- The practices for running language-model systems in production: evaluation, monitoring, versioning and cost control. Read more →
- Model Context Protocol (MCP)
- A standard way for AI applications to connect to tools and data sources, so an agent can use them without custom wiring each time. Read more →
- Multimodal model
- A model that handles more than one kind of input, such as text, images, audio and video. Read more →
- On-device AI
- Models that run on a phone, laptop or edge device, with no cloud round trip. Read more →
- Open-weight model
- A model whose trained weights are published, so you can run and adapt it on your own hardware. Read more →
- Prompt injection
- An attack in which hidden instructions in a document or message try to make an AI do something its owner did not intend. Read more →
- Quantisation
- Storing a model's numbers at lower precision so it uses less memory and runs faster, usually with a small loss in quality. Read more →
- RAG
- Retrieval-augmented generation: the system finds relevant passages in your documents and gives them to the model, so answers come from your facts. Read more →
- Red-teaming
- Deliberately attacking an AI system to find failures before real users or attackers do. Read more →
- Reinforcement learning (post-training)
- Training a model with rewards for good outcomes, such as a correct answer or a successful tool call, to improve behaviour beyond what examples alone teach. Read more →
- Sovereign AI
- AI whose data, models and operation stay inside a country or legal region. Read more →
- Speculative decoding
- A speed-up in which a small model drafts several words ahead and a large model verifies them together. Read more →
- Synthetic data
- Examples generated by a model to fill gaps in real data, always reviewed and kept out of the test set. Read more →
- Token
- A small piece of text, roughly three-quarters of a word, that models read and are priced by. Read more →
- Vector database
- A database built to find items by similarity of meaning, used in retrieval systems. Read more →
- Word error rate (WER)
- The share of words a speech recogniser gets wrong compared with a correct transcript. Read more →
Still unsure?
Missing a term, or lost in the jargon?
Ask in plain words. We reply in plain words, and good questions are added to this glossary.
- A plain-language answer
- Whether the idea matters for your project
- Links to deeper reading