Glossary
AI terms, in plain English
The words you meet when buying or building AI, defined without jargon.
- AI agent
- A system that follows a procedure, uses tools such as your CRM or inbox to do work, makes decisions within set limits, and escalates anything outside those limits to a person.
- Audit trail
- A complete, time-stamped record of what an AI system did and why, so any decision can be reconstructed and reviewed after the fact.
- Context window
- The amount of text a language model can consider at once. Larger windows let a model work with more of your documents in a single request.
- Deterministic backbone
- The plain-code part of an automation that handles routing, retries, logging and the audit trail, with AI used only at the judgement steps. It keeps the system reliable when a model has a bad day.
- DPIA
- Data protection impact assessment: a documented check of privacy risk that UK GDPR requires before higher-risk processing goes live. We complete one before any such AI workflow is deployed.
- Drift
- The gradual decline in accuracy as the real world moves away from what a system was set up for. Monitoring catches it early.
- Embedding
- A numerical representation of text that lets a system find related content by meaning rather than exact words. It is the basis of most retrieval systems.
- Eval
- A repeatable test that measures how well an AI system performs against known cases, so accuracy is measured rather than assumed before go-live.
- Exception queue
- Where an automation sends anything it is not sure about, so a person reviews the edge cases instead of the system acting on a guess.
- Fine-tuning
- Further training of a model on your own examples so it performs better on a specific task. Often unnecessary when good prompting and retrieval will do.
- GEO
- Generative engine optimisation: structuring content so AI answer engines such as ChatGPT, Claude and Perplexity cite it, alongside traditional search optimisation.
- Guardrails
- Rules and checks that keep an AI system inside safe limits, such as requiring human approval before money or messages leave the business.
- Hallucination
- When a model states something that is not supported by its sources. Grounding answers in your documents and requiring citations reduces it.
- Human in the loop
- A design where a person approves consequential actions before they happen, so autonomy is earned on evidence rather than assumed.
- Inference
- Running a trained model to get an output, such as a classification or a drafted response. Each model call is an inference.
- Latency
- The time a system takes to respond. It matters most for live interactions such as voice and chat.
- Least privilege
- Giving a system or person only the access needed for the task and no more, so the damage from any mistake or breach is contained.
- LLM
- A large language model: the kind of AI that reads and writes text, used for classification, extraction, drafting and reasoning over documents.
- MCP
- Model Context Protocol: a standard that lets an AI agent connect to tools and data sources in a consistent way across different applications.
- Observability
- The ability to see what an AI system is doing in production: what it ran, what it decided, and what it refused to decide.
- Orchestration
- Coordinating several steps, tools and model calls into one reliable workflow, with the order, the retries and the approvals defined.
- Prompt
- The instruction given to a model. Clear, specific prompts with examples produce more reliable results than vague ones.
- Prompt injection
- An attack where hidden instructions inside content try to make an AI system act against its rules. Careful input handling and guardrails reduce the risk.
- RAG
- Retrieval augmented generation: the model answers using your documents, fetched at the moment of the question. Good implementations also cite the source of each answer.
- Role-based access control (RBAC)
- Granting access according to a defined role rather than to individuals, so permissions stay consistent and auditable.
- Structured extraction
- Pulling specific fields, such as dates, amounts and names, out of unstructured documents and into a consistent format a system can use.
- Token
- A chunk of text, roughly a few characters, that models read and write in. Usage and cost are usually measured in tokens.
- Vector database
- A store of embeddings that lets a retrieval system quickly find the content most related to a question by meaning.
- Workflow automation
- Connecting AI and plain code to the tools your team uses so repetitive work is classified, routed, drafted and updated without a person doing it by hand.