Dictionary
AI, in plain English
The terms that actually come up when a business is deciding what to do about AI — what each one means, and why it matters commercially. No jargon defined with more jargon.
35 terms
Foundations
- Agent
- An LLM given tools and a goal, which takes multiple steps on its own rather than answering a single question.
- AI
- Software that performs tasks normally requiring human judgement — in practice today, almost always a language model.
- API
- A defined way for software to talk to other software — how AI gets connected to the systems you already run.
- Chatbot
- A conversational interface that answers questions but does not change anything in your systems.
- Context window
- The maximum amount of text a model can consider at once — everything it knows for that request must fit inside it.
- Fine-tuning
- Further training a model on your own examples to change its behaviour — usually not the answer you are looking for.
- Large language model (LLM)
- A model trained to predict the next piece of text, which turns out to be enough to summarise, draft, classify and reason about language.
- Prompt
- The instructions and context you send to a model — in production, usually assembled by code rather than typed by a person.
- System prompt
- Standing instructions given to a model on every request, defining its role, tone and limits.
- Token
- The unit models read and write, and the unit you are billed in — roughly ¾ of a word in English.
- Training data
- The text a model learned from — which determines what it knows, and when its knowledge stops.
Reliability
- Approval gate
- A checkpoint where an AI-proposed action requires human sign-off before it executes.
- Eval
- A repeatable test of whether your AI system is actually getting things right — the difference between improving it and guessing.
- Failure taxonomy
- A written, shared vocabulary for the specific ways your AI system goes wrong.
- Guardrails
- The constraints around a model that keep its behaviour inside acceptable limits.
- Hallucination
- When a model produces confident, fluent output that is simply wrong — the failure mode that makes AI risky in business processes.
- Human in the loop
- Designing the workflow so a person reviews or approves at the points where being wrong is expensive.
- Prompt injection
- An attack where text the model reads contains instructions that hijack its behaviour.
- Regression test
- A saved set of cases re-run on every change, so an improvement in one place cannot silently break another.
Operations
- Cost per call
- What a single AI interaction costs you — the number that decides whether a feature is viable at scale.
- Inference
- Running a model to get an answer — the thing you pay for, as opposed to training it.
- Latency
- How long the system takes to respond — often the difference between a feature people use and one they abandon.
- Model routing
- Choosing which model handles a request at run time, usually by difficulty, cost or availability.
- Vendor lock-in
- Depending on one AI provider so heavily that leaving becomes impractical.
Retrieval
- Embedding
- A numeric representation of meaning, letting a computer find text that is similar in sense rather than in wording.
- Hybrid search
- Combining semantic and keyword search, because each finds things the other misses.
- RAG (retrieval-augmented generation)
- Looking up relevant material from your own documents and giving it to the model, so answers are grounded in your data rather than its training.
- Vector database
- A store built to find the nearest [embeddings](/dictionary/embedding) quickly, so semantic search stays fast as your data grows.
Agents
- MCP (Model Context Protocol)
- An open standard for connecting AI agents to data sources and tools, so integrations are not rebuilt per vendor.
- Structured output
- Forcing a model to answer in a defined shape — JSON matching a schema — so software can rely on it.
- Tool calling
- Letting a model invoke functions you define — how an AI system does anything beyond producing text.
Delivery
- Pilot
- A limited production deployment with real users, run to find out what actually happens.
- Proof of concept
- A small build that answers a specific question about feasibility — not a small version of the finished product.
Governance
- Data governance
- Knowing what data you hold, who may see it, and where it is allowed to go — the prerequisite most AI projects skip.
- PII (personally identifiable information)
- Data that identifies a person — which constrains what you may send to a model and where.
Next step
Tell us what you’re trying to build
Most engagements start with a fixed-price audit, so the first thing you buy is a decision rather than a commitment.