AI Terminology
This is the shared vocabulary for the rest of the playbook, drawn from JMAN's internal AI Handbook so the same definitions and analogies hold whether you are in a design review or a client workshop.
Each term is defined twice: once precisely enough for an engineer implementing it, once plainly enough to say out loud in a scoping conversation. The business relevance column is the part most often missing from a glossary - it says when the term is the right answer and, more usefully, when it is not. Read the column that matches the room you are in.
Terms are grouped by function. Core Concept covers the AI/ML/GenAI hierarchy, Model & Data covers what models are built from and consume, Technique & Pattern covers how systems are built and prompted, and Quality & Risk covers how output is evaluated and governed. Every core topic has a deep-dive page linked from its name. Architectural patterns are defined here and covered in full in AI Architecture rather than duplicated.
Core Concept
| AI Term | Technical Definition for AI Engineers | Simplified Explanation for Consultants | Business Relevance / When to Use | Example |
|---|---|---|---|---|
| Artificial Intelligence | The broad field of systems performing tasks that normally require human judgement: language understanding, pattern recognition, prediction, planning. Includes non-learning approaches such as rule-based expert systems and search algorithms. The term dates to 1956. | The entire automobile industry. Everything else on this page is one vehicle or component within it. | Use when setting scope at the very start. Most of the time a narrower term is the accurate one, and using it avoids overselling. | A demand forecast, a contract extraction pipeline, and a routing agent are all AI, and share almost no implementation. |
| Machine Learning | A subset of AI where behaviour is learned from data rather than explicitly programmed. Covers supervised, unsupervised, and reinforcement learning. | The systems that learn from examples instead of being handed the rules. | Reach for ML when the rules are too numerous or too fuzzy to write down and you have labelled history to learn from. | Predicting which accounts will churn from two years of billing and support history. |
| Generative AI | A subset of ML that produces new content - text, code, images, audio - rather than a label or a number. Built on large language models and diffusion models. See AI Fundamentals for how the three nest. | The part of AI that writes, drafts, and summarises rather than scoring or classifying. | Use when the output is language or content. Do not use it where a deterministic calculation or a classical model is the right answer. | Drafting a first-pass investment memo from a data room. |
Model & Data
| AI Term | Technical Definition for AI Engineers | Simplified Explanation for Consultants | Business Relevance / When to Use | Example |
|---|---|---|---|---|
| Large Language Model (LLM) | A transformer trained on large text corpora by next-token prediction, then instruction-tuned. Knows only what was in its training data up to its cutoff, with no awareness of your systems unless you supply the context. | The engine. An incredibly well-read employee who has not seen today's emails or any internal document. | The default building block for any language task. Size the model to the task rather than defaulting to the largest. | Summarising 200 supplier contracts into a comparable clause table. |
| Foundation Model | A model pretrained on broad data and adapted to many downstream tasks by prompting or fine-tuning, rather than trained per task. See Large Language Models. | One general-purpose model you adapt, instead of a new model per problem. | Start here. Train or fine-tune something task-specific only once prompting and retrieval have demonstrably failed. | One model deployment sitting behind extraction, summarisation, and classification. |
| Token | The unit of text a model processes, roughly three-quarters of a word in English. Pricing, context limits, and latency are all denominated in tokens. | The meter. Everything the model reads and writes is billed by it. | Token maths is how you forecast cost before a build rather than after the first invoice. | A 40-page contract is roughly 20,000 tokens, so re-sending it every turn becomes the largest cost driver. |
| Context Window | The maximum tokens a model can attend to in a single call: system prompt, history, retrieved documents, tool results, and the output it generates. Fixed per model. | The desk. Everything the model needs has to fit on it at once, and once it is full, things fall off. | Drives retrieval and memory design. A bigger window is not a substitute for choosing what goes into it. | Ten retrieved policy documents plus a long history push the original question out of scope. |
| Embedding | A dense vector representation of text where semantic similarity maps to geometric proximity. Produced by a dedicated embedding model, not the generative one. | Turning meaning into coordinates, so "car" and "automobile" land next to each other. | The mechanism behind semantic search and RAG retrieval. | Matching a question about notice periods to a clause headed Termination. |
| Vector Database | Stores embeddings and runs approximate nearest-neighbour search over them, with indexing strategies that trade recall against latency. See Embeddings. | A librarian who understands meaning, not just keywords. | Needed whenever retrieval has to work on meaning rather than exact terms. | Searching for CEO and still returning documents that only say Chief Executive Officer. |
Technique & Pattern
| AI Term | Technical Definition for AI Engineers | Simplified Explanation for Consultants | Business Relevance / When to Use | Example |
|---|---|---|---|---|
| Prompt / Prompting | The instruction set given to a model: system prompt, user prompt, examples, and output schema. Prompts are production assets and should be versioned and evaluated like code. | How you ask. The same model gives very different answers depending on the phrasing. | The cheapest lever on output quality. Exhaust it before reaching for fine-tuning. | Adding three worked examples cuts format errors on an extraction task more than switching model does. |
| Fine-Tuning | Further training a pretrained model on a task-specific dataset to change behaviour, style, or domain handling beyond what prompting achieves. | Sending the employee on a specialist course rather than writing better instructions. | Only once prompting and retrieval have failed, and you have hundreds of high-quality examples. It does not reliably add knowledge. | Locking a consistent house tone across thousands of generated summaries. |
| RAG (Retrieval-Augmented Generation) | Retrieve relevant context at query time and supply it in the prompt, so output is grounded in current or private data without retraining. | Letting the employee look up the company files before answering. | The default answer to "the model does not know our data". Cheaper and more auditable than fine-tuning. | An HR assistant answering from the current policy set with the clause cited. |
| Tool / Tool Calling | External functions exposed to the model with typed schemas. The model selects a tool and its arguments, your code executes it and returns the result. | Giving the employee access to the company's systems. | Turns a model from something that writes into something that does. | Looking up live order status in the CRM instead of guessing from training data. |
| Agent | A model in a loop with tools and a goal, choosing the next action from results so far rather than producing one answer and stopping. | An employee who can run a task end to end, not just answer a question about it. | Use when the number of steps cannot be known in advance. If it can, a fixed pipeline is cheaper and more predictable. | Investigating a failed payment: check the gateway, check the account, draft the response. |
| Multi-Agent System | Several specialised agents coordinating under an orchestration topology, each with a narrower toolset and prompt than a single general agent. | A project team where each person owns one part of the job. | Use when one agent's prompt and toolset have grown too broad to stay reliable. It adds coordination cost. | A research agent, an analysis agent, and a reviewer agent producing a market summary. |
| Orchestration | The coordination layer sequencing prompts, retrieval, tool calls, and hand-offs, including retries and state management. See Multi-Agent Systems. | The workflow that decides what happens in what order. | Where reliability actually lives in production. Most failures are orchestration failures, not model failures. | Retrying a failed extraction with a narrower prompt before escalating to a human. |
| Model Context Protocol (MCP) | An open standard for exposing tools and data sources to AI systems through a consistent interface, so integrations are reusable across models and clients. | The USB-C connector of the AI world. | Cuts per-integration build cost when several assistants need the same underlying systems. | One MCP server for the data warehouse, used by an internal assistant and a client-facing app. |
| Memory | Retained state across turns or sessions: working state inside the context window, and durable facts written to and retrieved from a store. | Remembering a customer's preferences from previous conversations. | Needed for continuity and personalisation. Also the main place stale or sensitive data quietly accumulates. | An assistant recalling that the client reports in euros without being told each session. |
Quality & Risk
| AI Term | Technical Definition for AI Engineers | Simplified Explanation for Consultants | Business Relevance / When to Use | Example |
|---|---|---|---|---|
| Accuracy | How closely output matches ground truth. For generative systems this is measured with task-specific evaluations rather than a single score. See Hallucination & Groundedness. | How often it is right, measured rather than asserted. | Define the measure before the build. Without an evaluation set, "it seems good" is the only verdict available. | An extraction task scored on field-level exact match against 200 hand-checked documents. |
| Hallucination | Fluent output unsupported by the input or by fact, a direct consequence of next-token prediction having no model of truth. | A confident employee giving an answer they are actually unsure about. | The main reason a demo passes and production fails. Mitigate with retrieval, tools, validation, and review. | Citing a policy clause number that does not exist. |
| Explainability | The degree to which an output or decision can be traced and justified by a human, as opposed to an opaque result. | Being able to show why the system said what it said. | Required wherever a decision is challenged, audited, or regulated. | Showing the three retrieved passages an answer was drawn from. |
| Bias | Systematic skew in output caused by unrepresentative training data or prompt framing, producing inconsistent treatment across groups or inputs. | The system inheriting the imbalances of whatever it was trained on. | Test for it wherever output affects people. It is a delivery risk, not only an ethical one. | A screening assistant rating otherwise identical CVs differently by name. |
| Groundedness | The extent to which output is directly supported by the retrieved source material, measurable by checking each claim against the supplied context. | Whether the answer is actually backed by the documents. | The practical quality metric for a RAG system. Higher groundedness, fewer hallucinations. | Every sentence in a generated summary traced back to a source paragraph. |
| Guardrails | Programmatic constraints around a model: input validation, output filtering, schema enforcement, and policy checks applied at runtime. | The limits on what the system is allowed to do or say. | Non-negotiable for anything client-facing. Design them before launch, not after an incident. | Blocking a response containing personal data, and refusing out-of-scope questions. |
The Car Analogy
The AI Handbook frames the whole stack as a car. It is the fastest way to give a non-technical audience the shape of the system in one picture.
The analogy carries further than most: an engine with no fuel line is an LLM with no retrieval, and a car that drives itself still needs somewhere to be told to go.
Deep Dives
One page per core topic, in the order they build on each other.
The architectural patterns defined above - RAG, agents, multi-agent orchestration, MCP, fine-tuning, and guardrails - are covered in AI Architecture rather than repeated here.
On this page
Title
AI Terminology