The AI glossary for the person signing the contract.

39 terms that show up in AI proposals, each defined in plain language — then what it actually changes about a buying decision. Written for buyers, not for engineers.

Also available as plain Markdown.

Core concepts

Large language model (LLM)Token Context window Prompt Fine-tuning AI agent Tool use Model Context Protocol (MCP)Multimodal model Structured output

Retrieval & knowledge

Retrieval-augmented generation (RAG)Embedding Vector database Chunking Hybrid search Re-ranking Grounding Knowledge cutoff

Quality & evaluation

Hallucination Evaluation set LLM-as-judge Drift Observability Human in the loop

Cost & performance

Inference cost Prompt caching Distillation Open-weight model Frontier model Latency

Security & governance

Prompt injection Least-privilege retrieval No-training clause Data residency Data processing agreement (DPA)

How the work is bought

Forward deployed engineer (FDE)AI readiness Proof of concept (POC)Payback period

Core concepts

Large language model (LLM): A large language model is a neural network trained on very large amounts of text to predict what token comes next, which turns out to be enough to summarise, translate, classify, write code and follow instructions. It has no memory between calls and no access to your systems unless you give it one.
Token: A token is the unit a language model reads and writes — roughly three-quarters of an English word, so 1,000 tokens is about 750 words. Model pricing, context limits and latency are all measured in tokens rather than words or characters.
Context window: The context window is the maximum number of tokens a model can consider in a single call, counting the system prompt, the conversation, any retrieved documents and the answer it produces. Anything outside it does not exist as far as the model is concerned.
Prompt: A prompt is the text sent to a model to elicit a response; the system prompt is the standing part of it that sets role, rules and output format for every call. Prompt engineering is the practice of shaping that text until the behaviour is reliable.
Fine-tuning: Fine-tuning continues training an existing model on your own examples so it internalises a behaviour — a format, a tone, a classification scheme, a reasoning style. It changes how the model responds, not what facts it knows.
AI agent: An AI agent is a system where a model chooses which actions to take and in what order, calling tools and reacting to their results, rather than following a fixed script written by a developer. The defining property is that the control flow is decided at run time by the model.
Tool use: Tool use is the mechanism by which a model requests that your code run a specific function with specific arguments — a database query, an API call, a calculation — and then continues with the result. The model never executes anything itself; it emits a structured request your system chooses whether to honour.
Model Context Protocol (MCP): The Model Context Protocol is an open standard, published by Anthropic in November 2024, for connecting AI assistants to external tools and data sources through a common interface. It replaces bespoke per-integration glue with servers any compatible client can call.
Multimodal model: A multimodal model accepts or produces more than one kind of data — typically text plus images, audio or video — within the same request. In business use the common case is reading a document, screenshot or photograph and returning text.
Structured output: Structured output is a model feature that forces a response to conform to a supplied schema, usually JSON, so downstream code can parse it without defensive string handling. Modern implementations constrain generation rather than asking politely in the prompt.

Retrieval & knowledge

Retrieval-augmented generation (RAG): Retrieval-augmented generation is a technique that fetches relevant passages from your own documents at query time and supplies them to the model as context, so the answer is grounded in your data instead of the model's training. The model's weights are untouched; only what it is shown changes.
Embedding: An embedding is a list of numbers representing a piece of text, image or audio, positioned so that things with similar meaning sit close together in that numeric space. Comparing embeddings is how a system finds passages that are relevant without sharing keywords.
Vector database: A vector database stores embeddings and finds the nearest ones to a query embedding quickly, at scale. Dedicated products exist, and mature general databases — PostgreSQL with pgvector, for example — do the same job.
Chunking: Chunking is the process of splitting documents into passages small enough to retrieve precisely and large enough to remain self-contained. Strategy ranges from fixed token counts to splitting on document structure such as headings, clauses or table rows.
Hybrid search: Hybrid search combines keyword search with semantic vector search and merges the two result sets. Keyword search catches exact identifiers — part numbers, error codes, names — that embeddings blur together; semantic search catches paraphrases that keywords miss.
Re-ranking: Re-ranking takes the passages a first search returned and reorders them with a slower, more accurate model that scores each one against the query directly. The usual pattern is to retrieve twenty or fifty candidates cheaply, then re-rank and keep the best few.
Grounding: Grounding is the practice of tying a model's output to specific supplied source material, and usually of citing which passage supported which claim. A grounded system can be checked; an ungrounded one can only be trusted or not.
Knowledge cutoff: The knowledge cutoff is the date after which a model's training data ends, so it has no inherent knowledge of later events. Retrieval, tool use and web access are the ways a system supplies what the model does not have.

Quality & evaluation

Hallucination: A hallucination is a fluent, confident model output that is factually wrong or unsupported by the sources provided. It is a property of how these models generate text, not a bug that a future version removes entirely.
Evaluation set: An evaluation set is a fixed collection of representative inputs with known-good outputs, run against the system on every change to measure whether quality moved. It is the regression test suite of an AI system.
LLM-as-judge: LLM-as-judge uses a language model to score another model's outputs against a rubric, making large evaluation sets affordable to run. It is calibrated by checking its scores against human judgements on a sample.
Drift: Drift is the degradation of a system's quality over time as the world, the inputs or the underlying model change while the system stays the same. Provider model updates are a common and often unannounced cause.
Observability: Observability for an AI system means recording every call — the prompt, the retrieved context, the model version, the output, the cost and the latency — so a specific bad answer can be reconstructed afterwards. Standard application logging does not capture enough to do this.
Human in the loop: A human-in-the-loop design routes the model's output to a person for approval, correction or escalation before it takes effect. The model drafts; a human commits.

Cost & performance

Inference cost: Inference cost is what you pay each time the model runs, billed per input and output token at rates that differ between the two. It scales with usage, unlike the fixed cost of building the system.
Prompt caching: Prompt caching lets a provider reuse the processed form of a repeated prefix — a long system prompt, a fixed document set — across calls, charging a reduced rate for the cached portion and returning it faster. The savings depend on structuring prompts so the stable part comes first.
Distillation: Distillation trains a smaller, cheaper model to reproduce the behaviour of a larger one on a specific task, using the large model's outputs as training examples. The result is narrower but much faster and cheaper to run.
Open-weight model: An open-weight model is one whose trained parameters are published, so it can be downloaded and run on infrastructure you control. This is distinct from open-source in the software-licence sense; the weights are available, the training data usually is not.
Frontier model: A frontier model is one of the most capable models available at a given moment, typically from a major lab and typically the most expensive tier on offer. The label moves as the field does.
Latency: Latency is how long a request takes to answer; for language models it is usually reported both as time to first token and as total completion time, and measured at a percentile such as p95 rather than as an average. A p95 of four seconds means one call in twenty is slower than that.

Security & governance

Prompt injection: Prompt injection is an attack in which instructions hidden inside content the model reads — a web page, an email, a PDF, a support ticket — are followed as if they came from the user. It exists because a model sees instructions and data as the same stream of text.
Least-privilege retrieval: Least-privilege retrieval means an AI system can only retrieve documents the requesting user is already entitled to see, enforced by the same permissions as the rest of your stack. The alternative — one index with one service account — collapses every access boundary in the organisation.
No-training clause: A no-training clause is a contractual commitment that data you submit will not be used to train the provider's models. Enterprise and API tiers of the major providers offer this by default; consumer tiers often do not.
Data residency: Data residency is the requirement that data be stored and processed within a named jurisdiction. For AI systems it must cover inference too, not just storage, since a model call sends your data wherever the endpoint lives.
Data processing agreement (DPA): A data processing agreement is the contract governing how a processor handles personal data on a controller's behalf, required under GDPR and equivalent regimes. In an AI engagement it should name model providers as subprocessors, because they are processing on your behalf too.

How the work is bought

Forward deployed engineer (FDE): A forward deployed engineer is a software engineer who works inside a customer's environment — their systems, their data, their constraints — building and shipping software for that specific customer rather than a general product. The term was popularised by Palantir and has since become the standard label for the model applied AI consultancies use.
AI readiness: AI readiness is the extent to which an organisation's data, processes and access controls would let an AI system be built and run — as distinct from whether it has a use case. A readiness assessment measures that, and normally identifies which candidate use cases pay back.
Proof of concept (POC): A proof of concept is a time-boxed build whose only goal is to establish whether an approach works on real data, with no expectation that the code survives. A pilot is the next stage: the same capability run by real users on a limited scope, in production conditions.
Payback period: The payback period is how long a project takes to return its total cost — build plus running costs — in measured savings or additional revenue. For AI work it is normally quoted in months and should include inference spend and ongoing maintenance.

Next step

Knowing the words is not the same as knowing which ones apply to you.

The AI Opportunity Assessment works through your actual processes and data, and tells you plainly which of this vocabulary is relevant and which is somebody else's problem.

Book your assessment→Read straight answers first