AI glossary: 40 terms that matter at work.

AI terms from A to Z

AI agent

An AI agent can choose its next step, use a tool and continue working with the result. For a purchasing assistant, that might mean checking stock before drafting a reorder request; permission to place the order should be controlled separately.See AI solutions for business →

Batch

Batch processing handles a collection of jobs together, such as preparing descriptions for an entire product catalogue. It suits work that can wait, but a larger batch does not automatically make each individual job finish sooner.See AI solutions for business →

Classification

Classification assigns an input to one or more defined categories. Routing incoming requests works best when the categories reflect actual responsibilities and the system can flag cases that do not fit cleanly.See AI solutions for business →

Context window

A context window is the model’s working capacity for the tokens involved in a request. Adding an entire document archive can still produce poor answers if the useful passages are buried among irrelevant material.

Data processing agreement, DPA

A DPA sets out the obligations between a controller and a processor under Article 28 GDPR. It should cover the actual service, including security, subprocessors and what happens to personal data when the engagement ends.

Embedding

An embedding represents content as numbers that can be compared mathematically. It can help match a customer’s question to a relevant manual section, but a close match still needs to be checked for the right product and version.See AI solutions for business →

EU AI Act

The EU AI Act regulates AI through rules that depend on the use case and the organisation’s role. Before deployment, establish whether you are acting as a provider or deployer and which requirements apply. A general claim that a model is compliant does not settle the obligations of your application.See AI solutions for business →

Fine-tuning

Fine-tuning adapts an existing model through further training on selected examples. It may help produce consistent classifications or specialist responses, but success should be measured on cases that were not part of the training data.See AI solutions for business →

Function calling

Function calling lets a model request a defined operation with structured arguments. A support assistant might request shipment details, while the application decides whether the request is valid and which records the user may access.Connect AI to your existing systems →

GDPR compliance

GDPR compliance concerns the complete use of personal data, including collection, processing, access and deletion. Cloud and on-premises systems both need a lawful basis and appropriate safeguards; the hosting location cannot answer every compliance question.See AI solutions for business →

GPU/VRAM

A GPU accelerates many AI calculations, and VRAM is the memory available directly on that device. Hardware sizing must allow room for the workload as well as the model; fitting the weights into memory is only the starting point.Plan the right infrastructure →

Guardrails

Guardrails constrain how an AI application receives information, responds and takes action. A refund assistant might prepare a recommendation while a separate permission check prevents it from issuing payments above an approved limit.See AI solutions for business →

Hallucination

A hallucination is an AI output that presents invented or unsupported information as if it were established. A confident answer about a warranty can still be wrong, so important claims should be checked against the applicable document.See AI solutions for business →

Inference

Inference is the stage where a trained model processes a request and produces a result. It is the everyday workload your infrastructure must support, whether that means answering questions, recognising speech or assigning document categories.

Information extraction

Extraction turns relevant parts of a document into fields that software can use. When reading a purchase order, the system should distinguish an actual delivery date from a requested one and leave absent values empty.See AI solutions for business →

Knowledge distillation

Distillation uses a teacher model to help train another model, often a smaller one. A compact classifier may then handle a narrow business task efficiently, but its quality still needs checking against real examples.

Large language model, LLM

An LLM learns patterns from large amounts of language data and uses them to process and generate text. Its usefulness in a business application depends on the instructions, information and tools supplied around it, as well as the model itself.See AI solutions for business →

Latency

Latency describes how long a user waits for a system to respond. In a voice application, speech detection, network calls and audio generation all contribute to the pause the caller experiences.

Model Context Protocol, MCP

MCP provides a standard way for AI applications to connect to tools and information sources. A connector can expose a document search or business operation, while authentication and access permissions still require deliberate configuration.Connect AI to your existing systems →

Multimodal

A multimodal model works with more than one type of information, such as text and images. That can support questions about a photographed component, but support for pictures does not automatically include audio or video.

On-premises

An on-premises system runs on infrastructure at the organisation’s own premises. It can suit local integration or offline requirements, provided maintenance, security and sufficient capacity are covered.Plan the right infrastructure →

Open-weight model

An open-weight model makes its learned parameters available for others to run. Before adopting it, check the licence for your intended use and establish which supporting components and training information are actually available.See AI solutions for business →

Optical character recognition, OCR

OCR converts visible text in an image into text that software can search or process. Reading the characters on a delivery note is one step; identifying the correct item quantities is another.See AI solutions for business →

Private cloud

A private cloud provides cloud infrastructure for the exclusive use of one organisation. It can support dedicated workloads without requiring an in-house data centre, but the service agreement should clearly define isolation, administration and location.Plan the right infrastructure →

Processing on behalf of a controller

A service provider acts as a processor when it handles personal data under your organisation’s instructions. Check what it actually does with the data: processing for its own purposes can place that activity under a different legal role.

Prompt

A prompt tells the model what to do and supplies relevant context or constraints. For a sales summary, specifying the audience, required facts and output format gives reviewers something concrete to assess.

Quantisation

Quantisation uses lower-precision numerical representations to reduce a model’s resource requirements. A smaller memory footprint can be valuable, but check difficult production examples before assuming the compressed version behaves equally well.See AI solutions for business →

Reranking

Reranking takes an initial set of search results and applies a further relevance assessment. It can improve which passages reach an assistant, but it cannot recover a document that the first search never retrieved.See AI solutions for business →

Retrieval-augmented generation, RAG

RAG combines retrieval from an external knowledge source with a model-generated answer. A service assistant can consult the relevant maintenance instructions for each question, allowing documents to change independently of the model. Retrieved passages still need to support the answer actually given.See AI solutions for business →

Sampling

Sampling draws the next token from a probability distribution instead of always taking the highest-ranked option. Settings such as top-p change the available choices, so they should be tested against the consistency your task requires.

Semantic search

Semantic search retrieves content by meaning rather than requiring the same wording. It can connect a customer’s description of a fault to the right troubleshooting article, while exact identifiers often benefit from keyword matching alongside it.See AI solutions for business →

Speech-to-text

Speech-to-text produces a transcript from audio. Test it with the accents, product names and call quality your business actually encounters, especially where a single misheard digit changes the result.See AI solutions for business →

System prompt

A system prompt sets application-level instructions, such as the assistant’s scope and response rules. It guides behaviour, while enforceable controls must separately determine which data and actions are available.

Temperature

Temperature adjusts how concentrated the token probability distribution is during sampling. Lower settings usually make wording more predictable, but factual accuracy still depends on the model, evidence and task.

Text-to-speech

Text-to-speech turns written content into generated audio. A business voice should handle product codes and abbreviations clearly, while the surrounding application manages timing and interruptions.See AI solutions for business →

Throughput

Throughput measures the amount of work a system completes over time. For an invoice workflow, documents processed per hour may be more useful than a model’s token rate, particularly when validation and database updates are included.Plan the right infrastructure →

Token

A token is a unit a language model uses to represent text. Token counts depend on the tokenizer and content, so a page count or word count is only a rough guide to how much input a model can handle.

Vector database

A vector database stores numerical representations and supports similarity search over them. For company knowledge, it also needs a reliable connection to source records so updates, deletions and permissions remain consistent.See AI solutions for business →

Vision model

A vision model analyses visual input, such as a photograph of damaged packaging. Its suitability depends on image quality and the exact inspection task; performance on general pictures does not establish reliability for industrial quality control.

Voice agent

A voice agent combines spoken interaction with software that can answer questions or carry out defined tasks. A useful deployment needs clear escalation paths and should be tested with real call conditions, including silence, interruptions and misunderstood requests.See AI solutions for business →