AI Glossary

AI Glossary

The AI terms that come up in meetings, proposals and vendor presentations — in plain words, with what each one means for a business. Reviewed every Monday.

95 terms · last updated · 21 added this week

This glossary is here so you can quickly look up a term that comes up in a meeting or a vendor presentation. Alongside the short definition, I have tried to note what each term touches on the business side.

Type in the box to search, and use the group buttons to filter. Terms marked New were added within the last week.

Agent

A structure that plans its own steps towards a given goal, calls tools and judges the result. Unlike one question and one answer, it carries a task from start to finish.

Applications
Agent identity non-human identity New

An account and set of permissions given to an AI agent in its own right, separate from human users. What the agent does is tracked through this identity, and its rights are kept to what the job needs and limited in time.

In the field: If an agent connects to the ERP under an employee’s user name, it becomes unclear who actually made a change. A separate user with a narrow role keeps the audit trail clean.

Trust and governance
Agent observability tracing New

Recording every step an agent takes while it runs — which model it called, which tool it used, what data it looked at, how long it took and how many tokens it spent. It is the common ground for debugging, cost tracking and audit.

In the field: “Why did the agent cancel this order?” has no answer without a step-by-step trace.

Cost and operations
Agent2Agent protocol A2A New

An open protocol that lets AI agents from different vendors, or built with different tools, find each other, hand over tasks and exchange messages. MCP connects an agent to tools and data; A2A connects agents to each other.

In the field: For your purchasing agent to talk to a supplier’s order agent, or to the sales agent in your CRM, in a shared language, a standard like this is needed.

Applications
Agentic AI New

An umbrella term for systems that can plan their own steps and call tools. It is often used as a marketing heading.

In the field: If it appears in a proposal, ask: which tools does it call, which decisions does it make alone, and which come back for approval?

Applications
AI bubble New

A term for the debate over whether investment and valuations in the field have run ahead of the revenue actually being produced.

In the field: However that debate ends, tying your own investment to one measurable job is the safest route.

Cost and operations
AI literacy New

Staff who use an AI tool, or make decisions based on its output, knowing what the tool can and cannot do and what risks it carries. The EU AI Act expects organisations using such systems to take measures that support this literacy.

In the field: A short usage guide and a session on “which data never goes into which tool” reduce the risks of shadow AI and data leakage together.

Trust and governance
AI slop New

An everyday term for the mass of quickly generated content that is weak on accuracy and originality.

In the field: Asking who the source is, for information from a search result or a supplier document, is now a routine check.

Trust and governance
AI-powered New

A label that appears in most product pitches and says nothing on its own.

In the field: What to ask when assessing a proposal: which job does it do, which model, where is the data processed, what is the error rate, and what happens when it is wrong?

Cost and operations
Alignment

The work of bringing a model’s behaviour in line with human intentions and values.

Trust and governance
Anomaly detection

Flagging records that fall outside the usual pattern: an unusual price, unexpected consumption, a repeated faulty entry.

Applications
Artificial general intelligence AGI

A hypothetical AI able to do almost any mental task a person can. It does not exist today, and there is no agreement on its definition or its timing.

Foundations
Artificial intelligence AI

The general name for making software do work that requires human intelligence — classifying, predicting, understanding language, proposing decisions. Not one technology, but a heading.

Foundations
Benchmark

A model’s score on standard tests. It gives a general idea but says nothing about performance on your data; it does not replace your own evaluation.

Cost and operations
Chain of thought

The model producing the intermediate steps that lead to its result. It improves accuracy on multi-step calculation and logic, at the cost of time and money.

Language models
Chunking

Splitting long documents into small searchable pieces. The size of a chunk and where it is cut directly affect RAG quality.

Data and retrieval
Computer vision

The branch of AI that extracts information from images and video: recognising parts, spotting surface defects, reading labels.

In the field: The first application area that comes to mind in quality control and stock counting.

Foundations
Computer-use agent New

An agent that works in application screens the way a person does, looking at screenshots and using the mouse and keyboard. Useful for systems with no integration point; slower and more error-prone than working through an API.

In the field: An option for entering data into an old program that has no API. Running it in a separate environment with limited rights and putting critical steps behind approval are the safeguards vendors themselves recommend.

Applications
Context window

The total amount of text a model can take into account at once. What happens when the limit is reached depends on the application: the request may be refused, or the oldest part dropped or summarised.

In the field: The first number to check if you want a long contract or technical specification read in one go.

Language models
Copilot

An assistant that does not take the work over, but sits alongside the user and offers suggestions.

Applications
Data leakage

Company information leaving unintentionally — a customer list or a price table pasted into a tool.

Trust and governance
Data poisoning

An attack that corrupts a model’s output by deliberately inserting bad records into its training or retrieval data.

Trust and governance
Data residency

Which country data is physically held in, and whose law therefore applies to it.

In the field: One of the first clauses to check in a cloud AI contract.

Trust and governance
Deep learning

Machine learning done with multi-layer neural networks. On raw data such as images, sound and language it is clearly better than the alternatives.

Foundations
Deepfake New

A synthetic image or audio recording that imitates a person’s face, voice or movement.

In the field: Requiring a second channel to confirm a payment instruction that arrives in a manager’s voice is a concrete safeguard.

Trust and governance
Demand forecasting

Predicting future demand from past sales, seasonality and campaign effects. It works well on regular demand; on intermittent, lumpy demand it is not guaranteed to beat simple methods.

Applications
Digital twin New

A digital copy of a machine, a line or a plant, fed with real data, on which trials are run. AI is not a requirement, but the two are often mentioned together.

Applications
Distillation

Teaching a smaller model to behave like a larger one. The aim is to keep as much quality as possible while cutting cost and latency.

Language models
Embedding

Turning text into a sequence of numbers that represents its meaning. Meanings that are close to each other end up close in that number space.

Data and retrieval
EU AI Act

The European Union’s regulation classifying AI systems by risk level. It also concerns companies supplying products or services into Europe.

In the field: Obligations apply in stages: the prohibited practices and transparency rules are in force, while the rules for high-risk systems were pushed back by a 2026 amendment. The first step is to work out which risk class your own use falls into.

Trust and governance
Evaluation eval

Measuring the quality of a model or an application against a prepared set of questions and answers. Improvement without an evaluation set is guesswork.

Cost and operations
Explainability

Being able to understand why a model arrived at a particular result. For decisions subject to audit it is a requirement, not a preference.

Trust and governance
Few-shot prompting

Putting a few worked examples into the prompt to show the model the expected form. The cheapest way to correct behaviour without training.

Language models
Fine-tuning

Adapting an existing model to a particular job with a smaller, specific dataset.

In the field: For most enterprise needs, try RAG before fine-tuning: cheaper, faster and easier to keep current.

Language models
Foundation model

A large model trained once on very broad data and then adapted to many different jobs. Large language models are the best known members of this family.

Foundations
Frontier model New

A term for the largest and newest models, the ones representing the leading capabilities of the field. The line is not fixed; today’s frontier model is considered ordinary a few years later.

In the field: Regulation and safety debates mostly target this class; for everyday enterprise work, smaller and cheaper models are usually enough.

Language models
Generative AI

The kind of AI that produces new content — text, images, sound or code. Unlike traditional models that classify and predict, it creates the output itself.

Foundations
Guardrails

The rules that limit what a model may do: which data it can reach, which action it cannot take without approval, which subjects it does not answer.

Trust and governance
Hallucination

A model producing information that looks right but is wrong or invented. Every answer that is not tied to a source carries this risk.

Language models
Human in the loop

A design where the model proposes and a person decides. It is how trust is built in enterprise work; automation is increased afterwards.

Applications
Inference

Running a trained model to produce an answer. Training happens once; inference repeats with every request — that is where the ongoing cost sits.

Language models
Inference cost

What each question and answer costs to run. It looks small during a pilot and becomes the main budget line as usage spreads.

Cost and operations
Intelligent document processing IDP

Automatically extracting fields from documents such as invoices, delivery notes and order forms, and passing them into a system.

Data and retrieval
Jailbreak

Trying to get around the limits placed on a model through specially constructed prompts.

Trust and governance
Knowledge cutoff

The date where a model’s training data ends. It knows nothing after that date unless you give it the information.

In the field: If current regulation or pricing is being asked about, the answer must be tied to a source.

Language models
Large language model LLM

A model trained on very large amounts of text that produces text by predicting the next word. It is the engine behind ChatGPT and similar tools.

Language models
Latency

How long an answer takes to arrive. On a screen used on the shop floor, the difference between two seconds and ten decides whether it gets used at all.

Cost and operations
Machine learning ML

Methods that derive rules from example data instead of having the rules written by hand. Today it is the most widely used branch of AI.

In the field: The first place to look for work with plenty of history behind it: demand forecasting, scrap prediction, quality deviation.

Foundations
Mixture of experts MoE

A design where the model runs only part of its parameters for each question. It combines large capacity with lower inference cost.

Language models
Model card

A document describing what a model was trained for, its limits and its known weaknesses. The first document to ask for in a purchasing assessment.

Trust and governance
Model collapse New

The risk that quality degrades over time as models are trained on content produced by earlier models.

Language models
Model Context Protocol MCP

An open protocol for connecting AI applications to outside systems and data. It reduces the need to write every integration from scratch.

Applications
Model routing

Sending simple questions to a small, cheap model and hard ones to a powerful model. One of the most practical ways to bring cost down.

Cost and operations
Multi-agent system

Several agents sharing one job, each taking its own task. It breaks complex work into parts, but is harder to debug than a single agent.

Applications
Multimodal model

A model that can process images, sound or tables alongside text.

In the field: Opens the door to inputs such as technical drawings, label photos and scanned delivery notes.

Language models
Neural network

A mathematical structure that passes input through weighted connections to produce output. It is also the basis of today’s language models.

Foundations
On-device AI edge AI New

A model running on the device itself — a phone, laptop or handheld terminal — without going to the cloud or a central server. Data stays on the device and it works without a connection; in return, the model can only be as large as the device allows.

In the field: Worth considering for label reading, voice commands or photo classification in warehouses and on site, where the connection is weak.

Applications
On-premises model

A model running on your own servers. The data never leaves the building; the hardware, maintenance and updates become your responsibility.

Cost and operations
Open-weight model

A model whose weights can be downloaded and run on your own servers. Not the same as “open source”; the training data is usually not shared.

Language models
Overfitting

A model memorising the training data and then failing on new data. The most common reason a model that looked perfect in testing stumbles in the field.

Foundations
Parameter

The numerical weights a model adjusts during training. The count relates to capacity, but on its own it does not indicate quality.

Language models
Post-training New

The general name for the stages that turn a pre-trained model into one that follows instructions and works well and safely on a given job. Fine-tuning, reinforcement learning from human feedback and similar methods fall under this heading.

In the field: A company doing high-volume work in its own field can post-train an open-weight model on its own data to lower costs; in return it takes on training data, evaluation and maintenance.

Language models
Pre-training

The first and expensive training stage, on very large data, where a model gains its general language ability.

Language models
Predictive maintenance

Predicting machine failure from patterns in sensor data and maintenance history. It increases planned downtime and reduces unplanned downtime.

Applications
Process mining

Reconstructing how a process actually flows from the event logs in a system. It puts a number on the gap between the process on the map and the process on the floor.

In the field: The most honest way to see where approval and order processes wait in an ERP.

Applications
Prompt

The whole of the instruction and context given to a model. The quality of the result often depends more on the clarity of the prompt than on the model.

Language models
Prompt caching

Caching the long context repeated on every request so the same text is not processed again. It cuts cost and latency noticeably.

Cost and operations
Prompt engineering

The work of shaping a prompt so that it produces the result you want: task description, examples, formatting rules and limits.

Language models
Prompt injection

Diverting a model from its real task with instructions hidden in a document, email or web page it reads. A model cannot always tell data from instructions, so the problem has not been solved at the root.

In the field: An agent that reads a supplier’s PDF and creates a record in the ERP can be pushed into the wrong action by one hidden sentence in that document. Keeping the agent’s permissions narrow and putting critical actions behind approval are the basic safeguards.

Trust and governance
Quantization

Shrinking a model by storing its weights at lower precision. It is what makes running a model on ordinary in-house hardware possible.

Language models
Rate limit

The cap on requests or tokens a service accepts within a given period. At busy hours the bottleneck usually appears here.

Cost and operations
Reasoning model

A type of model trained to think for longer before answering. Good for hard analysis; unnecessary expense for simple work.

Language models
Red teaming New

Testing an AI system by deliberately acting like a real attacker — trying to trick it, push it into the wrong action or make it leak data. The aim is to find the gap before an attacker does; it can be done by people, by models, or both.

In the field: Before putting an assistant connected to the ERP live, trying to make it show records outside its permissions or act without approval is the most basic form of this test.

Trust and governance
Reinforcement learning

A model that learns by trying and correcting its behaviour according to the feedback it gets. Used in scheduling and game-like decision problems.

Foundations
Reinforcement learning from human feedback RLHF

Scoring a model’s answers against human preferences and training on those scores. It is one of the post-training methods that make a model useful and appropriate to talk to.

Language models
Retrieval-augmented generation RAG

Finding the documents relevant to a question first, then asking the model with those documents attached. The model answers from your data, not from its own memory.

In the field: Most enterprise question-and-answer applications are built this way; it also makes citing the source possible.

Data and retrieval
Reward hacking specification gaming New

A model or agent scoring high by exploiting a gap in the measure of success, without actually doing the intended work. The fault is usually less in the model than in a measure that does not fully describe the job.

In the field: An agent that reports its own tests as passing may have “passed” them by changing the tests. Tie the result to an independent check, not to the agent’s own report.

Trust and governance
Shadow AI

Employees using AI tools through personal accounts, without the company knowing. Opening an official route works better than banning it.

Trust and governance
Singularity New

The hypothesised moment when AI improves itself so quickly that it moves beyond human control. More a subject of debate than a scientific forecast.

Foundations
Small language model SLM

A language model kept small enough to run on limited hardware. On a narrow task it can come close to a large model at a fraction of the cost.

Language models
Superintelligence New

AI assumed to surpass humans in every field. It does not exist today; the word appears often in investment and policy debates.

Foundations
Supervised learning

Training a model on data labelled with the right answer. Used wherever you can say “this record is faulty, this one is not”.

Foundations
Synthetic data

Artificial data generated in place of real data, carrying the same statistical properties. Used where privacy matters or real examples are scarce.

Foundations
System prompt

The standing instruction, invisible to the user, that sets the model’s role and limits.

Language models
Temperature

The setting that decides how variable an answer will be. A low value gives consistent, repeatable output; a high one gives creative but unsteady output.

In the field: Kept low for work such as report and SQL generation, where repeatability is essential.

Language models
Text-to-SQL

Turning a question asked in everyday language into a database query. The generated query must be readable and reviewable.

Data and retrieval
Token

The smallest piece a model works with when processing text — a whole word or a fragment of one. Pricing is usually per token.

Language models
Tool use function calling

The model being able to call functions defined for it: running a report, querying a record, drafting an e-mail.

Applications
Training data

The dataset a model learns from. A model’s quality is largely the quality of this data.

Foundations
Transformer

The network architecture behind today’s language models. It rests on an “attention” mechanism that weighs how each piece of text relates to the others.

Foundations
Turing test New

The classic experiment measuring whether a machine can be told apart from a person in written conversation. It is not used as a yardstick for today’s models; it is a historical starting point.

Foundations
Unsupervised learning

Methods that find patterns and clusters in data without any labels.

In the field: Grouping stock records opened for the same part under different names is a typical example.

Foundations
Vector database

A database that stores embeddings and answers “which records are most similar to this one” quickly.

Data and retrieval
Vibe coding New

Building software by describing what you want in plain language and trying what the model produces, rather than writing it line by line. The term spread in 2025; it suits quick prototypes.

In the field: A practical route for a small internal tool; code that touches accounting or stock movements still needs review and tests.

Applications

If a term is missing, or you think one is explained wrongly, write to us; the glossary is reviewed every Monday.