Plain words for AI jargon.
56 terms that partners, directors and operations leads meet when their firm starts using AI, each defined in one to three sentences: how the models work, what gets built with them, the law and governance around them, and how firms turn them into value. Every term has its own link, so you can send a colleague straight to it.
The broad field of building computer systems that do tasks which normally need human judgement, such as understanding language, recognising patterns or making predictions. In most business conversations today, AI means generative AI tools such as ChatGPT, Claude, Gemini and Copilot.
Related: AI for consultancies · AI for professional services
The branch of AI in which a system learns patterns from examples rather than following rules a person has written. Almost all modern AI, including large language models, is built with machine learning.
AI that produces new content, such as text, images, audio, code or video, in response to an instruction. It predicts plausible output from patterns in its training data, which is why it can be fluent and wrong at the same time.
Related: AI team training
A machine learning model trained on very large amounts of text to predict the next token, which lets it draft, summarise, translate, classify and answer questions. GPT, Claude, Gemini and Llama are families of large language models.
A large model trained on broad data that can be adapted to many tasks rather than built for one. The EU AI Act calls these general-purpose AI models, and their providers have had specific obligations, such as technical documentation and a copyright policy, since 2 August 2025.
The unit of text a language model reads and writes, often a word or part of a word; in English a token averages roughly three quarters of a word. Model prices, rate limits and context windows are all measured in tokens.
The maximum amount of text, measured in tokens, that a model can take into account at once: your instructions, any documents you supply, the conversation so far and its reply. Anything outside the window is invisible to the model, which is one reason retrieval-augmented generation exists.
The instruction or question you give an AI model, together with any material you attach. Prompts that state the task, the audience, the format and the source material produce markedly better results, and writing them well is a skill that can be taught.
Related: AI team training
Standing instructions, set by whoever builds or configures an AI tool, that apply to every conversation: its role, its tone, its rules and what it must refuse. Users usually do not see it.
When an AI model states something false or invented with apparent confidence, such as a non-existent case citation, a wrong figure or a made-up quotation. It is a side effect of how the models generate text, so the remedy is grounding, checking and human review rather than trust.
Related: The AI safety boundary most businesses trust is only an interface
A language model trained to work through a problem in intermediate steps before it answers, which improves results on analysis, maths and multi-step tasks at the cost of time and tokens. Most leading providers now offer one, often labelled as a thinking mode.
Related: AI reasoning is going dark just as the law demands it
A model that can take in or produce more than one kind of data: reading a scanned contract, a chart or a photograph as well as text, for example, or answering by voice.
A list of numbers that represents the meaning of a piece of text or an image, so that passages with similar meanings sit close together. Embeddings let a system find the right paragraph in a document store by meaning rather than by exact keywords.
Running a trained model to get an answer, as opposed to training it. Inference is what you pay for, usually per token, when you use an AI service.
Further training an existing model on your own examples so that it follows a style, format or specialist task more reliably. It changes how a model behaves rather than reliably teaching it facts; for answers drawn from your own documents, retrieval-augmented generation is usually cheaper and easier to keep current.
A model whose trained parameters (its weights) are published, so anyone can download it and run it on their own hardware or cloud, as with Meta's Llama models. It gives more control over where data goes, but you take on the hosting, security and updates yourself.
Before the model answers, the system searches a set of trusted documents, such as your policies, past proposals or knowledge base, and passes the most relevant passages into the prompt. The answer is grounded in your own material and can cite it, and the documents can change without retraining the model.
Related: 90-day AI development programme · Managed AI engineering service
A database that stores embeddings and finds the entries closest in meaning to a query. It is the usual search layer behind a retrieval-augmented generation system.
Tying an AI model's answer to specific, supplied sources, such as retrieved documents, a database or a web search, rather than to what it absorbed in training. Grounded answers can be checked against their sources, which makes them far easier to trust in client work.
An AI system that is given a goal and works towards it in several steps, choosing which tools to use, such as search, email, a CRM or a spreadsheet, and acting on the results. The more an agent may do without a person approving each step, the more its permissions and checks matter.
Related: Enterprise AI deployment · 30-day AI sprint · AI agents are being trusted faster than they are being checked
A business process in which one or more AI agents carry out some of the steps, usually with fixed checkpoints where a person reviews or approves. It sits between a simple chatbot and full automation.
Related: 30-day AI sprint
The ability of a model to call external functions, such as running a search, querying a database or creating a calendar entry, by producing a structured request that the surrounding software carries out. It is how a model moves from writing about a task to doing it.
An open standard, introduced by Anthropic in November 2024, for connecting AI applications to data sources and tools through one common interface, so a connector built once works across many AI products. Anthropic donated it to the Linux Foundation's Agentic AI Foundation in December 2025, and it is supported by Claude, ChatGPT, Gemini and Microsoft Copilot, among others.
Related: Enterprise AI deployment
In Claude, a skill is a folder of instructions, scripts and reference files that the model loads only when a task calls for it, so a recurring piece of work, such as a proposal format or a house style, is done the same way every time.
Related: Enterprise AI deployment
Using AI inside an automated process so that steps which need judgement, such as reading an email, classifying a document or drafting a reply, can run without someone doing them by hand. Conventional automation follows fixed rules; AI automation copes with messy, unstructured input.
Related: 30-day AI sprint · 90-day AI development programme
An AI tool that works alongside a person, drafting, summarising or answering questions while the person stays in charge of the result. Microsoft 365 Copilot sits inside Office apps; ChatGPT, Claude and Gemini play the same role as standalone apps.
Related: AI team training
The way one piece of software asks another to do something. AI providers sell model access through APIs, which is how bespoke tools and automations are built on top of models such as Claude, GPT and Gemini.
A fixed set of test questions or tasks, with known good answers, that an AI system is run against to measure its accuracy before launch and after every change. Without evals, nobody can say whether a change made the system better or worse.
A design in which a person reviews, approves or corrects an AI system's output before it takes effect, for example signing off a drafted client letter before it is sent. It is the most common control for errors and for professional accountability.
Controls placed around an AI system to keep it within bounds: input and output filters, restricted permissions, blocked topics, required citations and approval steps. Guardrails reduce risk; they do not make a model incapable of error.
Related: The AI safety boundary most businesses trust is only an interface
An attack in which instructions hidden in content an AI system reads, such as a web page, an email or a document, trick it into ignoring its own instructions, for example by leaking data or taking an unwanted action. OWASP ranks it first in its Top 10 risks for applications built on large language models.
Related: AI agents are being trusted faster than they are being checked
The policies, roles, processes and records an organisation uses to decide how AI is used, by whom, with what data and with what checks. In a professional services firm it typically covers an AI policy, a register of tools and uses, impact assessments, training and audit trails.
Related: AI governance implementation · Free AI governance template
A written document that sets out how staff may and may not use AI at work: which tools are approved, what data may go into them, when output must be checked and who to ask. An external AI policy tells clients how the firm uses AI on their work.
Related: AI governance implementation · Our own AI policy · Free AI governance template
Staff using AI tools the organisation has not approved or does not know about, often personal ChatGPT or Claude accounts used with work material. The risk is client or personal data leaving the firm's control; the usual fix is an approved tool and a clear policy, because bans tend to push use out of sight.
Related: Enterprise AI deployment · AI governance implementation
The European Union's law on artificial intelligence, Regulation (EU) 2024/1689, which entered into force on 1 August 2024 and applies in stages. It regulates AI by level of risk and can reach UK firms that place AI systems on the EU market or whose AI output is used in the EU.
Related: AI governance implementation
The EU AI Act sorts AI uses into four levels: unacceptable risk (banned outright since 2 February 2025, such as social scoring), high risk (strict requirements), limited risk (transparency duties, such as telling people they are dealing with an AI) and minimal risk (no new obligations, which covers most everyday business use).
Related: AI governance implementation
Under the EU AI Act, an AI system used in a sensitive area listed in the Act, such as recruitment and managing staff, credit scoring or access to essential services, or one that is a safety component of a regulated product. Since the Digital Omnibus on AI, most high-risk obligations apply from 2 December 2027, and from 2 August 2028 for AI built into regulated products.
An amending EU regulation, in force since 27 July 2026, that simplified parts of the EU AI Act. It moved most high-risk obligations to December 2027 and August 2028 and softened the AI literacy duty, while most transparency duties, such as telling people they are dealing with an AI, kept their August 2026 start.
Under Article 4 of the EU AI Act, providers and deployers of AI systems must take measures to support the AI literacy of staff who use them: the skills and understanding to use AI well and to recognise its risks. The Digital Omnibus on AI softened the duty in 2026 from ensuring a sufficient level to supporting it, and documented training is the usual evidence.
Related: AI team training
A structured assessment that UK GDPR Article 35 requires before processing that is likely to result in a high risk to people's rights, such as large-scale profiling or large-scale use of special category data. The ICO counts innovative technology, including AI, among its triggers and recommends doing one whenever you are in doubt.
Related: AI governance implementation
A wider review than a DPIA of how a particular AI use could affect clients, staff, quality and reputation, covering accuracy, bias, confidentiality, accountability and what happens when it goes wrong. ISO/IEC 42001 builds one into its AI management system.
Related: AI governance implementation
A decision about a person made by a system with no meaningful human involvement, such as an automatic refusal of credit or a job application. UK data protection law restricts such decisions where they have legal or similarly significant effects; the Data (Use and Access) Act 2025 relaxed those rules from 5 February 2026, except for special category data, but people keep the right to contest a decision and to ask for a human to review it.
Where data is physically stored and processed. Enterprise AI platforms can keep data in a chosen region, such as the EU, which matters for client contracts, confidentiality and data protection obligations.
Related: Enterprise AI deployment
The international standard for an AI management system, published in 2023, which sets out how an organisation should govern, assess and monitor its use of AI. Organisations can be certified against it, much as they are against ISO/IEC 27001 for information security.
A voluntary framework from the US National Institute of Standards and Technology, published in January 2023, organised around four functions: Govern, Map, Measure and Manage. It is widely used as a risk checklist well beyond the US.
Systematic unfairness in an AI system's output, usually inherited from patterns in its training data or from how a task is framed, for example rating candidates differently by gender. It matters most where AI informs decisions about people, and it is tested for rather than assumed away.
How far people can understand why an AI system produced a given output. Professional and regulatory accountability usually needs a person who can explain and stand behind a decision, which is why AI output in client work is reviewed rather than simply passed on.
Related: AI reasoning is going dark just as the law demands it
A record of what an AI system was asked, which sources it used, what it produced and who approved it. An audit trail lets a firm show a client, an insurer or a regulator how a piece of work was done.
Related: AI governance implementation
A structured review of how ready an organisation is to use AI well, across areas such as strategy, data, skills, governance and current use, scored against comparable firms. Some people call it an AI readiness assessment. AssumeAI's free version takes nine questions and about two minutes.
Related: Take the AI maturity assessment
A firm's plan for where AI will create value, in what order, at what cost and under what rules, tied to its business goals rather than to particular tools. A useful one names specific use cases, owners and measures of success.
Related: AI strategy review · AI for consultancies
The sequenced delivery plan that follows from an AI strategy: which initiatives come first, what each needs and when each should pay back.
Related: AI priorities workshop · AI strategy review
Ranking candidate uses of AI by value, cost, risk and ease of delivery so that a firm starts with the ones that pay back soonest. In consulting and professional services, research, drafting and synthesis are usually among the early winners.
Related: AI priorities workshop · AI for consulting firms: where it pays back first
A proof of concept is a small, quick build that tests whether an AI approach works at all on real examples. A pilot then puts a working version in front of real users on real work, to show whether it is used and pays back before a wider roll-out.
Related: 30-day AI sprint
The point at which people actually use AI tools in their daily work, not merely have access to them. Most AI rollouts stall on adoption rather than technology, which is why training built around people's own work matters.
Related: AI team training · Why adoption is the most important thing
The business editions of AI assistants, such as ChatGPT Enterprise, Claude Enterprise and Microsoft 365 Copilot, with central administration, single sign-on, data controls and contractual commitments that personal accounts lack. Choosing and configuring one is often the first step in replacing shadow AI.
Related: Enterprise AI deployment
The value an AI initiative produces, typically hours saved, faster turnaround, more work won or fewer errors, set against what it costs to build, run and adopt. It is far easier to prove when a baseline is measured before the change.
Related: AI consultant cost in the UK · Case studies
Laws, standards, dates and product facts were checked against these sources on 21 October 2026. AI law in particular is changing quickly: treat the dates as a guide and take advice on your own obligations.