The AI glossary

Software that does work we used to assume needed a person, such as writing, summarising, reading images or judging what a request means.

A model trained on very large amounts of text to predict what comes next, which is what lets it write, summarise, translate and answer questions.

The trained file at the centre of an AI product, holding what it learnt from its training data, which the chat window or app is wrapped around.

A large general-purpose model trained once at great expense, then adapted by many companies for many jobs rather than built fresh for each one.

A model trained to work a problem through in steps before answering, which helps on maths, logic and planning at the cost of speed.

An AI system that works towards a goal in steps rather than answering a single question: it can plan, use tools such as search or a calendar, check its own results and try again.

A hypothetical AI able to match a capable person across most intellectual work rather than in narrow areas, with no agreement on whether or when it arrives.

The model design behind almost every modern AI system, able to weigh every part of an input against every other part at once.

A stack of simple mathematical units that pass signals to each other, loosely inspired by brain cells, and the shape almost every model takes.

The learnt numbers inside a model that hold what it knows, usually counted in billions and often quoted as a rough measure of size.

A model split into specialist sections where only a few run for any given request, giving large-model quality at smaller running cost.

A deliberately compact language model that trades some breadth for speed, lower cost and the ability to run on modest hardware.

The largest and most capable models available at any moment, usually the most expensive and the ones vendors lead their marketing with.

A set of models from one maker sharing a name and design, offered in sizes so you can trade cost against capability.

A model built to handle many kinds of task rather than one, which is also the category European regulation now names directly.

AI that produces new text, images, audio or code rather than only sorting or scoring things that already exist.

The broader field of software that learns patterns from examples instead of being given step-by-step rules by a programmer.

Machine learning built from many layers of connected calculations, which is what made today's language and image models possible.

The older name for getting software to work with human language, still used in job adverts and in established vendor products.

Able to work with more than one kind of input or output: text, images, audio, sometimes video.

The field concerned with getting software to interpret what is in an image or video: objects, faces, defects, movement.

Producing a new picture from a written description rather than retrieving or editing an existing one.

The kind of model behind most image generators, which starts from visual noise and refines it step by step towards the picture your description asks for.

AI you speak to and that speaks back, joining speech recognition, a language model and synthetic speech into one conversation.

Turning recorded or live speech into written words, which is what produces a meeting transcript or live captions.

Turning written words into spoken audio, used for narration, accessibility and assistants that answer out loud.

Reading the text inside a picture of a page, so a scan or a photograph becomes words a computer can search and copy.

A model that reads images and text together, so it can answer questions about a screenshot, chart, form or photograph.

Pulling structure and meaning out of documents such as invoices or contracts, rather than only converting them to plain text.

Producing moving footage from a written description or a still image, typically in short clips that are then edited together.

Changing an existing image by instruction, such as removing a background or replacing an object, rather than generating one from nothing.

Turning a written description straight into a video clip, the video counterpart of asking for an image in words.

Building a synthetic copy of a particular person's voice from samples, which raises a consent question before any practical one.

A presenter-style video built from a synthetic person reading a script, used for training and internal communications without filming anyone.

Producing music, sound effects or other audio from a written description, distinct from reading text aloud in a chosen voice.

The text, images and code a model learnt from: much of the public web, books, licensed sources, and sometimes users' conversations.

The long first stage of building a model, where it learns general patterns from an enormous body of material before anyone teaches it to be helpful.

Taking a trained model and training it further on a smaller, specific set of examples: your firm's writing style, a legal domain, a support script.

Reusing what a model learnt on one large task as the starting point for a different, usually smaller one, instead of starting from nothing.

Training by trial and reward rather than by copying examples: the system tries something, gets scored, and shifts towards what scored well.

The stage where people rank a model's competing answers and it is trained towards the ones they preferred, which is much of why assistants feel helpful.

Training a smaller model to imitate a larger one, to get most of the quality at a fraction of the cost and wait.

Storing a model's numbers at lower precision so it needs less memory and can run on smaller hardware, giving up a little accuracy.

Training material generated rather than collected, used where real examples are scarce, sensitive or expensive to label.

The kind of chip AI runs on, built to do enormous numbers of simple calculations at once, which is exactly the shape of the work models do.

Marking up examples so a model can learn from them, such as tagging which emails are complaints, usually slow and done by people.

A model adapted to one organisation's data or tone, which vendors sell as a premium tier and which needs upkeep as the underlying model moves.

Naming and dating the exact model behind a feature, so results can be reproduced and a change can be traced to a cause.

A vendor retiring a model version on a published date, after which anything built on it must move or stop working.

A vendor's published account of how a whole AI system behaves, including its known weaknesses and the testing it was put through.

A short standard summary of one model's intended uses, training approach and limitations, published so buyers can compare like with like.

Everything done to shape a model after its main training run, which is where most of its manners and refusals come from.

The date a model's training data stops, after which it knows nothing unless the information is searched for or supplied in the request.

What you type to an AI assistant: the request plus any context you include.

Standing instructions given to a model before your conversation starts, usually by the tool's developer: its role, tone, rules and boundaries.

The practice of writing and refining requests to get reliable results, which in practice looks more like briefing well than like engineering.

Showing a model two or three worked examples inside your request so it can match the pattern you want.

Asking for something with no worked examples at all, relying on what the model already learnt.

The amount of text a model can consider at once: your conversation, any documents you attach, and its own replies.

The unit models actually read and write: a short chunk of text, roughly three-quarters of an English word on average.

A setting that controls how predictable a model's output is.

A feature that lets an assistant carry facts about you across separate conversations, rather than starting blank every time.

When a model states something false as if it were certain: an invented citation, a wrong date, a plausible-sounding statistic.

Tying a model's answer to a specific, checkable source: a search result, your uploaded document, a database record.

The link or reference an AI answer offers as the source of a claim, which is only useful once you open it and check.

The intermediate reasoning a model works through before giving its answer, sometimes shown to you as thinking.

The hidden working a reasoning model produces on its way to an answer, which you usually pay for and count against limits without ever seeing it.

An AI product with its own chat surface that you go to and ask, handling many kinds of task rather than one.

AI built into a tool you already use, offering help in place rather than asking you to go somewhere else and describe the work.

A mode that spends minutes rather than seconds gathering and comparing sources, returning a written brief with references instead of a quick answer.

A loose label for building software by describing what you want and accepting the result, with little review of the code produced.

A vendor's bundle of models, tools and controls sold as one thing to build on, as opposed to a single finished application.

Letting an AI drive a screen the way a person would, clicking and typing in ordinary software that offers no other way in.

A side panel where a draft sits and is edited in place, so a document is revised rather than rewritten in the chat.

Attaching a document or image to a conversation so the AI can work from it, and the first point at which a data question arises.

What a product keeps from your previous sessions, which decides both how much it remembers and how much of your work it stores.

A saved assistant set up once with its own instructions and files, then reused by a team, sold under a different name by each vendor.

Before answering, the system searches a chosen set of documents, then writes its reply from what it found.

A way of representing text (or images) as a long list of numbers so that similar meanings end up numerically close together.

A database built to store embeddings and find the closest matches quickly, even across millions of documents.

Search that matches on meaning rather than on the exact words typed, so a query about annual leave finds the holiday policy.

The defined set of documents an assistant is allowed to answer from: policies, manuals, past tickets, whatever you have chosen to load.

Cutting long documents into passages small enough to search and feed to a model, which quietly decides how good the answers will be.

Getting your documents into a system so they can be searched, including the connectors, permissions and refresh schedule that keeps them current.

Reading a file's real structure, its headings, tables and columns, so that meaning survives instead of collapsing into a wall of text.

Combining keyword matching with meaning-based search, so exact terms such as part numbers still land while related wording is also found.

Matching the words as typed, which is precise for names and codes and unhelpful when the reader phrases things differently.

A map of how things relate to each other, such as which customer owns which contract, letting a system answer joined-up questions.

Search across every system an organisation uses at once, returning cited answers and showing each person only what their access allows.

The prepared copy of your content that search actually reads, which is why newly added documents take a while to become findable.

The facts about a document rather than its contents, such as owner, date and department, used to filter results and enforce access.

A product that replies with a written answer and its sources instead of a list of links, now a category of its own.

A citation pointing at the exact passage an answer came from, rather than the whole document, so a claim can be checked quickly.

Search that respects who is allowed to see what, so the same question returns different results for different people.

The full path from your documents to a grounded answer, covering ingestion, chunking, search and what is finally handed to the model.

A sequence of steps that runs in a set order, with each step's output feeding the next, whether people or software do the work.

Handing a repeated task to software so it runs on a trigger instead of waiting for someone to remember it.

Coordinating many automated steps across different systems, including the order they run in, what happens when one fails, and what gets retried.

An open standard for building connectors, so one connector can work across different assistants instead of being rebuilt for each.

Connectors let an AI assistant reach into your other tools: read a shared drive, check a calendar, update a spreadsheet.

An add-on that extends a tool you already use, rather than a separate product you open on its own.

An application programming interface: the route by which one piece of software talks to another.

A ready-made bundle of code a provider publishes so developers can use its API without writing the plumbing themselves.

A message one system sends to another the moment something happens, so the second system can react without repeatedly asking whether anything has changed.

The mechanism by which a model asks to run something outside itself, such as a search or a calculation, and then uses the result in its answer.

Making a model answer in a fixed shape with named fields, so another system can read the reply instead of a person having to.

A setting that forces a model's reply into JSON, the plain-text format software uses to pass data around.

Designing an automated process so a person approves or corrects specific steps rather than only seeing the result.

Tools built so the work is assembled visually rather than programmed, letting people who do not write code build and change automations.

Mostly visual building with the option to drop into code for the awkward parts, the usual compromise once automations get real.

A product whose job is connecting other products, holding the credentials and the retry logic so each pair does not need bespoke work.

The event that starts an automation, such as a form submission or a new file, and the first thing to check when nothing ran.

A secret string that identifies your account to a service, the usual first step in connecting tools and the usual first thing leaked.

The other products an AI tool has been given access to, which sets both what it can do and where your data reaches.

Older automation that copies a person's clicks and keystrokes across systems, widely deployed before AI and often what AI is now replacing.

A standard test used to score models against each other, useful for shortlisting and unreliable as a promise about your own work.

The wait between sending a request and getting the answer.

Running a trained model to produce an answer, as opposed to training it in the first place.

What it costs to run a model each time it answers, which is why heavy use is metered and the strongest models are rationed.

A cap on how many requests you may send in a given period, which protects the service and shapes how fast an automated job can run.

The ceiling a plan puts on how much you can do in a day, week or month before you are slowed, switched to a weaker model, or stopped.

Paying for what you actually use rather than a flat subscription, which suits uneven demand and makes forecasting harder.

A free tier is a usable product at no cost, with limits such as message caps or slower models; freemium describes the business model built around it, where the free version exists to sell the paid one.

Reusing the work already done on a repeated piece of context, which cuts both cost and wait when the same material is sent often.

Charging by the amount of text going in and coming out rather than per person, which makes long documents cost more than short questions.

A prepaid unit a vendor charges against, worth checking carefully because one credit rarely means the same thing across two products.

Paying a fixed amount per person per month, predictable to budget for and wasteful when only some of those people use it.

Paying for what you actually consume, which rewards light use and can surprise you in a busy month without a cap in place.

The amount of use a plan includes before extra charges begin, the number worth finding before comparing two headline prices.

A contractual promise about availability and response times, with agreed consequences, and the thing to ask for before depending on a tool.

Sending each request to whichever model suits it, keeping the expensive one for the hard cases and something cheaper for the rest.

The address your software sends requests to for a particular model, which is what you are really buying when you buy access.

Running a model on infrastructure someone provides, whether the model maker's, a cloud vendor's or your own.

A single point every AI request passes through, so keys, spending limits, logging and model choice are managed in one place.

How much work a system gets through in a given time, the number that matters for bulk jobs where a single reply's speed does not.

What one completed piece of work costs rather than what one person or one month costs, the unit that makes AI spending comparable to labour.

The reason a long document costs more to ask about than a short one, because everything you send is charged for on the way in.

How many requests a plan lets you run at the same time, usually the first limit a team meets when it moves past piloting.

What you are charged once an included allowance runs out, and the commonest unpleasant surprise on an AI invoice.

Running an AI model on computers you control, whether your own servers or a private cloud, instead of sending data to a provider's service.

A model running on the machine in front of you, so what you type never leaves it.

AI that runs on or beside the device producing the data, such as a phone, camera or vehicle, rather than sending it away to be processed.

AI that runs on a provider's servers and reaches you over the internet, which is how nearly every well-known assistant works.

Software whose code is published under a licence that lets anyone read, change and redistribute it.

A model whose trained parameters are published for anyone to download and run, unlike closed models used only through a provider's service.

A model you can only use through its owner's service, with the trained file itself never released.

What happens to the text, files and recordings you put into an AI tool: who can see them, how long they are kept, and whether they train the model.

A setting that tells a provider not to use your conversations to train its models.

A commitment about which country your data is stored in, usually the first question asked in regulated work and in the public sector.

Whose laws your data falls under, which is not settled by storage location alone because the provider's own jurisdiction can still reach it.

Running software on hardware you control, in your own building or data centre, rather than reaching it as a service over the internet.

A deployment with no connection to the internet at all, used where data must not leave the room and updates arrive by hand.

How long a vendor keeps your inputs and outputs before deleting them, and whether you can change that period.

A vendor commitment to discard your inputs and outputs immediately after answering, usually available on enterprise terms and worth getting in writing.

Stripping or replacing personal details before text reaches a model, so a name or account number never leaves your control.

Keeping each customer's data separated inside a shared service, the assurance behind a vendor saying your content stays yours.

A private route to a hosted model that avoids the public internet, sitting between ordinary cloud access and running the thing yourself.

A vendor offering to run the service in a chosen part of the world, which is how a data residency promise is usually delivered.

The work of making AI systems behave predictably and within bounds, covering both everyday failures and longer-term concerns.

The problem of getting a system to pursue what people actually intended rather than the literal instruction it was given.

Rules and filters wrapped around a model to keep its output within bounds: refusing harmful requests, staying on topic, not revealing private data.

Deliberately attacking an AI system before its users do: probing for ways to make it leak data, produce harmful content or ignore its instructions.

A request crafted to get a model to ignore its own rules, and the reason guardrails are treated as a control rather than a guarantee.

An organisation's stated commitments on fairness, safety and accountability in how it builds and buys AI, and the practices meant to deliver them.

The rules and roles an organisation sets for how AI may be used, who approves it and who answers for what it does.

The written document telling staff what they may and may not do with AI tools, usually the cheapest first control to put in place.

The specific ways an AI use could cause harm or loss, assessed case by case rather than treated as one general worry.

How far you can tell why a system produced a particular answer, which matters most where a decision has to be justified to someone.

Watching a live system's outputs over time so quality problems are noticed from evidence rather than from complaints.

Noticing when results get worse because the world or the inputs have changed, even though nothing about the system was altered.

The whole path a model takes in your organisation, from selection and testing through live use to eventual replacement.

Filtering what goes into or comes out of a system against a policy, the control behind a product refusing certain requests.

Marking generated content so it can later be recognised as machine-made, more reliable in some formats than in others.

A record travelling with a file describing how it was made and edited, aimed at showing where an image or document came from.

Constraining what a model may produce while it produces it, rather than checking and discarding the answer afterwards.

Structured tests that measure whether an AI system actually does its job, and the honest answer to how you would know if it stopped.

Using one model to score another's output at scale, cheaper than human review and needing its own checking before it is trusted.

An occasion where an AI system causes or nearly causes harm, which some regimes now expect to be recorded and reported.

Systematic unfairness in a system's outputs, learnt from its training data or its design, showing up as worse results for some groups.

A model declining to answer, sometimes correctly and sometimes not, and a common source of complaints when the boundary is drawn poorly.

What an AI investment returns against what it costs, counted honestly enough to include the time spent adopting and maintaining it.

One specific job you intend AI to do, defined narrowly enough that you could tell whether it worked.

The set of AI tools an organisation actually runs together, chosen to cover its work rather than bought as one vendor's suite.

A small build to test whether an idea works at all, deliberately short of production quality and often mistaken for it.

A limited live trial with real users and real work, the step where most of what a proof of concept missed shows up.

An AI system people depend on for real work, which brings expectations about reliability, support and cost that a trial never faces.

Staff using AI tools the organisation has not approved, usually with good intentions, and the reason many first policies get written.

The choice between paying for a finished tool and assembling your own, usually decided by whether the capability is core to you.

Everything a tool costs over its life, including setup, training, integration and the staff time to run it, not only the licence.

An organisation's stated plan for where AI is worth using and where it is not, which mostly earns its keep by ruling things out.

How far a tool has actually been taken up in daily work, as opposed to how many people were given access to it.

A small central team that sets standards and supports AI work across an organisation, rather than owning every project itself.

The formal process of getting a purchase approved, which for AI tools often takes longer than the evaluation that preceded it.

Buying and starting on your own with a card, without talking to anyone, which suits small teams and rarely fits regulated work.

Buying through conversations with a vendor's sales team, which is how enterprise terms, security review and custom pricing are reached.

How hard it would be to move off a tool later, worth weighing while switching is still cheap rather than when it is not.

One named person's access under a plan, the unit most software is still sold in and the one AI pricing is moving away from.

The gap between a promising trial and something the organisation runs, where a large share of AI projects quietly stop.

The person in a team who takes a tool up first and brings others with them, usually worth more to adoption than training does.

How long it takes before a tool is genuinely paying back, the measure that separates a quick win from a long programme.

The tier where a vendor's data handling, security and support terms usually change, which is why it is compared separately from features.

Where an organisation configures a tool centrally, including who has access, what is retained and which controls are enforced.

Taking a tool from a working pilot to everyone who should have it, which is a communication and training job more than a technical one.

Information that identifies a particular person, such as a name, email address or identity number, and the category most rules are written around.

The European data protection regime, which shapes how personal data may be used and is often applied by companies well outside Europe.

European legislation regulating AI by how risky its use is, placing the heaviest obligations on the highest-risk applications.

An independent audit report on how a vendor handles security and availability, commonly requested before a tool is approved.

An international certification for running an information security management system, often accepted alongside or instead of a SOC 2 report.

How an organisation decides who owns which data, who may use it and to what standard it is kept.

A durable record of who did what and when, which is what turns a claim about how a system was used into evidence.

Letting staff reach a tool with their existing work login, so access is granted and removed centrally rather than tool by tool.

The contract setting out what a vendor may do with personal data you send it, and usually a requirement before any real data moves.

A vendor agreeing to stand behind you if generated output leads to a legal claim, offered mainly on business plans.

The conditions attached to using a model or its output, including whether commercial use is allowed and what may be trained on.

Information carrying legal obligations of its own, such as health or financial records, where the rules decide the tool rather than preference.

Periodically rechecking who still needs access to what, and removing the rest, which audits ask for and organisations routinely postpone.

Another company your vendor passes your data to in order to deliver the service, which is why the published list is worth reading.

The standard set of questions a buyer sends a vendor before approval, and often the slowest step in adopting a new tool.

An authorised attempt to break into a system in order to find weaknesses first, usually reported on annually to customers.

Permission given freely and specifically for a stated use of someone's data, one of several lawful grounds and not always the right one.

A person exercising their right to see, correct or delete the data an organisation holds on them, within a deadline set by law.

Several agents working on one job with different responsibilities, which can help on large tasks and adds its own coordination problems.

A process where an AI decides the next step rather than following a fixed path, trading predictability for the ability to handle variation.

The parts agents play in a multi-agent setup, typically one planning and delegating while others carry out the individual pieces.

Splitting a large goal into steps small enough to act on, the step most agent failures can be traced back to.

The repeating cycle an agent runs of deciding, acting and checking the result, continuing until the goal is met or a limit stops it.

A toolkit for building agents that supplies the loop, the tool connections and the memory, so each project does not rebuild them.

Testing whether an agent completes real tasks correctly end to end, which is harder to measure than a single answer's quality.

A helper agent given one narrow piece of a larger job, keeping each part focused and its context manageable.

An agent that operates a web browser to complete tasks on sites that offer no other way to reach them.

An AI that works on a codebase across many steps, reading files and making changes, rather than suggesting one snippet at a time.

Confining what an agent can reach and change, so a mistake or a hijacked instruction is contained rather than loose in your systems.

An emerging way for agents from different vendors to talk to each other directly, aimed at the problem of tools that cannot cooperate.

How much an agent may do without asking, the setting that decides most of the value and most of the risk of using one.

The point where an agent passes work to a person or to another agent, and the moment worth designing most carefully.

An agent running in the background without anyone watching, which makes limits, logging and a stop control matter far more.

A reusable request with blanks to fill in, so the same job is asked for the same way every time instead of retyped.

A shared collection of tested prompts a team draws on, which is how prompting stops being one person's private skill.

Breaking a job into several requests where each one's output feeds the next, usually more reliable than asking for everything at once.

Telling a model what part to play, such as a cautious reviewer, to set the tone and standard of what comes back.

How reliably a model does what it was actually told, especially when the request has several conditions that must all hold.

A request organised into labelled sections rather than one paragraph, which makes long instructions easier for a model to follow.

Having a model check and revise its own answer before returning it, which catches some mistakes and costs an extra step.

A model working out the steps before doing any of them, which is what separates an agent from a single answer.

The visible record of steps a model took to reach an answer, useful for spotting where it went wrong and not always complete.

Deciding what a model should have in front of it for a task, now often more decisive than the wording of the request itself.

A setting controlling how widely a model draws when choosing its next word, the usual companion to a temperature setting.

Hidden instructions planted in content a model reads, aiming to make it ignore its real task, and the central security problem of AI tools.

A model revealing the private instructions it was set up with, exposing internal wording that was assumed to stay unseen.

The shape you ask an answer to come back in, such as a table or a list, and the simplest way to make results usable.

The conditions a delivered piece of work must meet to be considered done, written so that both sides can tell whether they hold.

Writing so that answer engines can extract a direct answer from your page, which favours plain statements over brand language.

Tools claiming to identify machine-written text, which are unreliable enough that decisions about people should not rest on them.

Telling readers that content was produced with AI, sometimes a legal duty and increasingly an expectation regardless.

A generated summary shown above search results, which can answer a question without a visit and changes what a ranking is worth.

Altering data so no individual can be identified from it again, which if genuinely achieved takes it outside most privacy law.

A vendor's catalogue of available integrations, which is the practical answer to whether a tool will fit your existing stack.

A point where a person must confirm before the process continues, placed where a mistake would be expensive rather than at every step.

Sending many requests together to be worked through when capacity allows, usually cheaper than asking for each one immediately.

The documented way an organisation writes, which is what an AI tool has to be given before its output can be consistent with everything else you publish.

The written argument for spending on something, setting out the cost, the expected return and what has to be true for it to hold.

The plan for keeping essential work running when a system is unavailable, which for a tool people now depend on daily means having a defined fallback.

California's consumer privacy law, giving residents rights over their personal data and reaching many companies outside the state.

The work of getting people to actually use a new tool, which is where most of the difference between a successful and a failed rollout sits.

Allocating a shared tool's cost back to the teams using it, which changes behaviour faster than any usage policy.

An agreed minimum you undertake to spend over a period, usually traded for a better rate and worth sizing against realistic use.

The check that a regulated system meets its legal requirements before it goes to market, which for higher-risk AI happens before launch rather than after.

The share of conversations an automated system finishes itself, distinct from deflection because a contained conversation may still have ended badly.

The tendency for material in the middle of a very long input to be used less reliably than material at either end, which is why a bigger context window is not the same as better recall.

Moving personal data between countries, which usually requires a specific legal mechanism rather than being a matter of where servers happen to be.

Data being taken out of an organisation without authorisation, whether by an attacker, a departing employee or a tool given too much access.

Collecting and keeping only what the purpose actually requires, which is both a legal principle and the cheapest way to reduce exposure.

The share of incoming enquiries resolved without reaching a person, the headline number in customer support AI and the one most loosely defined between vendors.

Controls that detect and block sensitive information leaving the organisation, increasingly pointed at what staff paste into AI tools.

A structured assessment of privacy risk carried out before a higher-risk processing activity begins, and often expected before an AI deployment touching personal data.

Running a process with the final actions disabled so the intended effects can be reviewed before anything real happens.

Keeping stored data unreadable without a key, so that copying the disk or the backup does not hand over the contents.

Protecting data while it moves between systems, which is what stops it being read by anything sitting on the network in between.

The defined route from an automated system to a person, including what context travels with the case.

What a workflow does with the cases it cannot complete, which is usually where the remaining human work concentrates.

Whether a customer's issue was settled in one interaction, a long-standing support measure that AI both improves and complicates.

Making content likely to be summarised and cited by generative systems, where being the source of an answer matters more than being clicked.

A fixed set of examples with known good answers, kept unchanged so that any drop in quality shows up as a difference rather than a feeling.

How long a case takes from start to finish, useful alongside quality measures and misleading on its own.

A category under European legislation covering uses such as employment and worker management or access to essential services, carrying the heaviest obligations before deployment.

United States legislation governing health information, whose obligations fall on specific categories of organisation rather than on everyone holding health data.

Designing a step so that running it twice has the same effect as running it once, which is what stops a retry from sending the same invoice again.

An outside firm engaged to set a tool up and make it work in your context, common where a product is capable but not self-explanatory.

A company that runs other people's models as a service, competing on price, speed and which models it carries rather than on the models themselves.

An international certification for running an AI management system, the AI counterpart to the security certification many buyers already ask for.

How encryption keys are created, stored, rotated and destroyed, which is where encryption usually succeeds or fails in practice.

A single control that stops an automated system immediately, which is the thing to establish before autonomy is increased rather than after.

The legal ground you rely on to process someone's personal data, of which consent is only one and often not the most appropriate.

Giving each person and system only the access it needs to do its job, which limits how far any single compromise reaches.

A lawful basis relying on a genuine business need balanced against the individual's rights, which has to be assessed and recorded rather than assumed.

A proposed file at a site's root that points AI systems at its most useful content, an emerging convention rather than a settled standard.

Facts an agent keeps across separate tasks and sessions, which makes it more useful over time and makes what it has stored worth being able to inspect.

Converting text from one language to another automatically, now good enough for most working use and still worth reviewing where wording carries obligation.

Paying someone else to run a system for you on an ongoing basis, rather than buying the software and operating it yourself.

Requiring a second proof of identity beyond a password, and the single control that removes most account-takeover risk.

Corrupting a model by tampering with what it learns from, which is a supply-chain concern wherever training data or a model is sourced externally.

The company that trains and owns a model, as distinct from the many products that resell access to it.

Pulling the people, companies, dates and amounts out of free text so they can be filed, matched or acted on.

A United States framework for identifying and managing AI risk, intended for voluntary use and organised around four functions that give an organisation ready-made headings.

Removing a leaver's access across every system they touched, which is easy to do partially and hard to do completely.

When a model learns its training examples too closely and performs worse on anything new, the commonest way a fine-tuning project quietly fails.

AI that forecasts or scores something, such as which customers will churn, as distinct from generative AI that produces new content.

The paid setup, configuration and training a vendor sells alongside its product, often a material share of first-year cost.

Paying to reserve a guaranteed amount of model capacity, which removes queueing at busy times and is charged whether you use it or not.

Replacing identifying details with references that can still be reversed with separate information, which reduces risk without removing the data from scope.

Using data only for the purpose it was collected for, which is what makes reusing an existing dataset to train a model a question rather than a given.

Granting access by job role rather than person by person, so joiners and leavers are handled by changing a role rather than remembering a list.

The internal register of what personal data an organisation handles, why and where it goes, which regulators expect to exist before they ask for it.

Asking for another attempt at the same request, which produces a different answer because the model does not work deterministically.

The share of cases actually solved rather than merely closed, which is why it is worth verifying against what customers do next.

The rules for what happens when a step fails, including how many attempts are made and how long the gap between them grows.

What you will do if a change has to be undone, written before the change rather than discovered during it.

A standard for creating and removing user accounts automatically from your identity system, so access follows employment without manual steps.

The explicit list of what an agent is permitted to do, defined in advance so that anything outside it fails rather than being attempted.

Training where the material supplies its own answers, such as hiding the next word and scoring the guess, which is what made training on the open web possible.

Model access with no capacity to manage, where you are charged per request and the provider handles the machines behind it.

A hard ceiling on what an account can consume, the control that turns an open-ended usage bill into a known maximum.

Pre-approved contract terms used to make an international data transfer lawful, and a routine item in vendor agreements.

The document defining what an engagement will deliver, by when and for how much, and the thing to fix in detail before work starts.

A case that completes end to end with no human touch, and the number most automation business cases are ultimately measured on.

Returning an answer word by word as it is produced rather than all at once, which changes how fast a tool feels without changing how fast it is.

Machine-readable markup describing what a page contains, which is how a search or answer engine knows a page holds a product, a FAQ or a definition.

The specific results a pilot has to produce to be judged worth continuing, agreed before it starts rather than argued about at the end.

Training on examples where the right answer is already attached, which is how most classification and prediction systems are built.

Exposure created by the vendors, models and libraries you depend on rather than by anything inside your own systems.

A firm whose business is connecting products that were not designed to work together, usually engaged for larger or older estates.

A written account of who might attack a system, how, and what it would cost you, used to decide which controls are worth the effort.

Cutting text into the chunks a model actually reads, which is why word counts and token counts never quite match and why some languages cost more.

A legal requirement to tell people they are dealing with an AI system or looking at generated content, separate from any duty about how well it works.

A limited window to evaluate a tool in real conditions, worth planning as a test with criteria rather than treating as free access.

A record of a document's earlier states, which matters more once drafts are produced quickly enough that the previous one is easy to lose.

A lower unit price at higher usage, which rewards consolidation onto one vendor and quietly raises the cost of leaving.

What an agent holds while a task is in progress, which is separate from anything it remembers between sessions and is the first thing to fill up.

An approach that verifies every request rather than trusting anything because of where it came from, which suits work spread across many cloud services.

China's national data-protection law, which governs how any business handling the personal data of people in China manages consent, cross-border transfers and security, whatever country the business itself is based in.

Japan's core data-protection law, enforced by the Personal Information Protection Commission, so any business handling personal data belonging to people in Japan needs to meet its consent and security requirements.

Singapore's data-protection law, enforced by the Personal Data Protection Commission, which sets the baseline rules on consent, notification and data security for handling personal data of people in Singapore.

India's data-protection law, whose provisions commence in stages, so a business handling the personal data of people in India needs to check which sections are already in force rather than assume the whole Act applies.

South Korea's framework law for AI, which pairs promoting the industry with preventing its risks and places duties for transparency, safety and operator responsibility.

China's interim rules for generative AI, which govern providers offering such services to the public in China.

Singapore's published AI governance guidance, extended with editions for generative and agentic AI, whose nine named dimensions a business can use to benchmark its own governance.

A popular name rather than an official title for Colorado's AI law, whose commencement was extended and whose provisions were then repealed and reenacted with different requirements.

The European rules for providers of general-purpose AI models, which is why a business building on a foundation model should check whether its supplier publishes the required technical documentation and training-data summary.

The working understanding needed to use AI tools responsibly, which European law turns from a good idea into a duty on providers and deployers to take measures supporting it among their staff.

The mechanism that lets a model weigh which parts of an input matter most for each word it produces, and the idea the whole transformer design is built around.

Sorting items into predefined categories, such as routing an incoming email as a complaint, a query or a renewal.

A common agent pattern where the model alternates between reasoning about what to do and taking an action, using each result to decide the next step.

The share of time a service is actually available, usually published as a percentage and worth reading alongside what the vendor commits to contractually.