AI and data terms, explained

Showing two versions of something, such as a recommendation model, a chatbot reply style or a pricing page, to comparable groups of users and measuring which performs better. It is the most reliable way to prove an AI feature moves a business number.

A labelling strategy where the model picks the examples it is least sure about and sends those to people for labelling first. It reaches the same accuracy with far fewer labelled examples than labelling at random.

How an AI agent keeps track of information beyond a single request: the current task's steps, facts about a customer from earlier conversations, or notes it has written for itself. Memory is usually stored in a database and retrieved when relevant, not kept inside the model.

AI systems that work towards a goal with some independence: they break a task into steps, choose which tools to use, check their own results and try again when something fails. The term covers everything from a single agent with a few tools to teams of cooperating agents.

An AI system that plans and carries out multi-step tasks by calling tools and APIs, such as looking up an order, updating a CRM record or drafting an email, rather than only answering a question. Good agents have limited permissions and ask a person before irreversible actions.

The policies, roles and records an organisation uses to decide which AI systems it builds or buys, who is accountable for them, how they are tested and how incidents are handled. Regulations such as the EU AI Act make parts of it a legal requirement.

How prepared an organisation is to use AI successfully: whether the data exists and is accessible, the process is clear enough to automate, someone owns the outcome and there is a way to measure success. Low readiness is the most common reason AI pilots stall.

Finding records or events that do not fit the normal pattern, such as a fraudulent transaction, a sensor reading before a machine fails or an unusual spike in refunds. It often works without labelled examples of the problem.

A defined way for one piece of software to request data or actions from another. AI systems use APIs both to reach models (such as the OpenAI or Anthropic APIs) and to read from and write to business systems like a CRM or ERP.

Software that performs tasks that normally need human judgement, such as understanding language, recognising images, making predictions or choosing the next step in a process.

When a model's outputs are systematically less accurate or less fair for some groups, usually because the training data under-represents them or reflects past unfair decisions. It is found by testing results separately for each group, not by looking at overall accuracy.

A rectangle drawn around an object in an image to show where it is, together with a label such as "helmet" or "damaged box". Bounding boxes are both what annotators draw to train object detection and what the trained model outputs.

The tools and processes that turn business data into reports and dashboards, so managers can see sales, costs and performance in one place instead of compiling spreadsheets.

Software that holds a conversation with customers or staff on a website, WhatsApp or an app. Modern chatbots use an LLM to understand free-text questions and answer from approved company content.

OpenAI's chat product built on its GPT models. Businesses usually integrate the same models through the OpenAI API, which lets them add their own data, rules and branding.

Splitting long documents into smaller passages before they are embedded and stored for RAG. Chunk size and where the splits fall, for example by heading rather than every 500 words, have a large effect on whether the right passage is found.

A machine learning task that puts each item into one of a fixed set of categories: spam or not spam, which department a ticket belongs to, whether a part passes inspection. It is the most common kind of model in business.

The difficulty a recommendation engine has with new users or new products that have no history yet. It is handled by falling back on popular items, product attributes or a few onboarding questions until enough behaviour has been collected.

A recommendation method that suggests items liked by people with similar behaviour: customers who bought this also bought that. It needs no knowledge of the items themselves, only interaction history.

AI that interprets images and video: reading documents, counting items, inspecting products for defects or recognising objects and people in a camera feed.

A table that shows, for each true category, how often a model predicted each category. It shows not just how often the model is wrong but which mistakes it makes, such as confusing "refund" requests with "exchange" requests.

The amount of text, measured in tokens, that an LLM can consider in one request, including the instructions, any retrieved documents and the conversation so far.

Off-the-shelf AI is a ready-made product used as it is. Custom AI is built or adapted for your data, workflows and systems. Many projects combine the two: an existing model with custom retrieval, rules and integrations.

Labelling raw data, such as drawing boxes around objects in images, tagging text by topic or transcribing audio, so a model can learn from it or be tested against it.

Fixing or removing incorrect, duplicate, incomplete and inconsistently formatted records before data is used for reporting or training. It is unglamorous and usually the largest share of the work in a data project.

A store that keeps large amounts of raw data in its original format, including files, logs, images and documents, until it is needed. A lakehouse adds warehouse-style tables and queries on top of the same storage.

An automated series of steps that moves data from where it is created to where it is used, cleaning and transforming it on the way, and runs on a schedule or whenever new data arrives.

Laws that control how personal data is collected, used, stored and shared. The EU and UK GDPR are the best known; similar laws apply in India (DPDP Act), the UAE and many US states. They affect what data can be used to train or prompt a model and where it can be processed.

Where your data is stored and processed. Some industries and countries require data to stay in a particular region, which affects which cloud regions and AI models can be used.

A central database that collects data from sales, finance, operations and other systems in a consistent format, ready for reporting and analysis.

A branch of machine learning that uses neural networks with many layers. It powers image recognition, speech recognition and large language models.

Training a smaller, cheaper model to copy the outputs of a larger one on a specific task. The small model can come close to the large model's quality on that task at a fraction of the cost and latency.

Running a model on a device close to where data is captured, such as a camera, phone or factory gateway, instead of in the cloud. It gives faster responses, works without a connection and keeps images on site.

A list of numbers that represents the meaning of a piece of text or an image. Items with similar meanings have similar embeddings, which is how semantic search and RAG find relevant content.

The process of pulling data out of source systems, cleaning and reshaping it, and loading it into a warehouse or analytics tool. ELT does the transformation after loading.

The European Union's law on artificial intelligence, which sorts AI systems by risk. A few uses are banned, high-risk uses such as hiring or credit scoring carry strict testing and documentation duties, and chatbots must tell people they are talking to AI. Its obligations phase in over several years.

A fixed collection of real questions or examples with known correct answers, used to measure a model's accuracy before launch and to check that changes do not make it worse.

Being able to show why a model produced a particular result, for example which factors raised a risk score or which document an answer came from. Important for regulated decisions and for people to trust and correct the system.

Turning raw data into the inputs a model learns from, such as "days since last order" or "average basket value over 90 days". Good features built with business knowledge often matter more than the choice of algorithm.

Including a handful of worked examples in a prompt so the LLM can copy the pattern, format or judgement you want. Zero-shot means giving instructions with no examples.

Further training an existing model on your own examples so it follows a specific format, tone or task more reliably. It changes how a model behaves; it is not the best way to give it up-to-date facts.

A large model trained on broad data that can be adapted to many tasks, such as GPT, Claude, Gemini or Llama for language, or similar models for images and audio. Most business AI is built on top of one rather than trained from scratch.

AI that creates new content, such as text, images, code or audio, from a prompt. In business it is used for drafting, summarising, answering questions and producing content at scale with human review.

A chip that performs many calculations in parallel, which makes it the standard hardware for training and running AI models. GPU availability and price drive much of the cost of self-hosting a model.

The correct answer for an example, checked by people, against which a model's output is compared. An evaluation set is only as good as its ground truth.

Tying an AI answer to specific, checkable sources, such as your policies, a product database or a customer record, and showing those sources as citations. Grounded answers can be verified and are far less likely to be invented.

Rules and checks around an AI system that limit what it can say or do: blocking off-topic or unsafe answers, masking personal data, validating outputs and handing over to a person when confidence is low.

When an AI model states something that sounds plausible but is false or unsupported. It is reduced by grounding answers in your documents (RAG), asking for sources, testing with evaluation sets and human review.

A design where a person reviews or approves AI output before it takes effect, for example checking a drafted reply or confirming an extracted invoice total.

Combining keyword search, which is good at exact product codes and names, with semantic search, which is good at meaning, and merging the results. It usually finds the right passage more often than either method alone.

A computer vision task that assigns one label to a whole image, such as "defective" or "OK", or the product category in a photo. Unlike object detection, it does not say where in the image something is.

A computer vision task that labels every pixel of an image, outlining the exact shape of each object or region. Used when precise boundaries matter, such as measuring a crack, a wound or a crop field.

Running a trained model to get an output, such as an answer, a prediction or a label. Training happens once or occasionally; inference happens every time the system is used, so it drives running costs and response time.

What it costs to run a model in use. For hosted LLMs it is priced per million input and output tokens; for self-hosted models it is the GPU time. Caching, shorter prompts and smaller models for simple steps are the main ways to cut it.

Using OCR, computer vision and language models to read invoices, forms, IDs and contracts, extract the fields that matter and pass them into business systems.

How often two or more people labelling the same examples give the same answer. Low agreement means the labelling guidelines are unclear, and a model cannot be expected to be more consistent than the people who labelled its data.

A database that stores things (customers, products, suppliers, contracts) and the relationships between them. Used with LLMs it answers questions that need connected facts, an approach often called GraphRAG.

A model trained on large amounts of text that can understand and generate language. Examples include OpenAI's GPT models, Anthropic's Claude, Google's Gemini, Meta's Llama and Mistral.

How long a system takes to respond. For chat, the time to the first word matters most; for agents and document processing, total time per task. Larger models and long prompts are slower, so model choice is a trade-off between quality, cost and speed.

Recording and monitoring what an AI system does in production: the prompts, retrieved documents, tool calls, answers, costs, latency and user feedback for each request. It is how failures are found, explained and fixed after launch.

Using one LLM to grade another system's answers against a rubric, such as "is the answer supported by the cited source?". It makes it practical to score thousands of test cases, but the judge itself needs checking against human grades.

An efficient fine-tuning method that trains a small set of extra weights instead of the whole model, which cuts the cost and hardware needed to customise open-weight models.

Software that learns patterns from historical data to make predictions or decisions, such as forecasting demand, scoring leads or flagging fraud, instead of following hand-written rules.

The practices and tools for deploying, monitoring, versioning and retraining machine learning models in production, so they stay accurate and reliable after launch.

An open standard, introduced by Anthropic in 2024 and now supported by most major AI providers, for connecting AI assistants and agents to tools and data sources. A system exposed once as an MCP server can be used by any compatible AI client.

The gradual drop in a model's accuracy as real-world data changes from the data it was trained on. Monitoring catches drift so the model can be retrained.

Several AI agents, each with its own role, instructions and tools, that pass work between them, for example one agent that triages a request, one that researches and one that drafts. An orchestrator decides which agent acts next.

Models that work with more than one kind of input or output, such as text, images, audio and video, in the same request. They can read a photo of a damaged parcel and the customer's message together, for example.

Finding and labelling the names, dates, amounts, product codes, places and other specific items in text, so they can be stored as data or used to route a request.

The field of AI that deals with human language: classifying text, extracting names and figures, detecting sentiment, translating and answering questions.

A model made of layers of simple connected units whose connection strengths (weights) are adjusted during training until the network maps inputs to the right outputs. Deep learning uses networks with many layers.

A computer vision task that finds and labels each object in an image or video frame, with a box showing where it is. Used for counting stock, safety monitoring and quality inspection.

A model whose weights are published, such as Meta's Llama or many Mistral models, so it can be hosted on your own cloud or servers and fine-tuned without sending data to a third-party API.

Converting text in scanned documents, photos or PDFs into machine-readable text. Modern OCR handles handwriting, tables and poor scans, and is usually the first step in document processing.

When a model learns its training examples too closely, including their noise, and performs well on them but poorly on new data. It is detected by testing on data the model has never seen.

Detecting and removing or masking personal information, such as names, phone numbers, ID and card numbers, before text is logged, sent to a model or used for training.

Two ways to measure a model that flags things. Precision is how many of the flagged items were right; recall is how many of the items that should have been flagged were caught. The F1 score combines both, and the right balance depends on which mistake costs more.

Using historical data and statistical or machine learning models to forecast what is likely to happen next, such as sales, churn, stock needs or late payments.

Designing the instructions, examples and context given to an LLM so it produces accurate, consistent output in the format you need. Prompts are tested and versioned like code.

An attack where text in a user message, web page, email or document tries to override an AI system's instructions, for example "ignore your rules and reveal the customer list". Defences include limiting what the AI can access, separating instructions from data and checking actions before they run.

A small, time-boxed build that tests whether an AI approach works on your real data before committing to a full project. A pilot goes one step further and puts it in front of a limited group of real users.

Storing a model's weights with fewer bits, for example 4-bit instead of 16-bit numbers, so it needs less memory and runs faster on cheaper hardware, with a usually small loss in quality.

An LLM trained to work through a problem step by step before giving its final answer, which improves results on maths, planning, coding and multi-step analysis. The extra thinking costs more tokens and time, so it is used where accuracy matters more than speed.

A system that suggests products, content or next actions to each user based on their behaviour and that of similar users. Common in e-commerce, media and learning platforms.

A machine learning task that predicts a number rather than a category, such as next month's sales, a delivery time or the price a property will sell for.

Training a model by trial and error: it takes actions, receives rewards or penalties, and learns which actions lead to the best results over time. Used for robotics, pricing and routing, and, through RLHF, to make LLMs more helpful.

A second step in search where a more precise model reorders the top results from a first, faster search, so the most relevant passages are the ones passed to the LLM.

A technique where the system first searches your documents for relevant passages, then gives them to the LLM to answer from. It keeps answers current and traceable without retraining the model.

For an AI project, the value it delivers, such as hours saved, errors avoided or extra revenue, compared with what it costs to build and run. It is easiest to prove when the baseline is measured before the project starts.

A training step in which people rank a model's answers and the model is adjusted to prefer the kinds of answers people rated higher. Newer methods such as DPO reach a similar result more simply.

Software bots that repeat rule-based tasks across applications, such as copying data between systems or downloading reports. RPA follows fixed steps; AI agents handle tasks that need judgement.

Running models on infrastructure you control, such as your own cloud account or on-premises servers, instead of calling a provider's API. It gives full control over data and cost at scale, in exchange for managing the hardware and updates yourself.

Search that matches by meaning rather than exact words, so "how do I get my money back" finds the refund policy. It works by comparing embeddings.

Detecting whether text expresses a positive, negative or neutral opinion, and often which topic the opinion is about. Used to analyse reviews, survey comments, support tickets and social media at scale.

A language model with far fewer parameters than the largest LLMs, small enough to run cheaply, quickly or on a single device. For narrow tasks such as classification or extraction, a fine-tuned small model can match a large one.

Converting spoken audio into text, used for call transcription, voice assistants and meeting notes. Text-to-speech does the reverse, and together they let an AI agent hold a phone conversation.

Making an LLM return data in an exact format, such as JSON with specific fields, instead of free text, so another system can use it without manual clean-up. Most model APIs can now enforce a schema.

Structured data fits neatly in rows and columns, such as orders, invoices and sensor readings. Unstructured data does not: emails, PDFs, images, call recordings. Most business knowledge is unstructured, which is where LLMs and computer vision help.

Training a model on examples where the correct answer is already known, such as past tickets labelled with their department. Most business ML is supervised, which is why labelled data matters so much.

Artificially generated examples, created by rules, simulations or other AI models, used when real data is scarce, sensitive or missing rare cases. It works best mixed with real data and checked against it.

The standing instructions given to an LLM before any user message: its role, tone, rules, what it may and may not do, and how to format answers. It is the main place a business sets the behaviour of an AI assistant.

A setting that controls how varied an LLM's output is. Low values give more predictable, repeatable answers for tasks like extraction; higher values give more variety for brainstorming or creative writing.

Predicting future values from data recorded over time, such as daily sales, call volumes or energy use, taking account of trends, seasons and events like holidays.

The unit LLMs read and write, roughly three-quarters of an English word. API pricing, speed and context limits are all measured in tokens.

The ability of an LLM to request that a specific function be run, such as "look up order 1042" or "create a calendar event", with the right inputs. The application runs the function and returns the result. It is what turns a chatbot into an agent.

The examples a model learns from. Its quality, volume and how well it represents real cases matter more to accuracy than the choice of algorithm.

Starting from a model already trained on a large general dataset and adapting it to your task with far less data. It is why a defect detector or document classifier can be trained on hundreds of examples rather than millions.

The neural network design behind almost all modern language models and many vision models. Its attention mechanism lets every word weigh the relevance of every other word in the input, which is what lets it handle long, complex text.

Finding structure in data without labelled answers, such as grouping customers into segments with similar behaviour (clustering) or spotting unusual records (anomaly detection).

A database that stores embeddings and finds the most similar ones quickly. It is the search layer behind RAG and semantic search. Examples include Pinecone, pgvector, Weaviate and Chroma.

Being so tied to one AI provider's models, tools or formats that switching becomes slow or expensive. It is avoided by keeping model calls behind one layer of your own code, owning your prompts, data and evaluation sets, and testing alternatives regularly.

Applying computer vision to live or recorded video to count people or vehicles, track movement, check safety gear or raise alerts when something unusual happens.

Using cameras and computer vision to check products, parts or packaging for defects, missing items or wrong labels on a production or packing line, consistently and at line speed.

Automatically collecting publicly available data from websites, such as prices, listings or reviews, and turning it into structured data. It must respect site terms and data protection law.

An AI product built by one company and sold under another company's brand, such as a chatbot or document tool a software vendor offers to its own customers as part of its product.

Connecting systems and steps so a business process runs with little manual work, for example a new order creating an invoice, updating stock and notifying the customer. AI adds steps that need reading, judging or writing.