AI glossary

An agent activity report is a periodic summary of what AI agents have done. It shows closed cases, hand-offs to people and errors, drawing its data from the audit trail. It resembles a team's weekly report in which every number can be checked.

Agent autonomy is the range of decisions and actions an agent takes without human approval. The company sets it, from only drafting proposals to handling a case end to end. It works like a power of attorney that states which matters, and up to what amount, an employee may sign alone.

An agent identity is an AI agent's own account in the company's identity system, separate from human accounts. It works like an employee badge: you know who did what, and access can be revoked. Microsoft documents this type of account in its Microsoft Entra Agent ID service.

The agent loop is the cycle in which an agent picks a step, carries it out, checks the result and decides on the next one. It ends when the goal is reached, an obstacle appears or the step limit runs out. Each pass is a separate call to the model, so the number of steps drives the cost.

Agent memory is the information an agent keeps between steps or between cases: notes, history and agreed facts. The model itself remembers nothing, so the program around it stores the memory. For a company it is another place holding data, which needs rules for access, retention and deletion.

Agent observability is a record of every step an agent takes: what it read, which tool it called, how many tokens it used and how long the task took. After an error you can replay the run, like a call-centre recording. Nobody has to guess what went wrong.

Agent orchestration is how one program or agent splits work across several AI agents. It sets the order of steps, passes results along and assembles the final output. Think of a project manager who assigns tasks to the team and makes sure nothing gets stuck.

A sandbox is an isolated environment where an agent works without access to real systems and data, or with tightly limited access. It is like a trial period on copies of documents: a mistake never touches production. Do not confuse it with an AI Act regulatory sandbox, which means testing under a regulator's supervision.

A skill is a written procedure for an agent: a folder with a SKILL.md file, sometimes with templates and scripts. The agent loads it only when a case calls for it, and a plugin bundles several skills and connections into one package. Anthropic published the skills format as an open standard in December 2025.

A2A (Agent2Agent) is an open protocol through which AI agents from different vendors pass tasks and results to each other. MCP connects an agent to a system; A2A connects an agent to another agent. Google announced A2A in April 2025 and handed it to the Linux Foundation in June.

Agentic AI is the umbrella term for AI systems that plan and carry out tasks on their own. It covers single agents, teams of agents and the tools that control them. In vendor offers it is often a marketing label, so ask what the system actually does by itself.

An AI agent is a program that receives a goal and chooses the next steps itself to reach it. A language model picks each step, and tools fetch data and act in the company's systems. It works like a new employee with a written procedure and a key card to selected cabinets.

An agent harness is the program around the model that executes its requests, enforces permissions and records the course of work. The model itself only writes text, including requests to use a tool. The harness decides whether a request is allowed or must wait for human approval.

AI agent management reporting means management reports that agents compile from data in company systems. It works like a controller who gathers figures from sales, finance and the warehouse every Monday. Every number should link back to its source so management can verify it.

AI agent security is the set of safeguards that limit what an agent can do in company systems. Agents take actions, so a malicious instruction hidden in an email can end in a data leak. The basics are a dedicated agent identity, least privilege, human approval for risky steps and a complete audit trail.

AI and your team is about how work gets redistributed when AI tools and agents take over some tasks. People move from repetitive steps to checking results, handling exceptions and managing customer relationships. The change works when the team understands the tool, has been trained and helps set the rules.

An AI assistant answers questions and drafts content, but a person carries out every action. An AI agent handles a case itself across several steps and systems, within the permissions it has been given. The assistant is an adviser at your desk; the agent is an employee who gets the job done.

AI data leakage happens when confidential information leaves the company through an AI tool. An employee pastes a contract into a public chatbot, or an agent sends a file to the wrong address. It resembles an email with an attachment sent to the wrong person, except it can happen automatically and in many cases at once.

AI governance is the set of rules, roles and procedures for AI in a company. It defines who decides on AI use, how risk is assessed and who answers for the outcome. Like corporate governance, it covers a policy, a system inventory, regular reviews and a way to report problems.

A hallucination is a confident-sounding model answer that is false, such as a regulation that doesn't exist. It is like a new hire who fills in from memory instead of saying “I don't know”. At work, sources, checks and human approval keep it in check.

AI implementation is moving an AI tool from a trial into everyday work within a specific process. It covers choosing the process, data access, oversight rules and training people. It resembles onboarding a new hire: a job scope, permissions, a supervisor and a probation period.

AI in customer service means AI models that sort incoming requests and draft replies. A chatbot talks to the customer on its own, while an assistant suggests replies to a human consultant who sends them. Companies usually start with cases where a mistake costs little, such as order status.

AI in finance means agents and models that match invoices, reconcile accounts and explain results. A predictive model forecasts cash flow, while a language model explains variances in plain words. An accountant approves exceptions and postings, because responsibility for the financial statements stays with people.

AI in HR means tools that screen CVs, schedule interviews and answer staff questions. The AI Act classifies AI used to select candidates or evaluate employees as high-risk (Annex III). The obligations, including human oversight and bias checks, apply from 2 December 2027.

AI in manufacturing and logistics forecasts demand, plans routes and predicts machine failures. Computer vision checks quality on the line, and an agent prepares orders to suppliers. The models rely on data from production systems, sensors and the warehouse, so data quality decides the outcome.

An AI incident is an event where an AI system or agent caused harm or could have: it sent data to the wrong recipient, ran a wrong operation or was manipulated. Incident response is a plan written in advance. It says who stops the agent, who checks the log, who informs customers and who fixes the rules.

AI literacy is the knowledge that lets employees use AI consciously: understand what a tool can do, where it goes wrong and what not to paste into it. Since Article 4 of the AI Act was amended with effect from 27 July 2026, companies must take measures to support these skills among their staff. They need not guarantee a specific level, and the Commission and Member States are to help them, especially SMEs.

An AI model is a learned set of patterns that turns input data into an output: an answer, a forecast or a classification. It resembles an experienced employee who knows the patterns from thousands of cases, though it can repeat fragments of some of them word for word. The model does nothing on its own; it needs software that feeds it data and uses the result.

AI models in 2026 come in families from several vendors: GPT (OpenAI), Claude (Anthropic), Gemini (Google), Qwen (Alibaba), DeepSeek and Mistral. Each family has tiers: a flagship for hard tasks, and fast, cheap models for high-volume work. Choosing a model is like choosing a specialist: one for reviewing a contract, another for sorting email.

An AI Officer is the person responsible for AI rules, security and rollout in a company. A Chief AI Officer (CAIO) is the same role at executive level, alongside the heads of IT and security. In smaller companies the role is often combined with another position.

An AI pilot is a short, limited test of an AI tool in one process with a small team. Before it starts, the company measures how the process works today and sets a success criterion. After the test, management decides whether to scale it, fix it or stop it.

AI red teaming is a controlled attack on your own AI system: a team deliberately tries to push the model or agent into an error, a data leak or a forbidden operation. It is like a fire drill: better to find the weak spot in an exercise than in production, in front of a customer. The findings feed fixes to rules and permissions.

AI ROI is the gain from a deployment minus its costs, divided by those costs. The benefit is measured against the situation before deployment, for example the time and cost of one case. Without that baseline the result is an impression, not a number.

The AI supply chain is every third-party component an AI system is built from: the model, MCP servers, plug-ins, libraries and data. Any of them can carry a vulnerability or malicious code, like a subcontractor with a key to the office. So before connecting one, you check its source, version and permissions.

An AI system audit is an independent check that an AI system or agent works within the company's rules and the law. The auditor reviews its permissions, the data it uses and who approved its decisions. As in a financial audit, nobody takes it on trust: log records are compared with what was supposed to happen.

An AI system inventory is a list of every AI tool and agent the company uses: what it is for, what data it processes, who owns it and what risk it carries. It resembles a fixed-asset register. Without it, it is hard to tell which AI Act and GDPR obligations apply to the company.

AI, machine learning and deep learning nest inside each other like a company, a department and a team. AI is the whole field, and machine learning is software that learns rules from examples. Deep learning is machine learning on multi-layer neural networks; today's language models run on it.

An AI winter is a period when interest in and funding for AI research drop after promises go unfulfilled. One example: the British Lighthill report of 1973 judged the field's results critically, and the government cut its support. For a company it is a warning that an AI project must justify itself with results in a process, not with hype.

Algorithmic bias is a systematic error that makes a model treat certain groups or situations worse. It usually comes from the data: if a company hired mostly one candidate profile for years, the model will treat it as the standard. That is why model results are checked separately for different groups before they affect decisions about people.

Anomaly detection is searching data for events that depart from the norm, such as an unusual invoice or a login at three in the morning. It works like an experienced accountant who sees at once that an amount “doesn't fit”. The model flags suspicious cases, and a person decides whether it is an error, fraud or just an exception.

AGI (artificial general intelligence) is a hypothetical system that does most intellectual work at least as well as a person. Definitions vary; OpenAI's 2018 charter describes highly autonomous systems that outperform humans at most economically valuable work. A company's rollout plan should rest on what models can do today.

Artificial intelligence (AI) is software that infers from input data how to respond or act. The EU AI Act gives predictions, content, recommendations and decisions as example outputs of an AI system. In a company, AI reads documents, sorts cases and suggests next steps; the company decides which decisions stay with a person.

The attention mechanism is the part of a transformer that works out which words in a text relate to each other. It is like reading a contract with a highlighter: at the word “penalty”, your eye goes back to the deadline and the amount. This lets the model connect information that sits many pages apart.

An audit trail is a record of who did what, when and why in a system. For an AI agent it also covers the instruction, the tools used, the input data and the human decision. It works like a front-office correspondence log: a month later you can reconstruct every case.

Authentication checks who a user or agent is; authorization decides what they are allowed to do. At reception someone checks your ID, and your access card opens only certain floors. An AI agent should go through both steps, like an employee: first sign-in, then a permission check for every action.

Backpropagation is a method for identifying which connections in a neural network are responsible for an error in the output. Gradient descent then corrects them in small steps, like a company fixing the faulty stage of a process after a complaint. The method was popularised by Rumelhart, Hinton and Williams in a 1986 Nature paper.

A baseline is a measurement of a process before AI is deployed. Like stepping on the scales before a diet: without the first reading you cannot tell what changed. It records the time, cost and number of errors per case, using the company's own data.

A benchmark is a standard test for comparing models; an evaluation (evals) checks a model on the company's own examples. A leaderboard tells you who passed the general exam; an evaluation tells you who can handle your invoices. Before deployment, the second one matters more.

Bielik and PLLuM are Polish families of open language models, trained with a focus on Polish. Bielik is developed by SpeakLeash and ACK Cyfronet AGH, and PLLuM by the HIVE AI consortium led by NASK. On 21 May 2026 Poland's Ministry of Digital Affairs announced 11 new PLLuM models, from 4 to 70 billion parameters.

Big data means datasets so large, fast-growing or varied that a spreadsheet or a single database can no longer handle them. An example is a full year of transactions from every till in a retail chain. To use ready-made AI models, a company does not need “big” data, only the right data for a given process.

Computer use is an AI agent's ability to operate a computer the way a person does: it looks at the screen, clicks and types. A browser agent does the same but only inside a web browser; Anthropic released computer use in public beta on 22 October 2024. Both help with older systems that have no API, meaning no entry point built for other programs.

Business process automation hands repetitive steps of a process to software instead of people. Traditional tools follow rigid rules, while AI agents can also handle free-text documents and some exceptions. The result depends on whether the process is written down and someone owns the outcome.

A chatbot is a program that talks with users in a chat window, on a website or in a messaging app. Older chatbots follow a fixed script, while newer ones use language models. A chatbot usually only answers and does not act on its own in company systems the way an agent does.

ChatGPT is OpenAI's AI assistant, released to the public on 30 November 2022. It is like a capable intern: quick at drafting and summarising, but it doesn't know your systems on its own. Companies buy the Business (formerly Team) or Enterprise editions, which add an admin console and separate data terms.

DeepSeek and Qwen (Alibaba) are Chinese families of language models, many of them with open weights. DeepSeek's own app and API store user data in China. Run on your own infrastructure or in Microsoft Foundry, they send no data to the maker, but still need your own testing.

Claude is a family of language models and an AI assistant made by Anthropic. Companies use it in the app, through the API, or in the Microsoft, Amazon and Google clouds. As of 26 September 2026 the newest model is Claude Opus 5.5, released on 22 September 2026, while the top, more expensive tier is Claude Fable 5.1.

A model platform is a cloud service offering models from many vendors under one contract. It works like a wholesaler with a single account: switching models is a settings change, not a new tender. Examples: Microsoft Foundry, Amazon Bedrock and Gemini Enterprise Agent Platform (formerly Vertex AI).

A coding agent is an AI agent that reads, writes and tests program code on its own. One example is Anthropic's Claude Code, generally available since May 2025. A developer hands it a task like they would to a junior colleague, then reviews and approves the changes.

A command-line interface (CLI) is a text-based way to operate programs: instead of clicking, you type a command. AI agents handle a CLI well because a command is plain text. What matters for a company is which commands an agent may run and with what permissions.

An AI policy is an internal document that tells employees which AI tools they may use, with which data, and who approves new uses. It works like the rules for a company car: who may drive it, where to and on what terms. It works best when the rules are mirrored in system permissions.

Computer vision is the field of AI that lets computers recognise what is in photos and video. In a company it reads scanned invoices, counts stock on shelves and spots product defects on the production line. It works well with repeatable shots and less well in poor light or with unusual cases.

Context engineering is selecting everything a model sees for a task: instructions, documents, data and tools. It is like preparing a case file for a stand-in: the full set of documents, with no spare pages. For agents it often matters more than the wording of the instruction itself.

A context window is the limit of text a model can see at once, including its own answer. It works like a desk: whatever doesn't fit has to be filed away or summarised. As of September 2026, leading models have windows of around a million tokens.

The cost of AI in a company covers licences, token fees, integrations, maintenance, people's time and oversight. Licences and integrations are planned up front, while tokens and oversight grow with the number of cases. That is why cost is measured per resolved case, not by the token price.

The cost per agent task is the sum of charges for every step an agent takes on one case. At each step the model usually rereads all the work so far, so later steps cost more. Add the minutes of the person who checks the result to that amount.

A data processing agreement (DPA) is the Article 28 GDPR contract a company signs with a vendor that processes personal data on its behalf, such as an AI tool provider. It sets out what the vendor may do with the data, how it protects it and who its sub-processors are. It resembles the contract with an accounting firm you hand employee records to.

A DPIA is a data protection impact assessment that the GDPR (Article 35) requires before processing likely to create a high risk for people, for example when using new technologies. In practice it is a written analysis: what data, for what purpose, what risks and what safeguards. For an AI rollout, the company usually prepares it with its data protection officer.

Data protection in AI is control over which data goes into AI tools and where it ends up. What matters is the contract with the provider, a ban on training models on company data and where processing happens. An employee pasting a contract into a personal chatbot takes it out of the company as surely as on a USB stick.

Data quality means data are complete, correct and up to date. Data readiness means they can be pulled out of company systems and handed to a model in usable form. If customers have duplicate records in the CRM, AI will repeat that mess in its answers.

Data residency is the country or region where an AI vendor stores and processes a company's data. For example, on 26 February 2025 Microsoft announced that, under its EU Data Boundary programme, it stores and processes EU and EFTA customer data for Microsoft 365, Dynamics 365, Power Platform and most Azure services within those regions. Before a rollout, check whether the specific service and model fall inside that boundary and what exceptions apply.

Data science is collecting, cleaning and analysing data to answer a business question. It combines statistics, programming and knowledge of the industry. A data scientist often builds predictive models but starts from the question, not from the tool.

A decision tree is a model that reaches a result through a series of yes/no questions, such as “does the amount exceed the limit?”. A random forest combines hundreds of such trees and takes the majority answer or averages their results. A tree is easy to show an auditor like a procedure chart; a forest is usually more accurate but harder to explain.

Deep learning is machine learning on neural networks with many layers. Successive layers pick up increasingly complex patterns: in images from edges to faces, in text from words to the meaning of a sentence. Language models, speech recognition and document reading all run on this method.

A deepfake is an image, voice recording or video generated or altered by AI so that it looks real. A typical business risk is a fake call from the “CEO” ordering an urgent transfer, which is why payments are confirmed through a second channel. Under Article 50(4) of the AI Act, a company or public body that uses an AI system to create a deepfake must disclose that the content is artificial.

Demand forecasting is estimating how much customers will buy in a given period. The model looks at sales history, season, prices and promotions, like a planner with a spreadsheet, but weighs more factors at once. Purchasing and stock levels depend on the forecast, so the size of its error matters too.

A diffusion model creates an image or video by gradually removing noise until the scene from the prompt appears. It works like a sculptor revealing a shape from a block, step by step. Companies use it for graphics and mock-ups, but need rules on rights to the output and on labelling.

An embedding records the meaning of a text or image as a list of numbers. It works like an archivist who files letters about the same matter together, even when their titles differ. Search by meaning, rather than by keywords, is built on this.

An employee assistant is an agent that works for one person, who approves every result. A process agent is assigned to a process: it moves cases through each stage, while a person resolves the exceptions. It is the difference between help on request and a colleague with their own job description.

An enterprise AI assistant is the business edition of a chat with a model, e.g. ChatGPT, Copilot, Claude or Gemini. It works like a helper at the desk: it drafts and summarises when an employee asks. Newer editions include agent modes, so the line between an assistant and an AI agent is blurring.

The AI Act is EU Regulation 2024/1689 on artificial intelligence. It classifies AI systems by risk and places obligations on their providers and on the companies that use them. It entered into force on 1 August 2024 and applies in stages; in July 2026 Regulation 2026/1744 moved some of the deadlines.

An expert system is a program that makes decisions using rules written down by people, for example “if symptoms A and B, then diagnosis C”. It works like a thick binder of procedures that someone has to update by hand. It does well with stable rules but struggles with unusual cases and free-form text.

Explainable AI (XAI) is the ability to show why an AI system produced a given result. Think of an accountant: the total matters, and so do the documents it was calculated from. An explanation written by the model itself does not always show how it actually reached the result.

Fine-tuning is extra training of an existing model on examples from one field or company. It is like training an experienced employee in your company's style and procedures. It changes how the model answers; for up-to-date facts, RAG usually works better.

A foundation model is a large model trained on general data before anyone adapts it to a specific task. It is like a graduate with broad knowledge before on-the-job training. The EU AI Act uses a related term: general-purpose AI model.

The GDPR applies when an AI model or agent processes personal data, including when customer data is pasted into a chatbot. The company still needs a legal basis, limits the data it uses and informs people. In practice it asks: who is the processor, does the vendor train the model on our data, and where is that data stored.

Gemini is a family of language models and an AI assistant made by Google. Companies use it in Google Workspace, through the API and in Google Cloud. As of 26 September 2026, Google's documentation lists Gemini 3.8 Flash as the newest model in its fast Flash line.

Generative AI is a family of models that create new content: text, images, video, audio or code. It works like an assistant who drafts the first version of a letter, which an employee then checks and corrects. It differs from predictive AI, which does not create content but scores and forecasts.

A GPU is a graphics processor that performs thousands of simple calculations at the same time. That is why AI models are trained and run on it. For a company, the card's memory matters most: it decides how large a model fits on a single server.

Grounding means basing a model's answer on specified documents and citing the source of each piece of information. It works like a footnote in a report: the reader checks the source instead of taking it on trust. A citation doesn't prove the answer is right, but it makes checking faster.

Guardrails are safeguards on a model's input and output: filters, rules and checks that block forbidden instructions, personal data or off-topic answers. They work like a security gate: they do not replace permissions, they add another layer. On their own they will not stop a determined attacker.

A high-risk AI system is an AI Act category for uses that can seriously affect people, such as recruitment, credit scoring or critical infrastructure. These systems require, among other things, risk management, event logging and human oversight. After the 2026 amendment, obligations apply from 2 December 2027 for Annex III systems and from 2 August 2028 for AI in Annex I products.

The 1956 Dartmouth summer research project is regarded as the start of AI as a field of science. Waves of enthusiasm and disappointment followed: the first AI winter, expert systems and a second winter, the return of neural networks and finally language models. For a board, the lesson is that AI promises have run ahead of results more than once.

Human in the loop is a setup in which a person approves an AI decision before it is carried out. It works like a manager signing off a transfer that an accountant prepared. The AI Act (Article 14) requires human oversight of high-risk systems, not approval of every single decision.

Human on the loop is a setup in which the AI acts on its own while a person watches over it. The person reviews results, receives alerts and can stop the system at any moment. Think of a shift supervisor who does not sign off every parcel but sees the conveyor and has a stop button.

Human oversight is an AI Act requirement (Article 14) for high-risk systems. A person with the right knowledge must understand how the system works, spot errors and be able to disregard its output or stop it. It is more than clicking “accept”: the overseer needs time, authority and information to genuinely assess the decision.

Indirect prompt injection is an instruction hidden in content the agent reads while working: an email, a PDF file or a web page. It resembles an invoice with a note saying “pay to a different account” that the accountant mistakes for the boss's order. It is most dangerous when the agent has tools to send or write data.

Inference is the work of an already trained model: it reads a question and generates an answer. Training is like learning a trade, and inference is the everyday work on cases. A company pays for inference through token fees, a subscription or the running of its own servers.

ISO/IEC 42001 is an international standard from December 2023 that describes an AI management system in an organisation: policies, risk assessment, impact assessment and supplier oversight. It works much like ISO 27001 for information security, so a company can get certified. The certificate confirms the management process, not the quality of a particular model.

The Jevons paradox is the observation that cheaper use of a resource can increase its total consumption. William Stanley Jevons described it in 1865, using coal and more efficient steam engines as the example. With AI, a cheaper token can raise the bill because the company uses AI in more and more processes.

A kill switch is a way to stop an agent or cut its access to tools in one move when it does something unwanted. It works like blocking a stolen payment card: it does not undo the damage, but it stops it from growing. The company decides in advance who may use it and how fast.

The knowledge cutoff is the point up to which a model's training data reaches. It is like an employee back from long leave: they only know what's new from the documents you hand them. Vendors publish this date; for Claude Opus 5.5, Anthropic lists June 2026.

An LLM, or large language model, is an AI model trained on vast amounts of text. It works like a well-read assistant: it predicts the next words based on what it has learned. It does not check facts on its own, so at work it needs sources and human review.

Liability for AI errors answers the question of who is responsible for harm caused by an AI decision. A model cannot be held liable, so, as with a letter drafted by an intern, the company that used it answers for the result. From 9 December 2026, under the new EU Product Liability Directive, the producer of defective software, including AI, is liable for harm to natural persons.

Temperature is a setting that controls how randomly a model picks the next words. Low values give more predictable answers, like a form; high values give freer ones, like a brainstorm. Some newer models don't let you change it, e.g. Claude models from Opus 4.7 onwards.

A local AI model runs on the company's own servers rather than in a provider's service. Data stays in the company, but hardware, updates and security are your responsibility. It is like running your own server room instead of renting cloud capacity.

Machine learning (ML) is a method in which a program derives rules from examples instead of receiving them from a programmer. It resembles a new employee who, after hundreds of invoices, learns which department each one goes to. The quality of the result depends on the quality and number of those examples.

An MCP server is a program that exposes the data or actions of one system, such as a CRM or a mailbox, to AI agents using the MCP standard. It defines which tools the agent sees and what it can do. Before connecting one, IT checks who wrote it and what permissions it receives.

Microsoft Copilot is Microsoft's AI assistant in Word, Excel, Outlook and Teams. With the paid licence it reads email and files like an assistant holding keys only to that employee's cabinets. Since August 2026 this has been the name of what used to be Microsoft 365 Copilot.

A model API is the technical entry point through which a program, not a person, gives a model its tasks. It works like a service counter: the company system hands in a case and the answer comes back to that system. API use is usually billed per token rather than per user.

Model confidence is a score the system assigns to a result, for example 0.92 when matching an invoice. The decision threshold is the cut-off: above it the system acts alone, below it the case goes to a person. The process owner sets the threshold and tests it on past cases, because the score can be overconfident.

MCP (Model Context Protocol) is an open standard for connecting AI agents to systems, data and tools. Instead of a separate connection for every app–system pair, each system gets one plug that follows the standard. Anthropic published MCP in November 2024, and since December 2025 it has been developed by the Agentic AI Foundation under the Linux Foundation.

Parameters are the numbers inside a model that store what it learned in training. They work like an employee's experience: you can't see it in their file, yet it shapes every answer. More parameters usually mean greater capability, but also higher cost and slower responses.

Model training is the stage in which a model adjusts its parameters on many examples until its results are accurate enough. It resembles onboarding a new employee: costly, but done rarely, after which the learned knowledge is used for a long time. Everyday use of a finished model is inference; companies usually pay for that, not for training.

A multi-agent system is several AI agents with different roles, tools and permissions, handling one case together. A lead agent or a fixed workflow coordinates them. It is like a department with divided duties: one person takes in the case, another checks it, a third approves it.

A multimodal model accepts images, audio, video or PDF files as well as text. It reads a scanned invoice, a damage photo or a call recording the way an employee goes through attachments. One tool then handles documents that used to need separate OCR software.

Natural language processing (NLP) is the field of AI that teaches computers to read, understand and produce text and speech. In a company this means, for example, sorting emails, extracting data from contracts and summarising meeting notes. Today's large language models are the latest stage of NLP.

A neural network is a model built from layers of simple computing units that pass numbers to each other and strengthen or weaken the signal. It resembles a document moving through departments: each adds its assessment and the last one issues the decision. The strengths of the connections are parameters the network sets during training.

NIS2 is Directive (EU) 2022/2555 on the cybersecurity of companies and public bodies in key sectors. In Poland it is implemented by the amended National Cybersecurity System Act, in force since 3 April 2026. Covered organisations manage risk and report incidents, and the head of the organisation, such as the management board, is responsible for this.

The NIST AI RMF is a voluntary framework for managing AI risk, published by the US institute NIST on 26 January 2023. It organises the work into four functions: Govern, Map, Measure and Manage. It is a voluntary framework for your own process, not a certificate, a regulation or a checklist.

Non-determinism means a model may answer the same question slightly differently each time. It is like an employee who describes the same case in different words the second time. Even temperature 0 doesn't guarantee identical answers, so a process should check the substance of the result, not its wording.

An open-weight model is one whose weights, the learned parameters, the maker publishes for download. A company can run it on its own infrastructure and fine-tune it on its own data. A licence sets the terms of use, while the training data and training code usually stay private.

An organizational knowledge base is an organized collection of procedures, documents and answers used by both people and AI agents. The agent searches it for information, so the quality of the base limits the quality of its work. An outdated procedure in the base leads to mistakes just like an outdated binder.

Overfitting happens when a model memorises its training data instead of learning general rules. It resembles an employee who knows last year's cases by heart but gets lost at the first new one. The model scores well on familiar data and poorly on new data, which is why it is checked on test data.

A per-seat licence is a fixed fee per person per month, like a phone subscription. Usage-based pricing grows with every request, like paying for each photocopied page. Some business plans combine both, so the budget has to include both parts.

Per-token pricing is the rate a model provider charges for text, quoted per million tokens. Input, what the model reads, costs less than output, what it writes. Reasoning models also bill their reasoning tokens, usually at the output rate.

A perceptron is the simplest neural network: it adds up weighted input signals and answers “yes” or “no”. Frank Rosenblatt presented it in a 1957 report and a 1958 paper. It works like a clerk with a single scoring table: points are added for features of an application and checked against a threshold.

Permission-aware retrieval means the AI searches only the documents the person asking is allowed to open. RAG is the retrieval of document fragments that the model receives before it answers. Without this control, a salesperson asks about bonuses and gets an excerpt from the payroll.

Predictive (analytical) AI is a family of models that estimate what will happen based on historical data. It forecasts demand, scores credit risk and flags customers likely to leave. It answers with a number or a label, not with new text as generative AI does.

A predictive model is a machine learning model that estimates a future outcome or risk from historical data. It answers questions such as how much we will sell in March or which customer will pay late. It gives an estimate or a probability, not a certainty, so the company decides at what threshold to act.

The principle of least privilege is the rule that everyone gets only the access needed for the task. It applies to people, service accounts and AI agents; NIST describes it in control AC-6. An agent that reads invoices needs no right to make payments, just as an intern needs no key to the safe.

A process owner is the person accountable for the outcome of an entire process, such as complaint handling. They decide what an AI agent may do on its own and what needs human approval. Without one, nobody answers for the agent's mistakes or keeps the rules up to date.

A prompt is the instruction a person or program sends to a model: the question, task, data and expected format. It works like a brief for a new colleague: the clearer the goal and context, the fewer rounds of corrections. The same prompt can give a different result in another model.

Prompt caching stores the repeated opening of a request, for example terms and conditions attached to every case. The provider keeps it for anything from a few minutes to a day, and reading it again is cheaper and faster. It works like a letter template: you do not retype the fixed part, you only add the new details.

Prompt engineering is the skill of writing instructions so a model gives a useful result. It is like writing a good purchase specification: goal, example, constraints and format. For agents, the broader practice of context engineering, choosing all the material the model sees, is becoming more important.

Prompt injection is an attack in which someone feeds an AI model an instruction that changes its behaviour against the owner's intent, such as “ignore previous instructions”. The model reads instructions and data through the same channel, so it cannot always tell them apart. The OWASP Top 10 for LLM Applications 2026, published in August 2026, still ranks it as risk number one.

The AI Act separates the provider, who develops an AI system or has it developed under its own name, from the deployer, who uses it. A company buying a ready-made tool is usually a deployer, like the buyer of a forklift. It becomes a provider, however, if it commissions a system under its own name, or puts its brand on or substantially modifies someone else's high-risk system.

A reasoning model is a language model that works through the steps of a solution before answering. It is like an analyst who does the maths on scrap paper before writing the conclusion. It analyses and plans better but uses more tokens; some models, e.g. Claude Opus 5.5, always reason.

Reinforcement learning (RL) is learning by trial and error, with a reward for a good result. It resembles a bonus scheme: the employee finds their own way to earn the bonus, sometimes by cutting corners. The same method is used to fine-tune language models, partly based on human ratings.

RAG (retrieval-augmented generation) is a method where a model receives matching passages from company documents before it answers. It works like an employee who checks the rulebook before replying to a customer. The model doesn't learn from the documents, and the answer can point to its source.

Robotic process automation (RPA) is software that repeats human clicks and keystrokes in applications according to a fixed script. It works well with stable forms but fails when a screen changes or a document looks unusual. An AI agent copes with variations, but needs controlled permissions and approvals.

RBAC is access control in which permissions are granted to a role rather than to an individual. An employee or an AI agent receives a role, such as "accountant", and with it a set of permissions. Changing the role changes access at once, which makes revoking rights and preparing for audits easier.

A SIEM is a system that collects logs from servers, applications and accounts in one place, correlates them and raises an alert on a suspicious pattern. Microsoft Sentinel is one example. When a company uses AI agents, their action logs can feed the SIEM next to human logs, so the SOC sees the whole picture.

A SOC (Security Operations Center) is a team that monitors a company's security, usually around the clock, detects attacks and responds to incidents. It works like building security with CCTV. For the SOC, an AI agent with system access is one more account whose actions should show up in the records.

Segregation of duties is the rule that no single person carries out an entire risky operation alone. Whoever prepares a payment does not approve it, and an AI agent does not approve its own work either. It protects against mistakes and fraud, because two parties see every important operation.

Shadow AI is employees' use of AI tools the company does not know about or has not approved, such as a personal chatbot for translating contracts. It usually comes from the lack of a better option, not from bad intent. The company loses control over where its data goes.

Single sign-on (SSO) is one sign-in that opens many company applications. Identity is confirmed by an identity provider, a central directory of employee accounts. When someone leaves, IT disables one account and access disappears from every connected application.

A small language model (SLM) usually has from a few hundred million to around 15 billion parameters. It runs faster and cheaper, often on an ordinary server or laptop. It works well for narrow, repetitive tasks such as sorting incoming mail, and less well for complex reasoning.

A subagent is an agent to which the main agent delegates a separate part of the task, such as searching documents. It works with its own instructions and permissions and returns only the result. It is like a case manager who sends a query to the archive and gets back a finished summary.

In supervised learning, a model receives examples with the correct answer, such as emails labelled “complaint” or “order”. In unsupervised learning, it receives unlabelled data and finds groups on its own, such as similar customers. The first needs work to label data; the second produces findings that a person must interpret.

A supervisor agent is an AI agent that assigns work to other agents and checks their results. It usually does not act in company systems itself; it escalates to people whatever it cannot resolve. It works like a coordinator who hands out assignments, collects finished work and checks its quality.

A system prompt is a standing instruction for the model: its role, rules, tone and limits. It works like a job description an employee reads before taking the first case. It is not a security control: a clever instruction can get around it, so permissions are set separately.

A token is a piece of text a model works with: a word, part of a word or a single character. It is like paying a translator per character rather than per page. Tokens set both the price of using a model and the limit of text it can take in at once.

Tokenization is cutting text into tokens before a model processes it. Each vendor cuts differently, and sometimes changes the method in a new model version. So the same document costs different amounts, and a token count only makes sense alongside the model name.

Agent tools are named functions through which an agent reads data and acts in systems, such as "fetch invoice". The model requests a tool by giving its name and parameters (tool calling), and the program around the model executes it. It is like a request form: the employee submits it, and the department carries it out after checking.

TCO, or total cost of ownership, is everything a tool costs over the whole time you use it. Like a company car: beyond the purchase price, the company pays for fuel, servicing and insurance. For AI that means licences, tokens, integrations, maintenance, training and oversight time.

Training data are the examples a model learns from. Test data are separate examples used to check whether it learned well. It works like an exam: the questions must not repeat the classroom exercises, or the score will be inflated.

The transformer is the neural network architecture behind most of today's language models. It works like a reader who takes in the whole page at once and sees what connects to what. Google researchers described it in 2017 in the paper “Attention Is All You Need”.

The Turing test is a trial Alan Turing proposed in 1950: a judge holds a written conversation with a person and with a machine. If the judge cannot tell which is which, the machine passes. Today's chatbots converse very fluently, so for a company fluent conversation alone does not prove that a system understands the process.

A vector database stores embeddings and quickly finds those closest in meaning to a question. A regular database finds an invoice by its number; a vector database finds contracts about late-delivery penalties. Regular databases such as SQL Server 2025 and PostgreSQL with pgvector now offer vector search too.

WebMCP is a proposed browser API through which a website exposes named tools to AI agents. Instead of clicking through the page, the agent calls something like "place order" with parameters. A W3C community group is developing the specification, and since June 2026 websites can trial it in Chrome 149.

A workflow is a predefined path of steps, with AI used at selected points. An agent decides for itself which steps to take and in what order. Anthropic recommends starting with the simplest solution and increasing complexity only when needed.

A world model is AI that predicts the effects of actions in an environment, such as a robot arm moving. JEPA is Yann LeCun's approach: the model predicts a simplified representation of the scene rather than every pixel. In 2026 world models are used mainly in robotics, autonomous vehicles and simulation.

Zero trust is a security principle under which no account, device or program is trusted just because it sits inside the company network. Every access request is checked separately, like a badge scanned at every door rather than only at the entrance. The principle covers AI agents and their identities too.