Learn · 50 terms

Dictionary

This dictionary explains, in plain words, fifty terms that leaders meet when they talk about AI. You do not need to be an expert; you need to understand the words the same way as the people you work with.

The fundamentals

The basics for understanding the AI landscape.

AI, artificial intelligence: The field of making computer systems do tasks that normally need human intelligence.; AI is an umbrella term, not one technology. Most of what a business meets today is machine learning and, within it, generative AI and large language models.; Picture it
AI is to machine learning what "transport" is to "the car".; See alsoMachine learning Generative AI LLM
Machine learning: The engine of most modern AI: a model that is trained on large amounts of data to recognise patterns itself, instead of being programmed with explicit rules.; A subcategory of AI. The model is not told the rules; it finds them in the examples it is trained on, which is why the quality of the data decides the quality of the model.; Picture it
If AI is transport, machine learning is the combustion engine.
Generative AI: AI that creates new content: text, images, music, code.; Generative AI produces something new, as opposed to analytical AI, which mainly analyses and finds patterns in existing data. Both have their place in a business.
LLM, large language model: A machine learning model that is trained on enormous amounts of text and built to understand, generate, summarise, and translate human language.; At its core an LLM is an extremely advanced predictor of the next word. Everything else it appears to do comes from being very good at that.; Picture it
A specific, top-of-the-line engine, a V8 turbo, if machine learning is the engine type.; See alsoThe transformer architecture Emergent abilities
Chat, chatbot: The window a person types into. It is the interface, not the engine.; Tell the chat apart from the model behind it. A chat answers the question it is given and waits for the next one, so the person drives every step; a knowledge worker takes a goal and works toward it.; Picture it
The chatbot is the dashboard, wheel, and pedals. The model is the engine.; See alsoAI agent Agentic knowledge worker
Hallucination: When a language model produces an answer that is wrong, invented, or meaningless, and presents it with confidence.; It happens because the model is a pattern generator without an awareness of truth. Understanding this is the start of managing the risk; grounding reduces it, because the model answers only from sources it has looked up, and shows them.; See alsoGrounding RAG
Structured data: Information in fixed fields and columns: databases, spreadsheets, name, date, amount.; It is machine-friendly: easy to filter, join, and calculate with, and good for control and compliance. Think of a spreadsheet of orders that can be counted straight away.
Unstructured data: Everything without a fixed schema: emails, PDF reports, chat logs, images, audio, video.; The knowledge sits in sentences, tables, and figures, so it has to be read, searched, and extracted before it can be counted. Here AI makes the difference: it finds and summarises the knowledge, pulls out fields such as names, amounts and deadlines, and turns them into usable structure.; See alsoStructured output Data pipelines and the data warehouse

The toolbox: from theory to use

The terms you need to turn AI potential into real solutions.

AI agent: A system that does not only answer a question but acts to reach a goal: it breaks a task into steps and uses tools to carry them out.; On this site, an agent is called a knowledge worker when it has a company's business language, records, tools, access, rules, and a logbook, and works with an operator.; Picture it
A self-driving car, as against a car with a very good map.; See alsoAgentic knowledge worker Tools Agent harness
Agentic, not assistive: An assistant answers the prompt it is given. An agentic system works toward a goal: it plans, does the work, checks it, and returns for sign-off.; When an assistant makes your slides or builds a budget sheet, the work is still yours: you prompt every step. An agentic knowledge worker takes the goal once and comes back when the job is done, or with the one question it could not answer alone. It can also assist. The data, the access, and the guardrails it works with set how far it may go.; See alsoAI agent Agentic knowledge worker The three sign-off zones
RAG, retrieval-augmented generation: A method where the agent looks things up in your own sources right before it answers, instead of guessing: documents and structured data alike.; The model itself does not change; it consults the sources on the fly. The lookup can be exact (the item number "A100" only), fuzzy ("A-100" counts as "A100"), or semantic, so it matches on meaning: "largest customer in Aarhus last quarter" finds the right row across fields.; Picture it
An employee with the handbook and the spreadsheet open, rather than one answering from memory.; SourcesLewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (2020); See alsoGrounding Hallucination
Grounding: The AI answers only from what it just looked up in your sources, and shows where it found it.; File and page, or table, row, and column: the source comes with the answer, so you can check it. It curbs hallucination because the model may not state anything it has no source for, and when there is no source it says "I don't know" or points to where the answer may be found. For example: "3 months, Staff policy, page 12, section 3.2."; See alsoRAG The logbook
Fine-tuning: Retraining an existing language model on your own data, which permanently changes what it knows.; It is expensive and complex, and rarely the place to start. For most companies, grounding the model in good sources works better than changing the model.; See alsoRAG
Context window: The model's short-term memory: the largest amount it can take into account at once, measured in tokens.; The window sets the complexity of the task an agent can hold in its head at one time. Larger windows allow longer documents and longer jobs; good instructions and lookups keep the window filled with what matters.; See alsoTokens and prices Agent harness
UI versus API: The UI is what a person sees and clicks; the API is the contract that lets programs talk to each other.; Work through the interface is personal productivity; work through the API is automation that scales. A knowledge worker that posts entries or answers questions for a whole team works through APIs.
Structured output: Asking the model to answer in a fixed, machine-readable format, such as JSON, with named fields that are checked against a schema.; Because the same fields come back every time, a system can read the answer, and code can check it before it goes on.; SourcesOpenAI: Structured outputs; See alsoFunction calling Unstructured data
Tools: The functions an agent may call to get a task done: look up, calculate, draft, write to a system.; Without tools, a language model can only answer from what it learned in training. With tools, it can fetch current data, calculate with it and work in your systems. Each tool is one function with one defined use.; See alsoFunction calling Access and access control MCP server
Function calling: Instead of writing free text, the AI fills in a small form with the fields a tool needs; the system runs the lookup or the action and returns the result in the same fixed shape.; Fixed fields every time means less guessing and more repeatable, safer actions. A knowledge worker uses it to ask the accounting system for a supplier balance, or to propose an entry.; SourcesAnthropic: Tool use; See alsoTools Structured output
Prompts: system and user: The user prompt is what a person types in the chat; the system prompt is the standing, often hidden instruction that sets the model's role, rules, and purpose for the whole conversation.; The system prompt is one of the main tools of control. It holds the rules for what to do. The guardrails that stop the model where it must not go are enforced outside it.; See alsoRules and guardrails Skills
Prompting frameworks: A shared structure for how an agent should behave: role, goal, context, tools, output format.; The framework makes answers consistent and repeatable, so a domain expert can build a custom agent by filling it in, or write a simple brief that a platform team turns into one. Answers become consistent and easier to check. In Gathena's work the frameworks are skills, playbooks, and runbooks.; See alsoSkills
Tokens and prices: The unit a language model reads text in, often part of a word; commercial models are priced per token in and out.; Prompt and answer both count. A long document in the context window costs tokens every time it is read, which is one reason a well-built knowledge worker retrieves what matters instead of pasting everything.; See alsoContext window Inference
Open source models: Models whose recipe, and often the trained model itself, is made freely available.; They give more control over where data goes, and demand technical expertise to run. Be exact about which layer is open and under which licence before you call a tech stack "open source".
Cloud versus on-premise: Cloud: you use models running on a provider's servers, flexibly and without buying hardware. On-premise: you run them on your own servers, with full control.; On-premise can be a requirement for extremely sensitive data. For most companies the practical answer is their own cloud project, where the data stays theirs and the company controls access.; See alsoAccess and access control Data pipelines and the data warehouse
Inference: The moment a trained model applies what it learned to new input and produces an answer. Every question to an LLM is inference.; Its cost and speed decide what is practical: a knowledge worker that runs every night on thousands of records has a different budget from a chat that answers one person.; See alsoTokens and prices
Chain of thought: Getting the model to show its working: how it reads, interprets, and acts, so a person can follow how it reached its output.; It is useful for checking a knowledge worker's reasoning, and it is part of what is written to the logbook with every proposal.; SourcesWei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (2022); See alsoThe logbook

Strategic terms

Here the focus moves from technology to people and organisation.

Domain knowledge: A person's deep, experience-based understanding of a field, an industry, or a process.; With AI, domain knowledge matters more, not less. The model can reason; your bookkeeper knows what a correct entry looks like in your company. Everything around a knowledge worker is there for one purpose: to bring that knowledge into the work.; See alsoThe company dictionary Ask the expert
Intention: The purpose or desired result behind a request. When you state it clearly, it gives the AI a direction.; A question without an intention gets an answer to the question. A question with an intention gets an answer to what you were trying to do.; See alsoContext
Context: All the surrounding information that gives an intention its meaning and removes ambiguity: data from a document, the industry's terms, the role the AI should take.; Context is what a new colleague is given on their first day and what a model never has on its own. Most of what a knowledge worker needs is context that you write down once.; See alsoWhat a knowledge worker needs Context window

The breakthroughs behind it

The technical steps that made the current AI wave possible.

The transformer architecture: A model structure from 2017 that is exceptionally good at understanding context across long stretches of text. Almost every modern LLM is built on it.; It lets a model keep a long document in view and connect the end of it to the beginning.; SourcesVaswani et al., Attention Is All You Need (2017); See alsoLLM
Self-supervised learning: The model learns from a simple task, "guess the next word", and checks and corrects itself against the text alone, so no person has to label every sentence.; This made training on the whole written internet possible.
Emergent abilities: Abilities a model develops as an unexpected side effect of mastering its core task at enormous scale.; To get good at guessing the next word, the model had to learn grammar, logical reasoning, and a working picture of the world. Nobody programmed those in.; SourcesWei et al., Emergent Abilities of Large Language Models (2022); See alsoLLM
RLHF, reinforcement learning from human feedback: A way of polishing a trained model: people judge which of two answers is better, and the model is rewarded for answers that resemble the preferred ones.; It makes a raw language model more helpful and safer to talk to.; SourcesOuyang et al., Training language models to follow instructions with human feedback (2022)

Gathena's own words

The words this site uses for what a knowledge worker needs, and what each one means.

Agentic knowledge worker: An AI agent that is set up to do a job in your company: a model with your business language, its own records, tools, access, rules and limits, a logbook, and an operator who signs off.; It takes a goal, breaks it into steps, uses tools, checks its own work, and comes back when the job is done, or with the one question it could not answer alone. People are not part of it; they make the decisions.; Picture it
A brain on a computer screen: the brain is the model, the computer is everything around it.; See alsoWhat a knowledge worker needs The operator AI agent
What a knowledge worker needs: Everything a knowledge worker needs to do real work: the same things a new colleague gets on their first day.; It needs your business language, its own records, tools, access, rules and limits, a logbook, and an operator who signs off. What matters is having all of them.; See alsoAgent harness The company dictionary Access and access control Rules and guardrails The logbook
Agent harness: The technical name for the computer around the brain: the software that gives a model its instructions, tools, access, memory and working loop.; Claude Code, Codex, GitHub Copilot and the agents in Microsoft Copilot are harnesses. Gathena puts your business language, records, tools, access, and rules into a harness so it can do a job in your company.; See alsoWhat a knowledge worker needs Tools
The company dictionary, ontology: A shared record of what you call things, what each thing means in your company, and how the things connect. You write it down once.; A "customer" is someone with a signed contract; an "open invoice" is one that is unpaid past its due date; an invoice belongs to a supplier, a supplier to a project. The technical word is ontology: a dictionary and a map in one. Every knowledge worker you add uses the same one.; See alsoThe data foundation Domain knowledge
The data foundation: One data warehouse in your own cloud that brings together the data from your systems. You can trust it because every figure can be traced to its source.; Knowledge workers take their data from it. Beside it is your company dictionary, which builds on Gathena's template: a general dictionary and map of how work is done. With the template, each new job is easier to set up than the last, and a job that runs again is easier still. Both belong to you and stay in your own cloud.; See alsoData pipelines and the data warehouse The company dictionary
Access and access control: What a knowledge worker may read and do. The company controls the access, gives a knowledge worker only what its job needs, and can take it back at any time.; The finance knowledge worker can read the accounting system and propose entries; it cannot send email. Every use of that access is written to the logbook.; See alsoRules and guardrails The logbook
Rules and guardrails: Rules tell a knowledge worker what to do; guardrails stop it where it must not go, whatever anyone tells it.; Rules say: follow the bookkeeping policy, use these terms, write in this tone. Guardrails say: no payments, no deleting records, nothing goes outside without sign-off. A rule can be argued with. A guardrail cannot.; See alsoPrompts: system and user The three sign-off zones
The three sign-off zones: The three levels of operator involvement: acts alone, asks first, the operator signs off.; The operator is always in the loop when money leaves the company, when something goes to an outside party, when a commitment is made, or when a rule is unclear. A step moves from "asks first" to "acts alone" only after it has proved reliable, and the way back stays open.; See alsoThe operator Earned automation
The operator, human in the loop: The person in charge of a knowledge worker, who gives the instructions, reviews the work, makes the judgement calls, and signs off.; An AI is a machine and cannot be held accountable. People can: the operator answers for the work, on the company's behalf. The word says who is in charge.; See alsoThe three sign-off zones Accountability and AI literacy
The logbook: The record of sources, actions, and decisions: which access was used, what was proposed, who decided, on which sources.; Every answer shows its source. Every handover between knowledge workers is written here, so a draft stays a draft until the right person has signed it off.; See alsoGrounding Chain of thought
Earned automation: The rule that a step moves to "acts alone" only when it has earned it: stable inputs, a known method, exceptions that are understood, and someone who owns the result.; Corrections improve the dictionary, the rules, and the methods, not the model. Learning builds up; autonomy is granted step by step and never taken for granted.; See alsoThe three sign-off zones Evals
Accountability and AI literacy: Accountability: you stay responsible for work AI does, not the AI or its vendor. AI literacy: your staff understand the AI they use, which the EU AI Act asks businesses to support.; Accountability rests on access the company controls, guardrails that hold, people who understand the system, and a logbook of what was done. In practice, people take that responsibility by reviewing the work; the law rarely prescribes a click for every output.; SourcesRegulation (EU) 2024/1689, the AI Act, Article 4Regulation (EU) 2016/679, the GDPR, Article 22Datatilsynet; See alsoThe operator
Data pipelines and the data warehouse: Pipelines that run on a schedule, read the systems you already use, and load their data into one data warehouse in your own cloud, where every figure can be traced to its source.; The pipelines keep the raw files as evidence and report when a source changes. Knowledge workers take their data from the warehouse, so an answer uses the same figures as the month-end report. AI does much of the structuring and cleaning.; See alsoThe data foundation MCP server Builder agent
MCP server: A server that an AI tool uses to reach your data and systems; each person can sign in with their own login, so they see only what they may see. MCP, the Model Context Protocol, is the shared standard for it.; With an MCP server in your cloud, an employee in Claude Team can ask "why did costs rise?" and get the answer with the records behind it. The employee sees only what they have access to.; SourcesModel Context Protocol specification; See alsoTools Access and access control Data pipelines and the data warehouse
Builder agent: The agent that builds and maintains the data pipelines, their tests, and the MCP servers. It works inside the project's own rules and asks before every change.; It works in Claude Code with five things in place: a written scope before any pipeline, a data contract for each table, guardrails and evals, decision records, and scripts that show a plan and change nothing until the operator tells them to apply it. The operator says yes; production needs a second, explicit yes.; SourcesAnthropic: Claude Code; See alsoData pipelines and the data warehouse Evals
Evals: The checks that score a knowledge worker's work against known answers. They run before anyone trusts the work, and again whenever something changes.; Evals show when a step has earned the move from "asks first" to "acts alone", and they show that a change of model or rules is safe before it goes live.; See alsoEarned automation Builder agent
Skills, playbooks, and runbooks: Written, versioned instructions that a knowledge worker follows: a skill for one capability, a playbook for how a kind of work is approached, and a runbook for the exact steps of a job.; They are shared between knowledge workers, so a better way to check an invoice is available to all of them the next day. Records are not shared; they stay where they belong.; See alsoPrompting frameworks The data foundation
Ask the expert: A specialist puts what they know, with the company's own context, into a knowledge worker that colleagues can ask. Experts still check the answers; colleagues come to them with better questions.; Knowledge is no longer tied to one person, experts spend their time on the hard cases, and a rule change is written down once, for everyone.; See alsoDomain knowledge Agentic knowledge worker