AI Glossary

AI systems that can autonomously plan and execute multi-step tasks by calling tools, browsing the web, writing and running code, or interacting with external APIs — without a human approving each action. Agentic AI is the shift from "AI as autocomplete" to "AI as a colleague that gets things done." Frameworks like CrewAI and LangChain are common building blocks.

An engineer or team member who leads AI adoption within their organisation — not necessarily the most senior, but curious, comfortable with ambiguity, and willing to experiment. AI champions prototype tools, bring findings back to the team, and create the social proof that shifts culture. A structured AI adoption programme needs 2–3 champions with protected time before the whole team engages.

The policies, processes, and controls a company puts in place to ensure AI systems behave safely, fairly, and in compliance with regulations. Good AI governance covers model auditing, bias detection, data lineage, access controls, and incident response. Increasingly required by enterprise procurement teams and regulators like the EU AI Act.

A company's capacity to successfully adopt AI — measured across data quality, infrastructure maturity, team skills, executive sponsorship, and process clarity. Low AI readiness is the #1 reason enterprise AI projects stall. A structured assessment (like the CEO Scorecard) is the fastest way to identify gaps before committing budget.

A model confidence score that accurately reflects the true likelihood of correctness — when the model says it is 80% confident, it should be right roughly 80% of the time. Calibration is critical for automated decision-making: you cannot threshold on confidence if confidence is systematically over- or under-stated. System One models like Jev are trained specifically for calibrated outputs; chat LLMs often require post-hoc temperature tuning.

The maximum amount of text (measured in tokens) a language model can process in a single request — including the prompt, conversation history, retrieved documents, and the model's response. A larger context window allows for more complex reasoning but increases cost and latency. GPT-4o supports 128K tokens; Claude 3.5 Sonnet supports 200K.

An open-source Python framework for orchestrating multiple AI agents working together as a "crew." Each agent is assigned a role, goal, and set of tools — and they collaborate to complete complex tasks. CrewAI is one of the most popular agentic frameworks for production multi-agent systems.

A systematic assessment of data quality and completeness run before committing to an AI project timeline. In practice, 60–70% of AI project time is data work — cleaning, labelling, formatting, validating — so discovering gaps early prevents schedule slippage. A data audit surfaces missing values, inconsistent formats, access restrictions, and data-owner dependencies before they become blockers.

A named AI teammate that keeps a job description, memory, preferences, and skills across sessions — not a one-shot chatbot. Digital employees (also called "Bots" in Grok Bot or "Wakers" in QoderWake) share persistent compute environments, hand off work to each other, and stop for human approval on consequential actions. The key shift: scope an outcome and review the result, instead of prompting forever.

Numerical representations of text, images, or audio as high-dimensional vectors, where semantically similar content clusters together in vector space. Embeddings power semantic search, recommendation systems, and RAG pipelines. They are generated by embedding models (e.g., OpenAI text-embedding-3-large or Cohere embed-v3).

The process of continuing to train a pre-trained language model on a smaller, domain-specific dataset to improve its performance on a narrow task or adopt a specific tone and style. Fine-tuning is expensive and requires high-quality labelled data — for most enterprise use cases, RAG is a better first step.

A large AI model trained on broad data at massive scale that can be adapted to a wide range of downstream tasks. Examples include GPT-4, Claude 3.5, Gemini 1.5, and Llama 3. Building on a foundational model via prompting, RAG, or fine-tuning is almost always faster and cheaper than training a model from scratch.

A capability that allows language models to invoke external tools and APIs in a structured, type-safe way. Instead of generating free-form text, the model outputs a JSON object specifying which function to call and with what arguments. This is the mechanism that makes agentic AI reliable and auditable.

When a language model generates factually incorrect information with high confidence. Hallucinations happen because LLMs predict plausible-sounding tokens, not ground truth. Mitigation strategies include RAG (grounding answers in retrieved documents), constitutional prompting, and output validation layers.

A popular open-source framework for building LLM-powered applications, providing abstractions for chains, agents, memory, and tool use. LangChain accelerates prototyping but can introduce complexity in production. For simpler use cases, calling the model API directly is often more maintainable.

An open-source data framework for connecting LLMs to external data sources — structured databases, PDFs, APIs, and more. LlamaIndex specialises in indexing and retrieval, making it a natural fit for RAG pipelines. It integrates with most major embedding models and vector databases.

A neural network trained on vast amounts of text data that can understand and generate human language. LLMs underpin virtually all modern AI products — from chatbots to code assistants to document summarisers. The most capable LLMs in 2025 are GPT-4o, Claude 3.7 Sonnet, and Gemini 1.5 Pro.

AI systems that do not wait to be asked — they hold enough durable context (goals, preferences, prior decisions) to notice when something matters, do work in real tools, and interrupt you only when the cost of silence exceeds the cost of a ping. The shift from reactive chat to proactive agents requires watchers (email, calendar, CRM), a persistent computer, and a stop line for sensitive actions. Examples include Grok Bot for work and Meta Muse for consumer goals.

The discipline of designing and optimising the instructions given to a language model to reliably produce desired outputs. Effective prompts include clear context, role definition, output format instructions, and few-shot examples. Prompt engineering is often the highest-ROI activity in the first 90 days of an AI project.

A technique that connects a language model to an external knowledge base before generating a response. Instead of relying solely on training data, the model retrieves relevant documents at query time — making answers accurate, up-to-date, and auditable. RAG is the recommended starting point for most enterprise AI use cases before considering fine-tuning.

The explicit boundary that determines when an AI agent must pause for human approval before proceeding — typically any action that sends, spends, deletes, or touches production. Without a stop line, you have a liability with a chat UI, not a teammate. The rule from the field: start every agent on read-and-prepare work, then promote to consequential actions only after you have seen failure modes and set the approval boundary in writing.

A class of AI model optimised for fast, structured decisions rather than free-form text generation — borrowing Kahneman's System 1 framing. System One models (e.g., Jev, Laya) give up prose in exchange for type-safe answers with calibrated probabilities in 70–500ms. Use them for routing, classification, scoring, and guardrails inside a latency budget; keep an LLM for the human-facing reply.

The basic unit of text that language models process — roughly 4 characters or 0.75 words in English. LLM pricing is based on input and output token counts. Understanding tokens helps estimate costs: a 10-page PDF is approximately 5,000 tokens; processing 1 million such PDFs at GPT-4o pricing costs around $2,500.

A database optimised for storing and querying high-dimensional numerical embeddings — the mathematical representations of text, images, or audio. Vector databases power semantic search and are the backbone of RAG systems. Common choices include pgvector (PostgreSQL extension), Pinecone, Weaviate, and Qdrant.

A TypeScript toolkit for building AI-powered web applications, with unified APIs for streaming text, structured outputs, and tool use across OpenAI, Anthropic, Google, and other providers. The Vercel AI SDK dramatically reduces the boilerplate of integrating LLMs into Next.js or React apps.