Glossary
AI glossary for business
Plain-language definitions of 45 AI terms that come up in automation, AI agent and document AI projects, and why each one matters to a business.
Agentic AI
Agentic AI describes systems in which a language model plans steps, calls tools and acts on results to reach a goal, rather than answering a single prompt. The model works in a loop: decide, act, observe, repeat, within limits set by its designers. For a business, agentic AI can take on multi-step work such as triaging requests or reconciling records, provided it has clear permissions, approvals and logging.
AI agent
An AI agent is a software program that uses a language model to decide which actions to take, such as searching data, calling an API or drafting a reply, in pursuit of a defined task. Agents differ from chatbots because they can act in other systems, not just produce text. Well-scoped agents can handle routine operational work while people review exceptions and higher-risk decisions.
AI governance
AI governance is the set of policies, roles, controls and records an organization uses to decide how AI systems are chosen, built, monitored and retired. It covers data handling, acceptable use, risk review, human oversight and accountability for outcomes. Practical governance lets teams adopt AI with confidence, because the rules for what data can be used and who signs off are written down before problems occur.
Business process automation (BPA)
Business process automation is the use of software to carry out repeatable business workflows, such as invoice processing, onboarding or order handling, with less manual effort. Modern BPA combines rules, integrations and AI models that can read documents and handle variation. It reduces rekeying, delays and errors, and frees staff for work that needs judgment.
Chunking
Chunking is the process of splitting documents into smaller passages before they are embedded and indexed for retrieval. Chunks may follow fixed sizes, headings, paragraphs or document structure, and often carry metadata such as source, page and date. Chunking choices strongly affect whether a retrieval system finds the right passage, so they are one of the first things to tune in a RAG project.
Context window
The context window is the maximum amount of text, measured in tokens, that a language model can consider at once, including the prompt, any supplied documents and its own response. Larger windows let a model read more material in one request, but cost and latency rise with length and accuracy can drop on very long inputs. Knowing the limit helps decide between long-context prompting and retrieval.
Data residency
Data residency refers to the physical or legal location where data is stored and processed, such as servers located in Canada. It is often set by contracts, sector rules or internal policy rather than by a single law. For AI projects, residency determines which hosted models, cloud regions or on-premises options are acceptable for documents and prompts that contain personal or confidential information.
Embeddings
Embeddings are numerical vectors that represent the meaning of text, images or other data, so that similar items sit close together in vector space. An embedding model converts each passage or query into a list of numbers that can be compared mathematically. Embeddings power semantic search, clustering and retrieval-augmented generation, letting systems find relevant content even when the wording differs.
Entity resolution
Entity resolution is the process of determining when records from different sources refer to the same real-world person, company or object, despite spelling differences, missing fields or inconsistent identifiers. It combines matching rules, fuzzy comparison and sometimes machine learning. Accurate entity resolution is essential for sanctions screening, customer data cleanup and any analysis that joins public records or multiple internal systems.
Evaluation (evals)
Evals are structured tests that measure how well an AI system performs on representative tasks, using a set of inputs with expected outcomes or scoring criteria. They can be scored by exact checks, human reviewers or another model acting as a judge. Running evals before and after every change to a prompt, model or data source shows whether quality actually improved and catches regressions before users do.
Fine-tuning
Fine-tuning is the process of further training a pre-trained model on a smaller, task-specific dataset so it adopts a particular format, style or skill. It changes the model's weights, unlike prompting or retrieval, which only change its inputs. Fine-tuning suits consistent output formats or specialized classification; it is usually a poor way to teach a model facts that change often.
Generative AI
Generative AI refers to models that produce new content, such as text, images, audio, code or video, based on patterns learned from training data and the instructions they are given. Large language models are the most widely used type. In business, generative AI is applied to drafting, summarizing, extracting data from documents and building assistants, usually with review steps for anything customer-facing.
Graph RAG
Graph RAG is a form of retrieval-augmented generation that uses a knowledge graph of entities and their relationships, extracted from source documents, alongside or instead of plain text search. The system can follow links between facts spread across many files. It helps with questions such as tracing a supplier to its shipments and customers, which similarity search alone tends to miss.
Grounding
Grounding is the practice of tying a model's answer to specific, verifiable source material, such as retrieved documents, database records or tool results, rather than its general training data. Grounded systems typically return citations so a reader can check each claim. Grounding reduces hallucination and makes AI output auditable, which matters for legal, financial and policy questions.
Guardrails
Guardrails are the controls placed around an AI system to keep its inputs, outputs and actions within acceptable limits. They include input filtering, output validation, topic restrictions, permission checks, spending caps and approval steps before sensitive actions. Guardrails let an organization deploy AI agents and assistants without relying on the model alone to behave correctly in every case.
Hallucination
A hallucination is output from a language model that is fluent and confident but factually wrong or unsupported, such as an invented citation, figure or policy. It happens because models generate likely-sounding text rather than looking facts up. Retrieval, grounding with citations, structured checks and human review are the main ways businesses reduce the risk in production systems.
Human-in-the-loop (HITL)
Human-in-the-loop is a design pattern in which a person reviews, corrects or approves an AI system's output at defined points before it takes effect. Typical checkpoints include low-confidence extractions, payments, customer communications and changes to records. HITL lets organizations automate the routine bulk of a process while keeping accountability for decisions that carry financial, legal or reputational risk.
Hybrid search
Hybrid search combines keyword search, such as BM25, with semantic vector search and merges the results into one ranked list. Keyword matching catches exact terms like product codes, names and clause numbers, while vector search catches paraphrases and related concepts. Most production retrieval systems use hybrid search because neither method alone handles both kinds of query reliably.
Inference
Inference is the act of running a trained model to produce an output, such as generating a response to a prompt, as opposed to training the model. Inference cost depends on model size, input and output length, hardware and request volume. For organizations running their own models, inference capacity and speed largely determine the hardware or cloud budget.
Intelligent document processing (IDP)
Intelligent document processing is the use of OCR, layout analysis and AI models to classify documents and extract structured data from them, such as invoice totals, contract dates or form fields. Unlike template-based capture, IDP can handle varied layouts and unstructured text. It replaces manual data entry and feeds downstream systems, with human review for low-confidence fields.
Knowledge graph
A knowledge graph is a data structure that stores entities, such as people, companies, products or documents, as nodes and the relationships between them as edges. It makes connections explicit and queryable, for example who owns which company or which lot went into which shipment. Knowledge graphs support ownership analysis, compliance work and graph-based retrieval for AI assistants.
Large language model (LLM)
A large language model is a neural network trained on very large amounts of text to predict the next token, which lets it understand instructions and generate, summarize, translate and classify language. LLMs are the core of most generative AI tools and AI agents. They are general-purpose, so business value comes from connecting them to the right data, tools and checks.
LLMOps
LLMOps, an extension of MLOps, is the set of practices for deploying, monitoring and maintaining applications built on large language models. It covers prompt and model versioning, evaluation suites, cost and latency tracking, logging, feedback collection and safe rollout of changes. Without it, AI features tend to drift in quality after launch and become hard to debug or improve.
LoRA (low-rank adaptation)
LoRA, short for low-rank adaptation, is a fine-tuning method that trains a small set of additional weights while keeping the original model frozen. The resulting adapter is a fraction of the model's size and can be swapped in or out at serving time. LoRA makes fine-tuning open-weight models affordable on modest hardware and lets one base model serve several specialized tasks.
Model Context Protocol (MCP)
The Model Context Protocol is an open standard that defines how AI applications connect to external tools, data sources and prompts through MCP servers. An MCP server exposes capabilities, such as querying a database or creating a ticket, in a consistent format any compatible client can use. It lets an organization expose an internal system once, securely, and reuse it across several AI tools and agents.
Multi-agent system
A multi-agent system is an AI design in which several specialized agents, each with its own instructions and tools, cooperate on a task, often coordinated by an orchestrator agent or a fixed workflow. One agent might research, another draft and a third check the result. Splitting work this way can improve reliability on complex processes, but adds cost and needs careful testing.
Multimodal AI
Multimodal AI refers to models that can take in or produce more than one type of data, such as text, images, audio and video, within a single system. A multimodal model can read a scanned form, describe a photo or transcribe and summarize a call. This lets businesses automate work involving mixed media, such as inspection photos, receipts and recorded meetings.
OCR (optical character recognition)
Optical character recognition is technology that converts images of text, such as scanned pages, photos or image-only PDFs, into machine-readable text. Modern OCR also detects layout, tables and handwriting with varying accuracy. OCR is usually the first step in making paper archives searchable and in any document AI pipeline that works with scanned records.
On-premises AI
On-premises AI means running AI models and their supporting software on hardware the organization owns or controls, in its own office or data centre, rather than on a provider's cloud service. Prompts and documents never leave the organization's environment. It is chosen when privacy obligations, client contracts or policy rule out sending data to third-party services.
Open-weight model
An open-weight model is an AI model whose trained parameters are published so anyone can download, run and adapt it, subject to its licence terms. The training data and code may or may not be released, which is why the term differs from fully open source. Open-weight models make private and on-premises deployments possible without depending on a single vendor's API.
PIPEDA
PIPEDA, the Personal Information Protection and Electronic Documents Act, is Canada's federal privacy law governing how private-sector organizations collect, use and disclose personal information in commercial activity. It is built on principles such as consent, limiting use, safeguards and accountability, and some provinces have their own similar laws. AI projects that handle personal information should be reviewed against it; this definition is general information, not legal advice.
Private LLM
A private LLM is a large language model deployed in an environment the organization controls, such as its own servers or a dedicated cloud tenancy, so prompts and documents are not shared with a public AI service. It is usually built on an open-weight model. Private LLMs suit legal, healthcare, financial and public-sector teams with strict confidentiality or data residency requirements.
Prompt engineering
Prompt engineering is the practice of designing the instructions, examples and context given to a language model so it produces accurate, consistent and correctly formatted output. It includes role and task framing, worked examples, output schemas and step-by-step guidance. Careful prompting, tested against evals, is often the cheapest way to improve an AI feature before considering fine-tuning.
Prompt injection
Prompt injection is an attack in which text supplied to an AI system, directly by a user or hidden in a web page, email or document, contains instructions that try to override the system's intended behaviour. It is a particular risk for agents that read untrusted content and can take actions. Defences include limiting tool permissions, separating trusted instructions from data and requiring approval for sensitive actions.
Quantization
Quantization is a technique that stores a model's weights at lower numerical precision, such as 8-bit or 4-bit instead of 16-bit, to reduce memory use and speed up inference. It usually costs a small amount of accuracy, which should be measured on the actual task. Quantization is what lets capable open-weight models run on a single workstation or modest GPU server.
Reranking
Reranking is a retrieval step in which a second model rescores an initial set of search results by how well each one answers the query, then reorders them. The first search is fast but approximate; the reranker is slower but more precise. Adding a reranker is a common way to improve answer quality in RAG systems without rebuilding the index.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation is a technique in which an AI system first searches a collection of documents or data for relevant passages, then gives those passages to a language model to answer the question. The model answers from current, specific sources and can cite them. RAG lets organizations build assistants over their own contracts, policies and knowledge without retraining a model.
Robotic process automation (RPA)
Robotic process automation is software that mimics a person's clicks and keystrokes to complete rule-based tasks in existing applications, such as copying data between screens. It works well for stable, structured processes but breaks when layouts or inputs change. Many organizations now pair or replace RPA with APIs and AI models that can handle unstructured documents and exceptions.
Semantic search
Semantic search is a search method that matches queries to content by meaning rather than exact keywords, typically by comparing embeddings. A query for "terminating a lease early" can find a clause about "early surrender of premises". It makes internal knowledge bases, document archives and support content easier to search, and it is the retrieval step behind most RAG systems.
Small language model (SLM)
A small language model is a language model with relatively few parameters, typically small enough to run on a laptop, phone or single GPU. SLMs are less capable at broad reasoning than large models but can perform well on narrow tasks such as classification, extraction or routing, especially after fine-tuning. They offer lower cost, faster responses and easier private deployment.
Structured output
Structured output is model output constrained to a defined format, such as JSON matching a schema, instead of free-form text. Many model APIs and serving tools can enforce a schema so every response has the expected fields and types. Structured output is what makes it practical to pass AI-extracted data directly into databases, ERPs and automated workflows.
Tokens
Tokens are the units of text a language model reads and writes, usually whole words, parts of words or punctuation. Models measure context window size, speed and pricing in tokens, and an English word is typically a little more than one token, depending on the tokenizer. Estimating token volume is the starting point for forecasting the running cost of an AI application.
Tool calling (function calling)
Tool calling, also called function calling, is a model capability in which the model returns a structured request to run a named function with specific arguments, such as looking up an order or sending an email, instead of plain text. The application executes the function and returns the result to the model. Tool calling is the mechanism that lets AI agents act on business systems.
Vector database
A vector database is a data store designed to index embeddings and quickly find the vectors most similar to a query, often alongside metadata filters and keyword search. Dedicated products exist, and many general databases, such as PostgreSQL with an extension, offer the same capability. It is the storage layer behind semantic search and most retrieval-augmented generation systems.
Workflow orchestration
Workflow orchestration is the coordination of the steps, systems and people in a multi-step process, including ordering, retries, branching, scheduling and hand-offs. An orchestrator tracks the state of each case so nothing is lost when a step fails or waits for approval. In AI automation, orchestration keeps agents, integrations and human reviews reliable and auditable.
Want these ideas applied to your business?
A 30-minute call is enough to tell whether AI is a fit, what a first project would look like and what it would take to run it in production.