AI Glossary

The concepts of enterprise AI adoption in short definitions that start from practice, not the textbook.

Core concepts

LLM (large language model): An AI model trained on large volumes of text that generates output token by token as a continuation of its input. Proprietary models (such as GPT, Claude and Gemini) are accessed through an API, while open-weight models (such as Llama, Qwen and Mistral) can also run on a company's own infrastructure. In business settings, LLMs handle document processing, customer service responses and text generation.Token Context window Prompt Fine-tuning Generative AI Integration
Token: The basic unit of text a language model works with: both input and output are split into tokens, each typically a word fragment or a short word. API-based models are billed per token, so token counts directly determine running costs. Context window sizes are also expressed in tokens.LLM (large language model)Context window Embedding
Context window: The upper limit, measured in tokens, on how much text a language model can take into account at once — input and generated output combined. Content outside the window is not visible to the model, so long documents require chunking or a RAG approach. In practice this determines how much material can be passed in a single request.LLM (large language model)Token RAG (retrieval-augmented generation)
Hallucination: The phenomenon of a language model producing a confident statement that is false or unsupported. In multi-step systems the error can propagate: one component's hallucination becomes the next component's input. Production systems reduce the risk with output checks, source-grounded answering (RAG) and human oversight.LLM (large language model)RAG (retrieval-augmented generation)Guardrail Human oversight
Embedding: A representation of text or other data as a numeric vector, where content with similar meaning maps to nearby vectors. This enables semantic search: document passages related to a query can be retrieved by vector distance. RAG systems and vector databases are built on embeddings.Vector database RAG (retrieval-augmented generation)Token
Vector database: A database designed to store embeddings and retrieve them quickly by similarity. Given a query embedding, it returns the stored vectors nearest to it, so relevant document passages can be found by meaning. Used in RAG systems and enterprise knowledge bases; ChromaDB is one example.Embedding RAG (retrieval-augmented generation)LLM (large language model)
RAG (retrieval-augmented generation): An architecture in which the language model receives relevant passages retrieved from an organization's own documents before answering, and grounds its response in them. Retrieval is typically done with embeddings and a vector database. The approach ties answers to the company's knowledge base without retraining the model and reduces hallucination.Vector database Embedding Hallucination Fine-tuning Internal AI Knowledge Base
Fine-tuning: Further training of a pre-trained language model on an organization's own data so that it performs more accurately in a specific domain or on specific document formats. It is typically applied to open-weight models, which can run on own infrastructure and be meaningfully retrained. It is warranted when the task requires deep domain knowledge or handling of company-specific formats.LLM (large language model)Open-weight model RAG (retrieval-augmented generation)AI Document Processing
Prompt: The input handed to a language model: an instruction, a question or text to process, on which the model bases its output. How a prompt is phrased directly affects the quality and form of the output. In business applications, prompts are usually built from templates into which the system inserts the current data.System prompt Prompt engineering LLM (large language model)
Prompt engineering: The practice of designing and testing prompts so that a model reliably returns output of the desired format and content. Its tools include precise instructions, worked examples (few-shot) and prescribed output formats. In production systems, prompts are versioned and validated with measurements.Prompt Few-shot prompting System prompt
Few-shot prompting: A prompting technique in which a few solved examples are included alongside the instruction, and the model processes the new input following their pattern. It improves output consistency without fine-tuning, which makes it a common step before retraining a model is considered. Typical uses include fixing the output format of classification and data extraction tasks.Prompt engineering Prompt Fine-tuning
System prompt: A standing set of instructions for a language model that applies before every user request: it fixes the role, tone, constraints and output rules. It is handled separately from the user prompt and is typically not visible to the user. In production systems it is one of the mechanisms for enforcing behavioral rules and part of the defence against prompt injection.Prompt Prompt injection Guardrail

Agents and automation

AI agent: A software component built on an LLM that decides its own next step toward a goal: it calls tools, reads and writes data, and plans further based on intermediate results. An agent is justified when a task requires planning, adaptation and judgment at multiple steps. For a predictable workflow, a simpler pipeline with conditional logic and human checkpoints is sufficient.Agentic workflow Orchestration LLM (large language model)Human-in-the-loop AI Agent Development
Agentic workflow: A multi-step workflow executed by one or more AI agents, in which the next step depends on intermediate results. Reliable operation involves state persistence — checkpoints from which the process can resume after a failure — along with exception handling and checks that catch error cascades. The handoff between agents is an error-prone point, so production systems typically manage it with a typed, validated state schema.AI agent Orchestration Exception handling
MCP (Model Context Protocol): An open protocol that standardizes how AI models connect to external systems — databases, document stores, enterprise applications. One MCP server is written per data source, and any model that supports the protocol can connect to it, which reduces the number of custom connectors and the maintenance burden. Configuring MCP server permissions requires the same care as API gateway rules.AI agent API integration Agentic workflow Generative AI Integration
RPA: The automation of rule-based tasks with software robots: the bot clicks, copies fields, and moves data between systems, producing the same result on every run. It suits steps for which an exact flowchart can be drawn; where the decision depends on the input — unstructured text, varying formats, judgment calls — AI does the work. Most real-world processes are a mix of the two, which is why hybrid systems are common.AI agent Exception handling API integration Process Automation
Human-in-the-loop: A system pattern in which automated processing requests a human decision at defined points: uncertain cases are flagged and routed to a colleague with the problematic parts highlighted. High-confidence cases flow through automatically, while the decision on exceptions stays with a person. The EU AI Act assigns human oversight to high-risk systems; human-in-the-loop checkpoints are one way to implement it.Human oversight Exception handling Guardrail AI Document Processing
Orchestration: The coordination of multiple agents or processing steps: which one runs when, what state is passed between them, and what happens on partial failure. In multi-agent systems it becomes a distributed systems problem, involving concurrent execution, timeout handling and retry logic. Frameworks such as LangGraph or CrewAI manage the execution graph and the handoffs between agents.AI agent Agentic workflow Monitoring
Prompt injection: An attack technique in which processed content — an email, a document, a web page — contains instructions that the model executes as if they came from the user or the operator. The risk is higher for agents with tool access, because an injected instruction can read out data or trigger actions. Defenses include separating instructions from input data, narrowing permissions, and guardrails.Guardrail System prompt Prompt AI agent
Guardrail: A control built around an AI system that filters or limits inputs and outputs according to defined rules. Examples include checks on format and value ranges, hard per-run limits — maximum iterations, tokens and cost — and the routing of uncertain cases to human review. When a limit is hit, the system stops and escalates before errors or costs compound further.Prompt injection Human-in-the-loop Monitoring Hallucination

Adoption and operations

Proof of concept (PoC): A small-scale implementation that checks whether a planned solution works technically on the given data and in the given environment. In AI adoption, a proof of concept implements the selected use case in a limited scope and measures the outcome. It grounds the decision about full implementation.Pilot ROI (return on investment)
Pilot: A pre-implementation phase in which the system operates on real data, in a real process, within a limited scope. A pilot measures results under production conditions, for example what share of processed items flows through without human intervention. The measured results decide whether the solution proceeds to full implementation.Proof of concept (PoC)ROI (return on investment)Monitoring
ROI (return on investment): Return on investment expresses how the return of an investment relates to its cost. In AI projects, the calculation typically rests on the manual working time replaced and the volume data of the process, set against the cost of implementation and operations. The payback period shows how long it takes for the cumulative return to reach the spend.Proof of concept (PoC)Pilot AI ROI Calculator
On-premise: A deployment model in which the system runs on the company's own servers instead of a public cloud service. In AI systems it is typically paired with open-weight models, so the data never leaves the company environment. It is a common choice for processes handling sensitive data and for regulated environments; private cloud is the related alternative.Open-weight model Multi-tenant
Open-weight model: An AI model whose trained weights are publicly available, so the model can run on the company's own infrastructure. In on-premise or private cloud deployments this is what lets the data stay within the company environment. An open-weight model can be fine-tuned on the company's own data, and running it does not depend on an external provider's API.On-premise Fine-tuning LLM (large language model)
API integration: The programmatic connection of systems through an interface (API) provided by the target system. In AI adoption this is how a model connects to existing ERP, CRM, and document management systems, without replacement or downtime. Where the target system offers an API, the integration is typically built on it instead of an RPA bot; permissions stay in the existing system.RPA MCP (Model Context Protocol)Orchestration Generative AI Integration
GitOps: An operations method in which the full configuration of the platform and the workloads running on it lives in version control, and every change reaches the environments from there. Every deployment is traceable: it is visible who changed what and when. A rollback is the restoration of a previous, working state, instead of manual intervention.Blue-green release Multi-tenant Monitoring Enterprise AI Platforms
Blue-green release: A release technique in which the new version starts alongside the old one and only receives traffic after it passes the checks. The switch happens without downtime. On failure the rollback is immediate, because the old version keeps running.GitOps Monitoring
Multi-tenant: An architecture in which several business units or several workloads run on the same platform, with separated access and resources. The separation applies to access and data alike. Launching a new use case therefore needs no new infrastructure: it lands on the existing platform.GitOps Blue-green release On-premise Enterprise AI Platforms
Monitoring: The continuous observation of a production system: one place shows what runs, what it costs, and where it fails. In AI systems, monitoring also tracks output quality; when accuracy falls below a threshold, an alert is raised. The logging that accompanies it also forms the basis of EU AI Act documentation.Logging Human oversight GitOps

EU AI Act and compliance

EU AI Act: The European Union's AI regulation, which classifies AI systems by risk and ties the obligations to the risk category. It uses four categories: unacceptable, high, limited and minimal risk. The obligations become applicable in stages, and the regulation assigns separate obligations to the provider and the deployer of a system. The first step of practical preparation is the risk classification, per system, with written reasoning.Risk categories Provider and deployer Human oversight EU AI Act Compliance
Risk categories: The four risk categories of the EU AI Act: unacceptable, high, limited and minimal risk. The obligations follow from the category. Most AI applications running in companies — document processing, customer service assistants, internal knowledge bases — are limited or minimal risk; systems that make or prepare decisions about people typically fall into the high-risk category. The risk classification is worth running per system, with written reasoning, and it has to be re-run when the system changes.EU AI Act Provider and deployer Human oversight Logging EU AI Act Compliance
Provider and deployer: The two roles of the EU AI Act: the regulation assigns separate obligations to the provider and the deployer of a system. The scope of the obligations depends on the role as well, so the role is worth clarifying together with the risk classification, in writing. On a custom build the two roles can differ, and later the role decides which documentation is whose responsibility.EU AI Act Risk categories
Data protection impact assessment (DPIA): An analysis required by the GDPR that assesses the risks a data processing operation poses to the rights of natural persons. For an AI system handling personal data, the DPIA arises alongside, and separately from, the EU AI Act risk classification; the legal assessment belongs to legal counsel.EU AI Act Risk categories Data extraction
Logging: Making the system's decisions traceable: it is recorded what the input was, what the output became, and who approved it. The EU AI Act lists it among the obligations for high-risk systems, alongside risk management, data quality, technical documentation and human oversight. Logging is worth building as part of the system; a later check retrieves these log entries.EU AI Act Human oversight Monitoring Risk categories
Human oversight: Review points built into the process: who looks at the system's decision, what the reviewer sees, and where they can intervene. The EU AI Act lists it among the obligations for high-risk systems, alongside risk management, data quality, technical documentation and logging. In practice it is a system design question: the review points are worth building as part of the system, together with the logging.EU AI Act Human-in-the-loop Logging Exception handling

Document processing

OCR (optical character recognition): OCR turns the image of a scanned or photographed document page into text. On its own it cannot determine which value in that text is the total or the invoice number; in traditional systems, templates defined per layout answer that. In production document processing systems, OCR typically serves as a pre-processing step, with an AI model doing the field extraction. Digital PDFs skip the OCR step.Intelligent document processing Data extraction Classification
Intelligent document processing: A multi-step processing pipeline that reads incoming documents, classifies them by type, extracts their data, and routes them to the right process. Its typical stages are OCR pre-processing for scanned pages, AI-based data extraction, rule-based validation, and recording in the ERP system. Uncertain cases are routed to exception handling, where the decision is made under human review. Contracts, orders, forms, and documents attached to incoming mail can all run through such a pipeline.OCR (optical character recognition)Data extraction Classification Exception handling AI Document Processing
Data extraction: The processing step in which the system pulls the fields a process needs from a document: line items, amounts, currencies, parties, deadlines. Template-based extraction reads from predefined coordinates; AI-based (LLM) extraction interprets the document's content, so it also handles layouts it has not seen before. In production systems, extracted data passes rule-based validation before moving on in the process.OCR (optical character recognition)Intelligent document processing Exception handling LLM (large language model)AI Invoice Processing
Classification: In document processing, the step in which the system determines the type of an incoming document — contract, order, worksheet, form — and routes it to the right process or colleague accordingly. Documents arriving by email, from an ERP, or as scanned PDFs thus enter the same processing pipeline. In machine learning, classification is a broader concept: it assigns an input to one of a set of predefined categories.Intelligent document processing Data extraction OCR (optical character recognition)
Exception handling: The step in automated document processing that routes uncertain or discrepant cases to human review. The system validates extracted data against rules, flags discrepancies, and hands over the problematic document with the questionable fields highlighted; the decision stays with the reviewer. The step is needed because AI-based extraction fails less often than template-based systems on varied layouts, but when it does, the wrong value looks no different from a correct one.Intelligent document processing Data extraction Human-in-the-loop Human oversight