Glossary

An approach to AI where systems pursue goals autonomously over multiple steps — planning, acting, checking results, and adjusting — rather than responding to a single prompt. Agentic AI is the paradigm shift happening now in enterprise software: from AI that answers questions to AI that completes tasks.

A process that runs autonomously through multiple steps using AI agents — detecting a trigger, gathering information, making decisions, and taking actions — without a human driving each step. Agentic workflows replace manual process chains for recurring operational tasks where the rules are known and the volume is high.

A hypothetical form of AI that can perform any intellectual task a human can — reasoning, learning, and adapting across entirely different domains without being retrained. AGI does not exist today. All current AI, including the most advanced large language models, is narrow: built for specific tasks.

A software program that performs tasks autonomously by perceiving its environment, making decisions, and taking actions — without constant human supervision. AI agents can handle exceptions, trigger workflows, and adapt based on rules or learned patterns.

The coordination layer that manages how multiple AI agents work together on complex, multi-step processes. Orchestration handles task routing, sequencing, state management, and error recovery — the infrastructure that turns individual agents into a coherent operational system.

The challenge of ensuring an AI system behaves in accordance with its intended goals — not just technically correct, but consistent with the values, constraints, and outcomes its operators actually want. Misaligned AI optimizes for the wrong objective, often in ways that are not immediately visible.

An AI assistant embedded in a tool or workflow that helps a human user complete tasks faster — suggesting next steps, drafting outputs, surfacing relevant data, or flagging issues. The human stays in control; the copilot handles the cognitive grunt work.

AI evaluations are structured tests that measure whether an AI model or agent is performing correctly — catching errors, regressions, and edge cases before they reach production. Without evaluations, you are flying blind.

The policies, processes, and controls that ensure AI systems operate within defined boundaries — accurately, fairly, and in compliance with regulations. For European companies, AI governance increasingly means EU AI Act compliance: risk classification, human oversight requirements, and audit trails.

The coordination of multiple AI models, tools, APIs, and data sources in a production system. AI orchestration defines how these components are connected, sequenced, and managed — it is the infrastructure layer that makes multi-component AI deployments reliable and maintainable in enterprise environments.

A hypothetical AI that surpasses human intelligence across every domain — not just matching human capability, as AGI aims to, but exceeding it by a significant margin. ASI does not exist and remains a theoretical concept discussed primarily in AI safety research.

The execution of a large volume of transactions or computations as a single grouped job, typically on a scheduled basis rather than in real time. Overnight batch jobs for invoicing, reconciliation, and demand planning remain common in ERP-heavy operations, and AI inference must be designed to fit these windows.

The process of measuring an AI model's performance on standardized tasks to evaluate its accuracy, speed, reliability, and suitability for a specific use case. Benchmarks help compare models objectively and track performance over time — but only if the benchmark matches what the model actually needs to do in production.

Systematic errors in an AI model's outputs caused by skewed training data, flawed assumptions, or poorly designed objectives. Bias causes AI to produce results that are consistently wrong in a predictable direction — often in ways that are invisible until you look closely at outcomes across different inputs.

A prompting technique that instructs an AI model to reason through a problem step by step before producing a final answer. Chain of thought makes the model's reasoning visible, reduces errors on complex multi-step tasks, and makes it easier to identify where the logic breaks down.

A software interface that handles conversations — answering questions, routing requests, or guiding users through a process — using either predefined rules or AI language models. Modern AI-powered chatbots handle nuanced queries and maintain context across turns. Older rule-based chatbots follow fixed scripts and break outside their defined scope.

A conversational AI product built by OpenAI, based on its GPT large language models. ChatGPT can draft text, answer questions, summarize documents, write code, and reason through problems in natural language. It is a general-purpose tool — powerful across a wide range of tasks, but not designed for deep integration with specific business systems out of the box.

An AI task that assigns an input — a document, email, image, or data record — to one of a set of predefined categories. Classification is one of the most widely deployed AI capabilities in business operations: routing invoices, categorizing support tickets, labeling transactions, sorting incoming documents by type.

A family of large language models developed by Anthropic, designed with a focus on safety, accuracy, and following complex instructions reliably. Claude is widely used in business automation for tasks that require nuanced reasoning, structured output generation, and consistent behavior under varied inputs.

The output generated by a large language model in response to an input prompt. A completion is whatever the model produces — a sentence, a paragraph, a JSON object, a decision — based on the instructions and context it was given. The quality of a completion depends on the clarity of the prompt, the model's parameters, and how well the task fits the model's training.

The raw processing power required to train and run AI models. More compute means larger models, faster inference, and the ability to handle more complex tasks — but also higher cost. For businesses deploying AI, compute decisions shape the economics of every automated workflow.

The ability of an AI agent to operate a computer's graphical interface — clicking buttons, filling forms, reading screens — the same way a human would. Introduced by Anthropic with Claude in 2024, computer use enables automation of software that has no API, including legacy ERP systems.

A numerical value (typically 0–100%) indicating how certain an AI model is about a specific output — an extracted field, a classification decision, or a match result. Confidence scores determine when to trust automation fully and when to route a result to a human for review.

The process of adding structured metadata, context, or derived attributes to raw content — documents, records, emails, product data — to make it more useful for search, automation, and decision-making. Enrichment transforms unstructured information into something a system can act on.

The maximum amount of text an LLM can read and reason over in a single interaction. Everything outside the context window is invisible to the model. For processing long documents like 100-page contracts or multi-invoice batches, context window size is a hard technical constraint.

AI systems designed to conduct natural language conversations — understanding intent, maintaining context across multiple turns, and generating relevant responses. Conversational AI powers chatbots, virtual assistants, and voice interfaces. The quality gap between basic chatbots and true conversational AI is large.

The practice of artificially expanding a training dataset by creating modified versions of existing examples to expose a model to more variation than the original data contains. Augmentation improves model robustness without requiring additional real-world data collection.

The automated process of pulling specific data fields from documents, emails, or databases — invoice totals, delivery dates, product codes, contract terms — and converting them into structured records a system can process. Extraction eliminates manual data entry at the front door of document-heavy workflows.

The process of collecting data from source systems — ERPs, databases, APIs, file uploads, email — and loading it into a pipeline or platform where it can be processed. Ingestion is the first step in any data workflow: nothing gets analyzed, enriched, or acted on until it has been ingested.

Organized collections of data used to train, test, and evaluate AI models. The quality, size, and representativeness of a dataset directly determine how well a trained model performs in production. Bad data produces bad models — regardless of how sophisticated the training process is.

The practice of capturing, encoding, and automating the judgment calls that currently live in experienced employees' heads. Decision intelligence turns tacit operational knowledge — when to escalate, when to approve, when to flag — into rules and models that run consistently inside existing systems.

A subset of machine learning that uses neural networks with many layers to learn patterns directly from raw data — images, text, audio, sensor readings — without manually engineered features. Deep learning is the technology underlying most modern AI breakthroughs, including large language models, image recognition, and document understanding.

A deterministic AI system produces the same output every time it receives the same input. No randomness, no variation — the result is predictable and repeatable. In operations, this matters wherever consistency and auditability are non-negotiable.

Diffusion models are a class of generative AI that create images, audio, or video by learning to reverse a noise process — starting from random static and progressively refining it into coherent output. They power tools like Stable Diffusion and DALL-E.

The umbrella term for AI systems that can read, understand, and extract structured information from business documents — invoices, contracts, delivery notes, customs forms, and more. Document intelligence goes beyond OCR by understanding context, not just characters.

An embedding is a numerical representation of a piece of text, image, or data as a vector — a list of numbers that captures its meaning in a form a machine can compare and reason about. Vectorizing is the process of creating that representation; the embedding is the result.

The technical and architectural approaches for connecting AI systems to ERP platforms — SAP, Exact, Business Central, AFAS, and others — so that AI can read operational data, trigger actions, and write results back into the system of record without manual re-entry.

The process of identifying, investigating, and resolving cases that fall outside normal operational rules. In manufacturing, logistics, and wholesale, exceptions — invoice mismatches, stock discrepancies, delivery delays — consume a disproportionate share of operations team time and are the primary target for AI automation.

The automated detection and routing of operational anomalies — invoice mismatches, stock discrepancies, delivery errors — that fall outside the expected range and require human judgment or escalation. AI exception handling replaces the manual review queues that slow down finance, logistics, and procurement teams.

Explainable AI refers to methods and practices that make an AI model's decisions understandable — showing which inputs drove an output, how the model arrived at a conclusion, and why one option was chosen over another. It is the difference between a black box and an auditable system.

Few-shot learning is the ability of an AI model to perform a new task correctly after seeing only a handful of examples — sometimes as few as two or three. It is how modern language models adapt to specific formats, tones, or business rules without full retraining.

Fine-tuning is the process of taking a pre-trained AI model and training it further on a smaller, domain-specific dataset so it performs better on a particular task. It adapts a general model to your specific language, formats, and business logic.

A foundation model is a large AI model trained on broad, diverse data at massive scale — designed to serve as a general-purpose base that can be adapted to many different tasks. GPT-4, Claude, Gemini, and Llama are all foundation models.

A capability that allows LLMs to call external tools and APIs in a structured, reliable way — triggering database queries, writing ERP records, sending notifications — rather than only generating text. Function calling is the mechanism that makes AI agents operational rather than conversational.

Gemini is Google DeepMind's family of multimodal AI models — capable of processing text, images, audio, video, and code within a single model. It competes directly with GPT-4 and Claude, and powers AI features across Google's product suite.

Generative AI is a category of AI that produces new content — text, images, code, data — rather than simply classifying or predicting from existing inputs. It is the technology behind large language models, image generators, and code assistants.

GPT is the model architecture developed by OpenAI that underpins ChatGPT, GPT-4, and related systems. It uses the transformer architecture to predict and generate text, and is trained on large volumes of internet text before being fine-tuned for specific applications.

A GPU is a processor designed for parallel computation — originally for rendering graphics, now the dominant hardware for training and running AI models. Without GPUs, modern AI as we know it would not exist at the scale it does today.

When an AI model generates output that is factually incorrect or fabricated — stated with confidence, with no indication that it is wrong. In operational contexts, hallucinations are not a theoretical risk: an AI that invents a product code, a supplier name, or a contract term causes real downstream errors.

Human-in-the-loop (HITL) means a human reviews, approves, or corrects AI outputs at defined points in an automated workflow. It is not a sign that automation failed — it is how you build automated systems that can be trusted with consequential decisions.

Inference is the process of running a trained AI model on new inputs to generate outputs. Every time you send a document to an AI for processing, you are performing inference. It is the operational phase of AI — distinct from training, which is where models learn.

The industry category for AI-powered systems that automate the capture, extraction, validation, and routing of data from business documents. IDP platforms combine OCR, machine learning, and workflow automation to replace manual document handling in finance, logistics, and procurement.

A knowledge graph is a structured representation of entities and the relationships between them — products, suppliers, locations, processes — stored in a way that lets AI systems reason about connections, not just retrieve isolated facts.

A type of AI model trained on large volumes of text that can understand and generate language. LLMs are the foundation of AI agents, copilots, and document processing tools — they are what reads the invoice, understands the supplier email, and drafts the exception summary.

Latency is the time it takes for an AI system to return a response after receiving an input. In operational AI, it determines whether automated workflows feel instantaneous or whether they become a bottleneck that slows down the people and processes they are supposed to accelerate.

Llama is Meta's family of open-source large language models. Released with permissive licences, they can be downloaded, fine-tuned, and deployed on your own infrastructure — giving organisations full control over their AI stack without relying on a third-party API.

Machine learning is the discipline of building systems that improve their performance on a task by learning from data, rather than following explicitly programmed rules. It is the foundation of modern AI — from image recognition to language models to anomaly detection.

An open standard developed by Anthropic in 2024 that defines how AI models connect to external tools, databases, and services. MCP gives AI agents a consistent interface for reading files, querying systems, and triggering actions — without custom integration code for every tool.

Metadata is data that describes other data. The invoice total is data. The fact that it is an invoice, from supplier X, received on a specific date, tagged to a specific cost centre — that is metadata. It is what makes documents findable, sortable, and usable in automated workflows.

Mistral is a French AI company and the family of language models it produces — known for strong performance at relatively small model sizes and an open-weights approach that makes them practical to self-host. Mistral 7B and Mixtral 8x7B are widely used in production deployments.

The set of settings that control how an AI model behaves during inference — including temperature, response length, sampling method, and output format. Model configuration does not change the model's weights; it shapes how the model uses what it already knows.

The numerical weights stored inside a trained AI model that encode everything it learned during training. Parameters are not settings you configure at runtime — they are baked into the model during training and determine what it knows and how it reasons.

A configuration where multiple AI agents work together on a shared task — each handling a distinct role, sharing state, and passing work between each other. Multiagent systems are how complex operational workflows get automated end-to-end without forcing one agent to do everything.

An AI system that can process and generate multiple types of data — text, images, documents, audio — within a single model. Instead of routing a scanned invoice to one system and its metadata to another, a multimodal model handles both in one pass.

A technique for training a small set of learnable prompt parameters to make a single model perform well across multiple tasks simultaneously. Instead of fine-tuning separate models for each task, MPT optimizes shared prompts that steer one model to handle all of them.

The field of AI that enables computers to read, interpret, and generate human language. NLP is the foundation beneath every AI feature that touches text — document extraction, classification, search, summarization, and automated communication.

A mathematical model loosely inspired by the brain, made up of layers of interconnected nodes that learn to recognize patterns in data. Neural networks are the core architecture behind most modern AI — from image recognition to language models.

Technology that converts images of text — scanned documents, photos, PDFs with embedded images — into machine-readable characters. OCR is the foundational layer of document processing, but alone it only produces raw text. AI is required to understand what that text means.

The process of analyzing a document, piece of text, or data structure to identify its components and extract meaning from them. Parsing breaks down raw input into structured elements — sentence boundaries, field values, entity types — that downstream systems can use.

The initial training phase where an AI model learns general knowledge and language patterns from a massive dataset — before any task-specific tuning. Pre-training is what gives models like GPT or Claude their broad capabilities.

The use of technology to execute business processes with reduced human intervention. Process automation is the broad category that includes RPA, workflow automation, and AI-driven automation — each suited to different levels of process complexity and variability.

The input you give to an AI model to direct its response. A well-constructed prompt specifies role, context, instruction, output format, and examples — each element shaping how the model interprets the request and what it produces.

A technique where the output of one AI prompt becomes the input to the next, breaking a complex task into a sequence of focused steps. Each step in the chain does one thing well, and the results accumulate toward a final answer.

The practice of designing the instructions given to an AI model to produce accurate, consistent, and useful outputs. In operational AI, prompt engineering determines whether a model extracts the right fields from an invoice, applies the right rules to an exception, or generates a response that matches your process requirements.

A reusable, parameterized prompt structure where variable placeholders are filled with specific data at runtime. Prompt templates are how operational AI workflows apply consistent AI instructions to thousands of different documents or records without writing a new prompt each time.

An architecture that connects an AI model to external data sources at query time — retrieving relevant documents, records, or facts before generating a response. RAG is what allows an AI agent to answer questions based on your actual ERP data, contracts, and supplier records rather than its training data alone.

A generation of AI models — including OpenAI's o1/o3 and Anthropic's Claude 3.7 — that work through complex problems step by step before producing an answer. They trade speed for accuracy on multi-step decisions, making them better suited for operational edge cases where a wrong answer has real consequences.

A machine learning approach where an agent learns by taking actions in an environment and receiving rewards or penalties based on outcomes. The agent optimizes its behavior over time to maximize cumulative reward — without being told the correct answer upfront.

A training technique where human raters compare pairs of AI outputs and indicate which is better. Those preferences train a reward model, which then guides the AI to produce responses humans prefer — making the model more helpful, accurate, and safer.

Software that automates repetitive, rule-based tasks by mimicking human interactions with applications — clicking buttons, copying data, filling forms. RPA works well on stable, predictable processes but breaks when inputs vary or exceptions occur.

A search method that finds results based on meaning rather than exact keyword matches. Semantic search uses vector embeddings to measure conceptual similarity — so a query for "late delivery penalty" finds a contract clause about "delivery schedule non-compliance" even though no words match.

An NLP technique that classifies the emotional tone of text — positive, negative, or neutral. In business contexts, it goes beyond simple polarity to detect urgency, frustration, dissatisfaction, and risk in customer communications, supplier messages, and internal feedback.

A theoretical future point where AI surpasses human intelligence and begins improving itself recursively — leading to rapid, unpredictable technological change. The singularity is a concept from theoretical AI research, not a near-term operational reality.

Data organized in a predefined format with clear fields, types, and relationships — stored in databases, spreadsheets, or ERP systems. Structured data is what AI can query, compare, and act on directly, without a parsing step.

An AI response formatted according to a defined schema — typically JSON, XML, or a fixed template — rather than free-form text. Structured output is what makes AI responses usable by downstream systems without additional parsing.

A hidden instruction layer sent to an LLM before the user's input that configures the model's behavior, persona, constraints, and context. The system prompt is where AI agents receive their role definition, business rules, and operational guidelines.

A model configuration parameter that controls how random or deterministic an AI model's outputs are. Low temperature produces consistent, predictable responses. High temperature produces more varied, creative outputs. For most operational automation, low temperature is the right default.

An open-source machine learning framework developed by Google for building and training neural networks. TensorFlow is a tool for AI researchers and engineers — not something operations teams interact with directly.

The basic unit of text that AI language models process. Tokens are not words — they are chunks of characters, typically 3–4 characters each. Every input and output is measured and billed in tokens. The maximum number of tokens a model can process at once is its context window.

The labeled or unlabeled dataset used to teach an AI model how to recognize patterns, make predictions, or extract information. The quality, volume, and representativeness of training data directly determines how well a model performs on real-world tasks.

The neural network design that powers most modern AI language models, including GPT, Claude, and Gemini. Transformers use a mechanism called self-attention to process entire sequences of text simultaneously, making them far more effective at understanding context than earlier architectures.

Information that has no predefined format or schema — emails, PDFs, scanned documents, free-text notes, images. Most operational data is unstructured, and processing it at scale requires AI rather than traditional database queries.

A database designed to store and search high-dimensional numerical representations (vectors) of text, documents, or other data. Vector databases power semantic search and retrieval-augmented generation (RAG) by finding items that are conceptually similar — not just textually identical.

The use of software to execute sequences of tasks automatically, reducing or eliminating manual steps. Workflow automation ranges from simple rule-based triggers to AI-driven processes that can handle variation, make decisions, and route exceptions.

The ability of an AI model to perform a task it was never explicitly trained on, using only a description or instruction. A zero-shot model generalizes from its training to handle new task types without requiring labeled examples.