Enterprise AI glossary
The terms that show up in any automation proposal, explained in two or three lines and without marketing. Useful for reading a quote and understanding what you're buying.
26 terms
Fundamentals
- Artificial intelligence
- A set of techniques that let a system perform tasks requiring interpretation, such as understanding text, recognising an image, or estimating an outcome. In a company it is applied to concrete, repetitive tasks rather than as a general capability. Read more about this →
- Machine learning
- A branch of artificial intelligence where a model learns patterns from historical data instead of following hand-written rules. It is used to classify, detect anomalies, and estimate future values when enough history exists.
- AI agent
- A system that interprets a natural language request, retrieves the information it needs from a company's systems, and carries out a concrete task. Unlike a chatbot, which only converses, an agent acts and logs every action. Read more about this →
- Chatbot
- A conversational system that answers questions using information loaded into it beforehand. It does not access company systems, so it can explain how a process works but cannot report the status of one specific person's case. Read more about this →
- Large language model (LLM)
- A model trained on large volumes of text that predicts which words follow a given input. It is what enables interpreting and writing in natural language, but on its own it knows nothing about a company's internal information.
- Natural language processing (NLP)
- The field concerned with getting a system to interpret human language as it is actually written or spoken, with abbreviations, typos, and regional variants. It is what allows understanding the intent behind an informal message.
- Computer vision
- Techniques that let a system interpret images and video: identifying objects, reading a licence plate, detecting a defect in a part, or classifying an anomaly on a production line.
RAG architecture and models
- RAG
- Retrieval-augmented generation. An architecture that connects a language model to a company's actual documentation: before answering, the system retrieves the relevant fragments and answers only from that basis, citing where each fact came from. Read more about this →
- Embedding
- A numerical representation of a text's meaning. Two fragments about the same thing have similar embeddings even without sharing words, which is what makes it possible to search by meaning rather than by exact match.
- Chunking
- Splitting a document into fragments that stand on their own before indexing it. A bad split separates a condition from its exception and makes the system answer with half the information, so it largely determines the final quality.
- Vector database
- A database that stores embeddings and retrieves the fragments most similar to a query. It is the piece that makes searching by meaning across thousands of documents viable within an acceptable response time.
- Reranking
- A second ordering step over the retrieved fragments. Similarity search returns mathematically close candidates; reranking reorders them by actual relevance to the question, which is not always the same thing.
- Fine-tuning
- Retraining a model with your own examples so it adopts a specific way of answering. It is useful for teaching style or format, not for adding information that changes: a RAG architecture fits that better.
- Hallucination
- A plausible but incorrect answer a model produces when it lacks information. It is not a random error: it follows from the model predicting likely text rather than verifying. It is mitigated by connecting it to your own sources and requiring citations.
- Context window
- The maximum amount of text a model can consider at once, including the question, the retrieved documents, and its own answer. Once exceeded, the oldest information is no longer available for that query.
- Prompt
- The instruction a model receives, including the user's request and the rules defining how it should behave. In an enterprise system the end user doesn't write it: it is part of the design and sets the limits of what the agent may do.
Automation and integration
- Process automation
- Redesigning a workflow so a system handles the repetitive tasks and people step in only for exceptions. It requires the process to be explainable and the information to be accessible in some system. Read more about this →
- RPA
- Robotic process automation. It repeats a predefined sequence of steps over other systems' interfaces. It is precise and cheap as long as nothing changes: move a button or add a field and it breaks and must be rebuilt.
- OCR
- Optical character recognition: extracting text from an image or scanned PDF. AI-powered OCR additionally interprets the document's structure and recognises which field each value is, without per-supplier templates. Read more about this →
- API
- An interface that lets two systems exchange information programmatically. Whether a system has a documented API is what separates a predictable integration from a project of uncertain duration. Read more about this →
- Systems integration
- Connecting the tools a company already uses so a record is entered once and available everywhere. Before connecting them, you have to define which system is the source of truth for each field. Read more about this →
- Edge computing
- Processing data at the place where it is generated instead of sending it to the cloud. It lets a capture point keep working offline and sync afterwards without losing records.
Running it in production
- Confidence threshold
- The value above which the system accepts a result as valid. Anything below goes to human review. Raising it means more oversight and fewer errors; lowering it, the opposite. It is a business decision, not a technical one.
- Human in the loop
- A design where certain actions require a person's approval before being executed. The practical rule is that anything with outward consequences — money, commitments, customer communications — goes through someone.
- Traceability
- A record of what the system did, with which data, and when. Without traceability no organization can grant an agent permissions over its systems, because an error cannot be detected or reconstructed.
- Structured data
- Information organized into identifiable fields, such as date, amount, or supplier, rather than free text. Turning informal text into structured data is what makes it possible to query, control, and aggregate it.
Tell us which process you'd like to improve
We'll look at how it works today and send back a concrete diagnostic: what to automate, how, and what result to expect. No cost, no commitment.