AI terms, explained plainly.
No jargon defined with more jargon. What each term means, and why it matters when you are deciding something.
Terms worth knowing
No jargon defined with more jargon. What each term means, and why it matters when you are deciding something.
AI Agent
Software given a model, tools and permission to act - read a file, update a record, place a call - rather than only producing text. The difference from a chatbot is that it changes something.
AI Act
The EU regulation on artificial intelligence, in force across the Union with obligations phasing in through 2026 and 2028. It classifies systems by risk and attaches duties to each class.
AI Assessment
A structured review of processes to find where AI pays and where it does not, with estimated return per case. The point is the ranking, not the list.
AI Literacy
Knowing what these systems can and cannot do, where they fail, and how to check their output. Confidence without it is worse than no adoption.
API
The interface one piece of software exposes so another can use it. Whether your existing systems have usable ones largely determines what can be automated.
ASR
Automatic speech recognition: turning spoken audio into text. The first step in any voice agent, and where difficult audio degrades quality first.
Automation
Software doing a task a person used to do. Classic automation follows fixed rules; AI automation copes with variation, which is why it reaches processes rules could not.
Chain of Thought
Prompting a model to work through intermediate steps before answering. It improves reasoning on multi-step problems and makes the answer auditable.
Chatbot
A conversational interface. Older ones follow scripted trees; modern ones run on language models. Neither acts on your systems unless built as an agent.
Claude
A family of language models by Anthropic, strong on long documents and careful instruction-following. One of the options we benchmark, not a default.
Text completion
The underlying operation of a language model: predicting what comes next. Everything else - chat, agents, summarising - is built on top of it.
Context window
How much text a model can consider at once. Exceed it and earlier content is dropped, which is the usual cause of a model "forgetting" something you said.
Deep learning
Machine learning using many-layered neural networks. The technique behind current language and vision models.
Embedding
Text converted into a list of numbers that captures meaning, so similar texts sit close together. This is what makes semantic search possible.
Few-shot learning
Giving a model a handful of examples in the prompt to establish the pattern you want. Often more effective than a long instruction, and far cheaper than fine-tuning.
Fine-tuning
Further training on your own examples to fix tone or format. Usually unnecessary: RAG solves more problems for less money and without locking your data into a model.
Foundation model
A large general-purpose model trained on broad data, adapted afterwards to specific uses. The base layer everything else is built on.
Function calling
A model's ability to invoke code with structured arguments - look up an order, book a slot - instead of only writing about it. The mechanism that turns a model into an agent.
GDPR
The EU data protection regulation. For AI it governs what data may be processed, on what basis, and what rights people retain over it.
Gemini
Google's family of language models. Strong multimodal capability; another option we benchmark per case.
GPT
OpenAI's family of language models. Widely used, and the reason "GPT" is often used loosely to mean any language model.
Hallucination
A confident answer that is simply wrong. The reason serious deployments cite sources and keep a person on high-impact decisions.
Human-in-the-loop
A design where a person approves or corrects specific decisions. Where you place those gates determines both the risk and the value of an automation.
Inference
Running a trained model to get an answer, as opposed to training it. This is what you pay for per use, and what latency measures.
LangChain
An open-source framework for building applications on language models. Useful for prototyping; production systems often outgrow it.
LlamaIndex
A framework for connecting language models to your own data, largely for RAG. Solves indexing and retrieval so you do not build them yourself.
LLM
A large language model: software trained on text that predicts what comes next. Useful for language work, unreliable for arithmetic and for facts it was never shown.
MCP
Model Context Protocol: an open standard for connecting models to tools and data sources, so integrations are not rebuilt per model.
Machine learning
Software that derives its behaviour from data rather than from rules someone wrote. The broader field language models belong to.
Long-term memory
Storing information across conversations so a system recalls earlier interactions. Distinct from the context window, which resets each time.
Mistral
A family of models from the French company of the same name, including open-weight ones that can be self-hosted. Relevant when data residency rules out hosted APIs.
Model-agnostic
An architecture that separates what your system does from which model runs it, so providers can be swapped without a rebuild. It is what prevents lock-in.
Multi-agent system
Several agents with distinct roles working on one task. Powerful, and considerably harder to debug - worth it only when a single agent genuinely cannot cope.
Multimodal
A model handling more than text: images, audio, video. Useful for reading scanned documents and photographs of damage or defects.
N8N
An open-source workflow automation tool. Often the right place to orchestrate an agent's steps when the logic is mostly sequencing.
NLP
Natural language processing: the field concerned with software handling human language. Language models are its current state of the art.
NLU
Natural language understanding: extracting intent and meaning from what someone said. What decides whether a voice agent understood the caller.
Orchestration
Coordinating models, tools and steps into a working process. Most of the engineering in a real deployment is here, not in the model.
OpenAI
The company behind the GPT models. One provider among several we benchmark; being tied to none is deliberate.
Prompt
The instruction given to a model. Small changes in wording produce large changes in output, which is why prompts are tested rather than written once.
Prompt engineering
Writing and testing prompts systematically so results hold up across real inputs, not just on the examples you tried first.
RAG
Retrieval-augmented generation: the model answers from your documents rather than from memory. This is what makes an answer checkable against a source.
ROI
Return on investment. For AI, it should account for build cost, running cost and your team's time - not only the licence.
RPA
Robotic process automation: software following fixed rules through existing interfaces. It breaks on variation, which is exactly where AI agents take over.
System prompt
The standing instruction that defines a model's role and boundaries for every conversation. Where behaviour and limits are set.
Temperature
A setting controlling how varied a model's output is. Low for extraction and classification, higher for drafting - a wrong value here causes a lot of apparent unreliability.
Token
The unit models read and bill in, roughly three quarters of a word. Costs and limits are expressed in tokens, which is why the term appears on invoices.
Tool use
A model calling external functions to look things up or take action. The general term for what function calling enables.
TTS
Text-to-speech: turning text into spoken audio. What gives a voice agent its voice, and where naturalness is most noticeable.
Vector
The numeric representation of a text's meaning, stored in a vector database so similar content can be found by meaning rather than by keyword.
Voice AI Agent
An agent operating over the phone: it understands speech, answers, acts on your systems, and escalates what needs a person.
Zero-shot learning
A model performing a task with no examples, from the instruction alone. Works well for common tasks, less so for anything specific to your business.
Terms that mislead
Describes everything and commits to nothing. Ask which model, on what data, and what happens when it is wrong.
Rarely means unsupervised. Ask where the approval gates are, and whether you can move them.
Can mean fine-tuning, or simply RAG. The difference matters for cost, for lock-in, and for what happens to your data.
Meaningless without the test set. Accurate on what, measured how, compared to which baseline.