The AI glossary.
Vendor-neutral. Plain English.

AI systems that can take autonomous actions in real systems — reading, writing, deciding, and orchestrating multi-step tasks — rather than only producing text. The dominant frontier of AI work in 2026. An AI receptionist that books an appointment is agentic; a chatbot that only answers questions is not.

A software entity built on a large language model that can perceive its environment (read data), reason about it, and take actions on it. The general category that includes AI receptionists, AI SDRs, AI voice agents and AI assistants.

Outbound calling campaigns conducted by AI rather than human SDRs. In the UK, must comply with PECR, TPS/CTPS screening and ICO guidance. See is AI cold calling legal in the UK.

The set of legal, technical and operational requirements that an AI system must meet to handle personal data lawfully in the UK or EU. GDPR governs personal-data processing; PECR governs electronic direct marketing. Both apply to most AI customer-engagement deployments.

The coordination of multiple AI models, tools, data sources and human checkpoints into an end-to-end workflow. Where prompts produce single responses, orchestration produces complete business processes.

A voice-only AI agent that holds real-time conversations on the phone. Inbound, outbound or both. Narrower in scope than an AI receptionist (which is multi-channel).

A multi-channel AI agent that answers phone, WhatsApp, SMS and web chat — qualifying, booking, e-signing, escalating. The most-searched AI category in the UK in 2026.

The discipline of building AI systems that behave reliably, transparently and within explicit boundaries. For business deployments, AI safety includes guardrails against hallucination, prompt injection, data leakage and unauthorised actions.

An AI agent acting as a sales development representative — running outbound prospecting across phone, email and LinkedIn. The same compliance bar as human SDRs applies (PECR, GDPR, TPS in the UK).

An AI system that conducts spoken conversations — inbound or outbound, on the phone or in an app. The voice layer typically combines speech-to-text (STT), a language model and text-to-speech (TTS) into a single real-time loop.

The use of software to run end-to-end business processes that previously required manual handling. Broader than workflow automation; broader than RPA. The orchestration layer that runs operations.

A software agent that holds text-based conversations, typically on a website or in a messaging app. Distinct from an AI receptionist (multi-channel including voice) and an AI agent (action-taking, not just conversational).

The umbrella term for AI systems designed to hold natural multi-turn conversations — covering chatbots, voice agents, AI receptionists and AI assistants. Distinct from purely generative AI in that conversational AI maintains context across turns.

The automation of customer-relationship-management workflows — lead scoring, follow-up sequences, contact enrichment, opportunity routing — using a combination of rule-based triggers and AI-augmented decisions.

Numerical representations of text (or images, audio) that capture semantic meaning. The mathematical primitive that makes vector databases, semantic search and retrieval-augmented generation (RAG) work.

The process of further training a pre-trained large language model on a smaller, task-specific dataset. Used to specialise a general model for a particular domain, tone or capability. Distinct from RAG, which adds context at query time rather than retraining.

A capability of modern LLMs that lets the model call structured external functions during a conversation — looking up calendar availability, writing to a CRM, processing a payment. Function calling is what turns a chatbot into an agent.

A shorthand for frontier LLMs at the capability tier of OpenAI's GPT-4 — including Claude Opus, Gemini Pro and their successors. These are the models capable of running production-grade agentic workloads.

An LLM output that is confidently wrong — plausibly worded but factually incorrect. Mitigated in production by retrieval-augmented generation (RAG), grounding, constrained output formats and human review of high-stakes outputs.

The task of identifying what a user actually wants from their message — "book an appointment", "ask a question", "complain". The first step in most conversational-AI pipelines, and the foundation of routing and escalation.

A structured repository of an organisation's information — policies, FAQs, product details, procedures — that an AI agent consults during conversations. Typically wired to the LLM via retrieval-augmented generation (RAG).

A neural network trained on vast quantities of text to predict the next token in a sequence. The category of models underlying ChatGPT, Claude, Gemini and the agentic AI built on top of them.

An open standard introduced in 2024 by Anthropic for connecting AI models to external tools, data sources and actions in a structured, secure way. MCP-compatible agents can be wired to new data sources by configuration rather than custom code.

AI models that handle multiple input or output modalities — text, voice, image, video — in a single system. Modern voice agents are multimodal in that they accept audio input and produce audio output via integrated text and audio models.

The craft of designing the instructions given to an LLM to elicit reliable, high-quality outputs. A meaningful skill in 2026, though decreasingly the bottleneck as frontier models improve and orchestration frameworks abstract prompts away.

A pattern that retrieves relevant documents from a knowledge base and supplies them to the LLM as context before it generates a response. The dominant way to ground LLMs in current, organisation-specific information without retraining.

A training technique where humans rank model outputs and a reinforcement-learning algorithm tunes the model to produce more of the highly-ranked outputs. The technique that made ChatGPT conversational rather than only generative.

Software bots that mimic a human clicking through user interfaces. Used to automate legacy systems that have no API. Distinct from BPA, which orchestrates across multiple systems and includes AI.

The conversion of spoken audio into written text. The first stage of any voice-AI pipeline. Production-grade providers include Deepgram, Whisper (OpenAI) and Google Speech-to-Text.

The conversion of written text into spoken audio. The final stage of a voice-AI pipeline. Production-grade providers include ElevenLabs, OpenAI Realtime and Google Cloud TTS. Modern TTS is largely indistinguishable from human voice in blind tests.

The Telephone Preference Service (consumer) and Corporate TPS (business) — UK opt-out registers that AI cold-calling deployments must screen against before dialling. Calling a registered number without specific opt-in consent is a PECR breach.

A database optimised for storing and querying embeddings. The infrastructure layer that makes retrieval-augmented generation (RAG) practical. Pinecone, Weaviate, Qdrant, pgvector are common choices in 2026.

The generation of a synthetic voice that sounds like a specific human speaker. Used in branded AI receptionists where a known voice talent has licensed their voice to the brand. Subject to consent and likeness laws in most jurisdictions.

Meta's API for businesses to send and receive WhatsApp messages programmatically. The foundation layer for any AI WhatsApp chatbot. Requires business verification and a Meta-approved BSP (business solution provider).

A single trigger-action chain — "when X happens, do Y". The smallest unit of automation. Tools include Zapier, Make and n8n in basic mode. Subset of BPA.