The A-Z of AI: every conversational AI term you need to know in 2026

The conversational AI space moves fast, and so does its vocabulary. Whether you are a CX leader evaluating vendors or an engineer implementing solutions, this glossary covers every term you will encounter. Bookmark it. We keep it updated.

A

Agentic AI: AI systems that act autonomously toward goals, reasoning through multi-step problems and taking real actions in connected systems such as issuing refunds, updating orders, or escalating tickets. Unlike traditional chatbots, agentic AI does not follow pre-written scripts.

Agent Handoff: The process of transferring a conversation from an AI agent to a human agent, including passing full conversation context so the customer does not have to repeat themselves.

API (Application Programming Interface): A set of rules that allows software systems to communicate. AI agents use APIs to read and write data in CRMs, e-commerce platforms, helpdesks, and other tools.

Auto-tagging: Automatically categorising conversations by topic, sentiment, or intent without human intervention. Useful for reporting and routing.

Average Handle Time (AHT): The average duration to resolve a customer interaction, including conversation time and any post-conversation work. AI typically reduces AHT by 40-60%.

B

Bot: A broad term for any software that automates tasks. In CX, typically refers to rule-based chatbots that follow decision trees. Increasingly replaced by AI agents.

Brand Voice: The consistent tone, style, and personality an AI agent uses when communicating. Enterprise platforms allow configuring brand voice through system prompts.

Bulk Resolution: Handling multiple related customer queries simultaneously, for example notifying all affected customers during a service outage.

C

Canned Response: A pre-written reply template. Traditional chatbots rely on these; agentic AI generates contextual responses dynamically.

Chain-of-Thought (CoT): A prompting technique where the AI reasons through a problem step by step before giving an answer. Improves accuracy on complex queries.

Channel: A communication medium through which customers interact: web chat, WhatsApp, email, SMS, voice, social media, or in-app messaging.

Chunking: Breaking large documents into smaller pieces for efficient retrieval in RAG systems. Chunk size and overlap affect retrieval quality.

Classification: Sorting customer messages into predefined categories such as billing, shipping, or technical support. A foundational capability for routing and analytics.

Containment Rate: The percentage of conversations handled entirely by AI without human intervention. A key metric, but can be misleading if the AI deflects rather than resolves.

Context Window: The maximum amount of text (measured in tokens) an AI model can process in a single interaction. Larger context windows allow the model to reference more conversation history and documents.

Conversational AI: AI technology that enables natural, human-like dialogue between machines and people across text and voice channels.

Conversational Commerce: Using chat-based interactions to guide customers through product discovery, recommendations, and purchases.

CSAT (Customer Satisfaction Score): A survey-based metric, typically a 1-5 scale, measuring customer satisfaction immediately after an interaction.

Customer Effort Score (CES): A metric measuring how easy it was for a customer to get their issue resolved. Lower effort correlates with higher loyalty.

D

Data Grounding: Anchoring AI responses in verified, factual data sources rather than relying on the model's training data alone. A core strategy for reducing hallucinations.

Decision Tree: A flowchart-based logic structure used in rule-based chatbots. Each node presents a choice that routes the conversation. Limited in flexibility compared to AI agents.

Deflection: Redirecting a customer away from a human agent, often to a self-service resource. High deflection with low resolution indicates poor AI performance.

DPA (Data Processing Agreement): A legal contract between a data controller and processor that governs how personal data is handled. Required under GDPR.

E

Embeddings: Numerical representations of text that capture semantic meaning. Used in RAG systems to find documents similar to a customer's question, even when the exact words differ.

Escalation: Transferring a conversation from AI to a human agent, typically triggered by complexity, sentiment, or explicit customer request.

Entity Extraction: Identifying specific pieces of information (order numbers, dates, product names) from unstructured customer messages.

F

Failover: Automatic switching to a backup AI model or system when the primary one becomes unavailable. Critical for maintaining uptime in customer-facing AI deployments.

Few-shot Learning: Providing the AI model with a small number of examples to guide its behaviour on a specific task. More flexible than fine-tuning, less resource-intensive.

Fine-tuning: Adapting a pre-trained model with domain-specific data to improve performance on particular tasks. Increasingly replaced by prompt engineering and RAG for most CX use cases.

First Contact Resolution (FCR): The percentage of customer issues resolved during the initial interaction without requiring follow-up. A premium support metric.

Flow Builder: A visual tool for designing conversation paths and automation logic. Used in both rule-based and hybrid AI systems.

G

Generative AI: AI that creates new content (text, images, code) rather than selecting from pre-existing options. Powers natural-sounding AI agent responses.

Grounding: See Data Grounding.

Guardrails: Rules and constraints placed on AI agents to prevent unwanted behaviours such as discussing competitors, sharing incorrect pricing, or making promises the company cannot keep.

H

Hallucination: When an AI model generates plausible-sounding but factually incorrect information. Mitigated by grounding responses in verified knowledge bases and implementing guardrails.

Human-in-the-Loop (HITL): A system design where humans review, approve, or correct AI outputs. Common during initial deployment and for high-stakes decisions.

Hybrid AI: An architecture combining rule-based logic with generative AI. Rules handle predictable flows; AI handles open-ended conversation.

I

Inference: The process of running input data through a trained AI model to generate a prediction or response. Inference speed and cost are key factors in production AI.

Intent Detection: Identifying the underlying goal or purpose behind a customer's message. For example, 'Where is my order?' maps to an order-tracking intent.

Integration: Connecting the AI platform to external systems (CRM, e-commerce, helpdesk) so the agent can read data, take actions, and automate workflows.

J

JSON (JavaScript Object Notation): A lightweight data format used for structured communication between AI systems and APIs. AI agents parse and generate JSON to interact with external tools.

K

Knowledge Base: A structured repository of information (FAQs, product docs, policies) that an AI agent references to answer customer questions accurately.

KPI (Key Performance Indicator): A measurable value that demonstrates how effectively an AI deployment achieves its objectives. Common AI CX KPIs include resolution rate, CSAT, AHT, and cost per conversation.

L

Latency: The time delay between a customer sending a message and receiving a response. For conversational AI, sub-second latency is the goal.

LLM (Large Language Model): The foundation AI model that powers conversational AI. Trained on vast text datasets to understand and generate human language. Examples include Claude, GPT, and Gemini.

Low-code / No-code: Platforms that allow non-technical users to build and configure AI agents through visual interfaces rather than writing code.

M

MCP (Model Context Protocol): A standard protocol that enables AI agents to discover and use external tools dynamically at runtime. Think of it as USB-C for AI integrations: plug in a tool, and the agent knows how to use it.

Model Router: A system that directs conversations to different AI models based on query complexity, language, or cost optimisation rules.

Multi-model Architecture: An AI platform that supports multiple AI model providers simultaneously, enabling failover, cost optimisation, and best-model routing per conversation type.

Multi-turn Conversation: A dialogue that spans multiple exchanges. The AI must maintain context across turns to deliver coherent, relevant responses.

N

Natural Language Processing (NLP): The branch of AI focused on enabling computers to understand, interpret, and generate human language.

Natural Language Understanding (NLU): A subset of NLP focused specifically on comprehending the meaning and intent behind human language input.

NPS (Net Promoter Score): A loyalty metric based on the question: How likely are you to recommend us to a friend or colleague? Scored from -100 to +100.

O

Omnichannel: Supporting customers across multiple communication channels (web, WhatsApp, email, voice, social) from a unified platform, with shared conversation history.

Orchestration: Coordinating multiple AI agents, tools, and workflows to handle complex customer requests that span multiple systems.

P

Personalisation: Tailoring AI responses based on customer data such as purchase history, preferences, location, and past interactions.

Playground: A testing environment where teams can experiment with prompts, model settings, and conversation flows before deploying to production.

Prompt: The instruction or input given to an AI model that shapes its response. Prompt quality directly impacts output quality.

Prompt Engineering: The practice of crafting and optimising prompts to get the best results from AI models. A core skill for AI agent configuration.

Proactive Support: Reaching out to customers before they contact you, based on detected events like delivery delays, payment failures, or subscription renewals.

Q

Query Classification: See Classification.

Queue Management: Routing and prioritising conversations in a support queue, including AI triage to determine urgency and optimal routing.

R

RAG (Retrieval-Augmented Generation): An architecture that retrieves relevant documents or data before generating a response, grounding the AI in factual, up-to-date information. The primary technique for reducing hallucinations.

Rate Limiting: Controlling how many API requests or conversations an AI system processes per time period to manage costs and prevent abuse.

Resolution Rate: The percentage of conversations where the customer's issue was actually solved, not just contained or deflected. The gold-standard metric for AI agent performance.

Responsible AI: Designing and deploying AI systems that are fair, transparent, accountable, and aligned with ethical principles.

Routing: Directing incoming conversations to the right destination: a specific AI agent, a human team, or a specialised workflow based on topic, language, or priority.

S

Semantic Search: Search that understands meaning and context rather than just matching keywords. Uses embeddings to find relevant results even when wording differs.

Sentiment Analysis: Detecting the emotional tone (positive, negative, neutral, frustrated) of customer messages in real time. Used for routing, escalation triggers, and analytics.

SLA (Service Level Agreement): Contractual commitments on performance metrics such as uptime percentage, response time, and resolution time.

Slot Filling: Collecting required pieces of information from a customer during a conversation, such as order number, email address, or issue description.

System Prompt: The foundational instruction set that defines an AI agent's behaviour, personality, knowledge boundaries, and guardrails. Not visible to customers.

T

Temperature: A model parameter (0.0-2.0) that controls response randomness. Lower values produce more deterministic, focused answers; higher values increase creativity and variation.

Token: The basic unit of text processing for AI models. Words, punctuation, and subwords are all tokens. Pricing is often calculated per token processed.

Tool Use (Function Calling): The ability of an AI agent to invoke external functions or APIs during a conversation, for example looking up an order, processing a refund, or creating a ticket.

Training Data: The dataset used to train an AI model. For CX applications, this includes conversation logs, product information, and support documentation.

Transfer Learning: Applying knowledge gained from training on one task to a different but related task. The foundation of how pre-trained LLMs adapt to specific domains.

U

Utterance: A single message or statement from a user in a conversation. The building block of conversational AI training data.

Uptime: The percentage of time an AI system is operational and available. Enterprise SLAs typically guarantee 99.9% or higher uptime.

V

Vector Database: A database optimised for storing and querying embeddings (vector representations of text). Powers semantic search and RAG retrieval.

Voice AI: Conversational AI deployed on voice channels (phone, IVR, smart speakers). Requires additional speech-to-text and text-to-speech capabilities.

W

Webhook: An HTTP callback that sends real-time notifications between systems when specific events occur. AI platforms use webhooks to trigger actions in external tools.

Workflow Automation: Using AI to automate multi-step business processes, such as processing a return (verify order, check policy, issue refund, send confirmation).

X-Y

XAI (Explainable AI): AI systems designed to provide understandable explanations of how they reached a decision or response. Important for trust and compliance.

Z

Zero-shot Learning: An AI model's ability to handle tasks or queries it was not specifically trained on, relying on general knowledge and reasoning ability.

This glossary is a living document. As the conversational AI field evolves, so will these definitions. Bookmark this page and check back regularly.