AI Glossary: Key Terms for Conversational AI, Agents and Voicebots

A comprehensive reference for business and product teams navigating the world of artificial intelligence. From foundational concepts to agent architectures, voice AI and business metrics — every term you need to understand and deploy AI in your organization.

1. Fundamental concepts

Artificial General Intelligence (AGI) AGI refers to hypothetical AI systems that can perform a wide range of cognitive tasks at least as well as an average human, across many domains, not just one. It usually implies broad reasoning, adaptation and autonomy, rather than being limited to a single use case like translation or image recognition.

Artificial Intelligence (AI) AI is the broad field of building computer systems that can perform tasks which typically require human intelligence, such as understanding language, recognizing patterns, making predictions or taking decisions. In business, AI usually refers to applied machine learning models embedded into products, workflows or services.

Machine Learning (ML) Machine learning is a subset of AI where models learn patterns from data rather than following manually coded rules. Instead of if/then logic, ML systems adjust their internal parameters to improve performance on tasks like classification, prediction or recommendation.

Deep Learning Deep learning is a subset of machine learning that uses multi-layer neural networks to learn complex patterns from large volumes of data. These models automatically discover useful features (e.g. shapes in images, patterns in audio, structures in text) instead of relying on hand-crafted rules. Deep learning powers most modern speech recognition, image generation and large language models.

Neural Network A neural network is a layered mathematical structure inspired (loosely) by the brain, made of interconnected "neurons" that transform input data step by step. Each connection has a weight that determines how strongly one unit influences another; training adjusts these weights so the network produces better outputs over time.

2. Large models, tokens and training

Large Language Model (LLM) A large language model is a deep neural network trained on vast amounts of text to predict the next token in a sequence. In practice, LLMs can chat, summarize, translate, write code and act as the reasoning core of many AI products. "GPT", "Claude", "Gemini", "Llama" or "Mistral" are LLM families; "ChatGPT", "Copilot" or "Le Chat" are assistant products built on top of them.

Tokens Tokens are the basic units of text a model processes, such as words, sub-words or characters. Billing, context limits and usage metrics are usually expressed in tokens (not characters). In practice, 100-150 tokens correspond roughly to 75-100 words of English.

Weights Weights are the numerical parameters inside a model that define how much importance is given to different features of the input. During training, the learning algorithm iteratively adjusts these weights to reduce the gap between the model's prediction and the desired output.

Training Training is the process of teaching a model to perform a task by exposing it to data and adjusting its weights to reduce errors. It is compute-intensive, often requires very large datasets, and is typically done once or a few times at scale before deployment.

Inference Inference is the process of running a trained model to generate outputs (predictions, text, audio, decisions) for new inputs. From a business perspective, training is a fixed, capital-intensive cost, while inference is the recurring cost that grows with usage.

Compute "Compute" refers to the computational power required to train and run AI models, usually provided by GPUs, TPUs or specialized accelerators. For modern LLMs, compute capacity is often the main bottleneck driving costs, latency and scalability.

Memory cache / KV caching Caching reuses intermediate computations so the model does not need to recompute everything at each step. In transformer models, key-value (KV) caching stores past tokens' representations so generating the next tokens becomes faster and cheaper. This is a key optimization for real-time and high-traffic applications.

Context window The context window is the maximum amount of text (tokens) the model can consider in a single request, including prompt, system instructions and previous messages. A larger context window allows the model to keep longer conversations, process longer documents or handle more complex multi-step tasks.

Cost per token Cost per token is the unit price AI providers use to bill model usage. It often differs for input tokens and output tokens, and may also differ for "reasoning tokens" in advanced models. Optimizing prompts, responses and caching strategies can significantly reduce total cost per token for a product.

Reasoning model A category of LLM (like o1, o3 or Gemini Thinking) designed to work through explicit, step-by-step reasoning before producing its final answer — rather than answering directly. These models are generally slower and more expensive to run, but more reliable on complex logic, math or planning tasks.

Small Language Model (SLM) A language model significantly lighter than a typical LLM, often deployable locally or on edge devices rather than via a cloud API. SLMs trade off some of a large model's general capabilities for much lower inference cost and latency — relevant at scale or for specialized use cases.

3. Learning techniques and optimization

Fine-tuning Fine-tuning means continuing the training of an existing model on a more specific dataset to adapt it to a domain or task (e.g. customer support for veterinarians, internal knowledge of a company). It typically improves accuracy and tone in that niche while reusing all the general capabilities learned before.

Transfer learning Transfer learning uses a model trained on one task as the starting point for another, related task, reusing learned representations instead of training from scratch. Fine-tuning is a common form of transfer learning applied to large foundation models.

Distillation Distillation is a "teacher-student" technique where a smaller model learns to imitate a larger one by training on the larger model's outputs. The goal is to keep most of the quality while reducing size, latency and cost, which is crucial for edge devices or high-volume inference.

Quantization A model compression technique that reduces the numerical precision of its weights (e.g. from 32-bit to 8-bit or 4-bit) to shrink memory footprint and speed up inference, at the cost of a slight quality loss. Complementary to distillation: the two techniques are often combined to deploy capable models at lower cost.

4. Generative AI, diffusion and GANs

Generative AI (GenAI) Generative AI refers to models that can create new content: text, images, audio, video, code or 3D assets. They do not just classify inputs; they produce original outputs that follow patterns learned from training data.

Diffusion model A diffusion model gradually adds noise to training data and then learns to reverse this process, denoising random noise back into coherent images, audio or other media. This "reverse diffusion" is behind many state-of-the-art image and video generation systems.

GAN (Generative Adversarial Network) A GAN uses two neural networks: a generator that produces synthetic data and a discriminator that tries to distinguish real from fake. By competing, both networks improve, leading to highly realistic images, videos or audio. GANs are widely used for deepfakes and realistic media synthesis.

Hallucination A hallucination occurs when a model produces confident but incorrect or fabricated information. Hallucinations are inherent to current LLMs, especially on topics not well covered in their training or when prompts are ambiguous. Product teams mitigate them via retrieval (RAG), constraints and better evaluation.

5. Embeddings, RAG and vector databases

Embedding An embedding is a numerical vector representation of text, audio, images or other data that captures semantic meaning. Similar content ends up with similar vectors. Embeddings are the backbone of semantic search, recommendation, clustering and retrieval-augmented generation.

Vector database A vector database is optimized to store and search embeddings efficiently using similarity metrics (e.g. cosine similarity). It enables fast "find me the most similar documents" queries, which are essential for RAG systems, recommendation engines and personalization.

RAG (Retrieval-Augmented Generation) Retrieval-augmented generation combines a generative model with an external knowledge base. Before answering, the system retrieves relevant documents (using embeddings and a vector database) and feeds them into the model so it can ground its answer in up-to-date or private data. RAG is critical for enterprise use cases where accuracy and freshness matter.

6. Agents, tools and automation

AI Agent An AI agent is a system that not only generates text but can plan, decide and take actions to achieve a goal, often across multiple steps and tools. Unlike a simple chatbot, an agent can call APIs, interact with databases, update CRMs, schedule meetings or launch workflows autonomously under constraints defined by the business.

Agentic AI Agentic AI refers to architectures that give models structured autonomy: the ability to set sub-goals, choose tools, monitor progress and adapt plans in real time. In practice, this means AI systems that behave more like digital coworkers than static chat interfaces, operating inside defined safety and compliance boundaries.

Tool use / Function calling Tool use is the ability of a model to call external functions (APIs, internal services, databases) from within a conversation. The model decides when and how to call a tool (e.g. "create_lead", "book_meeting", "check_inventory"), receives the result, and then continues the interaction with up-to-date information or completed actions.

Multi-agent system A multi-agent system orchestrates several specialized AI agents that collaborate to complete a process. For example, one agent qualifies a lead, another negotiates a meeting time, and a third handles post-call follow-up and CRM updates. This mirrors human teams and is powerful for complex workflows.

Agentic Loop The observe → decide → act → verify cycle an agent repeats autonomously until it reaches its goal. This looping mechanism is what distinguishes an AI agent from a simple question-answering system: the agent evaluates the outcome of its action and adjusts its next step accordingly. Source

Orchestration The control layer that decides which agent runs, in what order, and with what data, within a multi-agent system. The orchestrator doesn't perform the task itself: it coordinates the specialized agents that do. Source

Context Engineering The discipline of precisely selecting which information is fed into an agent's context window at each step of its reasoning, rather than providing everything at once. Framed by parts of the industry as having superseded simple prompt engineering in 2026 for complex agentic systems. Source

Agent Card / Skill Within the A2A protocol ecosystem, an Agent Card is a structured description of the capabilities an agent exposes to other agents: which tasks it can perform, with which input and output formats. It plays a role similar to API documentation, but designed to be read by other agents.

Sub-agent / Handoff / Agent Mesh Three task-delegation mechanisms in a multi-agent system: a sub-agent is a specialized agent invoked by a primary agent; a handoff is the transfer of a conversation or context from one agent to another; an agent mesh refers to the overall architecture connecting multiple specialized agents together.

Grounding The act of connecting an LLM's responses to real, verified data to reduce the risk of hallucination. RAG is one grounding technique among others — grounding is the general concept, RAG a specific implementation.

Chunking Splitting a document into segments (chunks) before indexing it in a vector database. Chunk size and splitting logic directly affect the quality of a RAG system's results: chunks that are too large dilute meaning, chunks that are too small lose context.

OAuth / BYOK OAuth is the standard authorization protocol that lets an agent access a third-party service with the user's consent, without ever seeing their password. BYOK (Bring Your Own Key) refers to the practice of supplying your own API key to an AI tool rather than using shared access — useful for cost control and compliance.

Agent governance and security

Guardrails Hard-coded constraints — not simple prompt instructions — that block certain dangerous actions by an AI agent: spend caps, allow-lists of domains, restricted write permissions. Unlike a text instruction, a guardrail cannot be bypassed by a clever rephrasing. Source

Human in the loop (HITL) A checkpoint in an agentic workflow where explicit human validation is required before an agent executes a sensitive action — sending a payment, deleting data, external communication. HITL remains the reference safeguard for high-impact actions. Source

Kill switch An emergency stop mechanism that immediately cuts off an agent's execution if abnormal or dangerous behavior is detected. Source

Prompt injection An attack that hides malicious instructions inside content an agent processes (an email, a web page, a document) to make it execute actions not intended by its operator. It is one of the main vulnerabilities of agentic systems connected to external sources. Source

Sandboxing Isolating an agent's execution environment — restricted network access, a walled-off file system — to limit potential damage in case of unexpected behavior or a successful attack. Source

Trust Level / Least Privilege A security principle whereby an AI agent is granted only the access and permissions strictly necessary for its task, never more. An agent handling customer questions, for instance, has no reason to have write access to the billing database. Source

Workflow automation (AI) AI-driven workflow automation uses models and agents to execute end-to-end business processes: from inbound lead capture to qualification, appointment scheduling, routing to the right team and CRM enrichment. Instead of just answering questions, the system actually progresses the business process to completion.

Benchmark (AI models) Benchmarking AI models means systematically comparing them on standard tasks (e.g. reasoning, coding, multilinguality, latency, cost) to choose the best fit for a given product. In practice, teams benchmark models from OpenAI, Anthropic, Google, Mistral and others on their real business use cases, not only on public leaderboards.

7. Conversational and voice AI

Conversational AI Conversational AI covers systems that can understand and generate natural language in an interactive way across channels (web chat, WhatsApp, voice, email). Modern conversational AI goes beyond scripted chatbots, using LLMs, memory and tool use to provide more fluid, context-aware experiences.

Chatbot (modern) A chatbot is a conversational interface that interacts with users via text or messaging apps. Modern chatbots can rely on LLMs, RAG and tools to answer questions, guide users and trigger actions. Compared to an AI agent, a chatbot is often more constrained to Q&A and guided flows, with less autonomy over external systems.

Voicebot A voicebot is an AI agent that talks with users over the phone or other voice channels in real time. It handles speech recognition, language understanding, reasoning and speech synthesis under tight latency constraints, enabling use cases like inbound call routing, appointment booking or support triage without a human operator.

Speech-to-Text (STT) Speech-to-text converts spoken audio into written text. It is the first step in most voice AI systems and must be accurate, fast and robust to noise, accents and domain-specific vocabulary.

Text-to-Speech (TTS) Text-to-speech converts written text into natural-sounding audio. Modern TTS can generate expressive, low-latency voices that feel close to human speech, which is critical for customer experience in voicebots and virtual agents.

Latency Latency is the time it takes for a system to respond to a user action. In voice AI, latency must be very low (often under a few hundred milliseconds) to keep conversations natural and avoid people talking over the bot or abandoning the call.

VAD — Voice Activity Detection Detecting the precise moment a user actually starts speaking, distinguishing their voice from background noise or silence. It's a prerequisite step for speech-to-text: without reliable VAD, the system transcribes noise or clips the start of sentences.

Barge-in A voicebot's ability to detect that the user is interrupting it while it's speaking, and to stop immediately to listen. A voicebot without barge-in keeps talking over the user, which instantly breaks the sense of natural conversation.

Turn-taking / Turn detection Managing who speaks when between the human and the AI in a voice conversation — deciding when one has finished speaking and when the other can respond. It's one of the hardest friction points to solve for a voicebot to feel natural on the phone.

Diarization Distinguishing between multiple speakers on the same audio stream. Particularly useful for call centers and multi-party call analysis, where each turn of speech needs to be attributed to the right person.

Wake word A trigger word or phrase that activates a voice assistant's listening (e.g. "Hey Busony"). Lets a system stay idle and only record/process audio after detecting the keyword, for privacy and resource-consumption reasons.

8. Product, business and safety

Prompt engineering Prompt engineering is the practice of designing and structuring inputs to a model to obtain better, more reliable outputs. It includes instructions, examples, constraints, roles and formatting, and is a key lever for improving quality without changing the underlying model.

Few-shot / Zero-shot prompting Two prompt engineering techniques that differ by how many examples are given to the model: zero-shot prompting asks for a task with no examples, relying only on instructions; few-shot prompting supplies a few representative examples in the prompt to guide the format and quality of the expected answer.

Lead qualification (AI) AI-driven lead qualification uses models and agents to assess how valuable an inbound prospect is based on their answers, behavior and context. It can score leads, ask follow-up questions, and decide whether to route them to sales, propose a meeting or handle them via self-service.

AI scheduling AI scheduling refers to agents that automatically find and book meeting slots across calendars, time zones and constraints. Combined with conversational AI, it allows leads or customers to confirm appointments directly in chat or over the phone without human intervention.

Omnichannel AI Omnichannel AI delivers consistent, connected experiences across multiple channels (phone, chat, email, social messaging). The same underlying agent or knowledge base can follow the user from one channel to another, preserving context and history.

AI safety AI safety focuses on preventing AI systems from causing harm, intentionally or unintentionally. It covers topics like misuse prevention, robustness against attacks, bias reduction, and ensuring systems behave within acceptable norms, especially when they are autonomous or high-impact.

Alignment Alignment refers to making AI systems' behavior match human values, goals and constraints. In practice, this means training and governing models so they follow policy, respect regulation, and act in ways that serve users and organizations, not just optimize a technical objective.

RAMageddon "RAMageddon" is an informal term describing the global memory chip shortage driven in part by the AI boom. Large-scale training and inference require massive amounts of RAM, which also affects availability and prices for gaming, consumer electronics and traditional IT.

9. Agentic optimization, GEO & protocols

SEO — Search Engine Optimization Optimization for traditional search engines (Google, Bing). Still fundamental — generative engines and AI agents rely on the same authority and relevance signals as classic SEO. SEO is the foundation layer on which GEO and AEO are built. Our SEO & GEO approach →

GEO — Generative Engine Optimization Optimizing your content to be cited in AI-generated answers from ChatGPT, Perplexity or Gemini. The goal is no longer to appear in a list of links, but to be the synthesized source in the response. GEO relies on content structure (H1/H2/H3, lists, FAQ), Schema.org data, E-E-A-T score and brand consistency across the web. Our GEO service →

AEO — Agentic Engine Optimization (also: Agentic AI Optimization) Optimizing your digital presence to be discovered, cited and selected by autonomous AI agents that complete tasks on users' behalf — purchasing, booking, comparing. This is the third era of search: after SEO (classic search engines) and GEO (generative AIs), AEO prepares your site for the agentic web. It involves exposing structured endpoints, adopting WebMCP and making your site agentically actionable. Our AEO service →

MCP — Model Context Protocol Open standard introduced by Anthropic in November 2024 allowing AI agents to connect to external services in a standardized way. It is the "HTTP protocol" of the agentic web: rather than coding an integration for every "AI × tool" pair, you expose your service once via an MCP server, and any compatible agent can read, write and act on it. Adopted by OpenAI, Google, Automattic (WordPress), Stripe, Supabase and hundreds of others. Learn more →

WebMCP — Web Model Context Protocol Extension of MCP for web browsers, incubated at the W3C. Allows websites to expose structured actions (typed tools) directly to AI agents in the browser, without extensions or server-side integration. WebMCP-compatible sites are "agentically actionable" by Google, Gemini and future web agents — a strong AEO signal. WebMCP guide →

UCP — Universal Commerce Protocol Protocol announced by Google at NRF 2026, co-developed with Shopify, Target and Walmart. Allows any AI agent to discover, compare and complete a purchase without leaving the AI interface. It is the commerce counterpart to WebMCP: an open standard to make catalogs and carts accessible to autonomous shopping agents.

A2A — Agent-to-Agent Protocol Google's protocol allowing AI agents to collaborate directly with each other without human intervention. An "orchestrator" agent can delegate sub-tasks to specialized agents (pricing, logistics, support) via A2A, passing context in a structured way. Foundation of multi-agent agentic commerce.

AP2 — Agent Payments Protocol Agentic payment protocol launched by Google on September 16, 2025, with more than 60 partners (American Express, Mastercard, PayPal, Salesforce). AP2 cryptographically proves that a human actually authorized a purchase carried out by an AI agent, via "mandates" — an Intent Mandate (authorizing the intent) and a Cart Mandate (authorizing the final cart). It complements UCP by securing the last step: payment. Source

ACP — Agentic Commerce Protocol A rival standard to AP2, developed by OpenAI and Stripe and launched on September 29, 2025. It powers Instant Checkout in ChatGPT among other things. Unlike AP2, the merchant remains the "seller of record" in the ACP architecture. Source

Mandate A cryptographically signed digital proof that precisely defines what an AI agent is authorized to buy: scope of authorization, budget, approved suppliers, possible recurrence. It is the central trust mechanism in agentic payment protocols like AP2. Source

x402 An emerging machine-to-machine payment protocol, cited within the same terminology cluster as ACP, UCP and AP2. It aims to enable direct micro-payments between AI agents without human intervention on every transaction. Source

Merchant of record The entity that is legally and fiscally responsible for a transaction, even when it is executed by an AI agent on a customer's behalf. A key compliance concept in agentic commerce: it determines who handles VAT, refunds and legal liability for the sale. Source

AX — Agent Experience The AI agent equivalent of UX (User Experience). AX measures the quality of an agent's interaction with a site or service: data readability, endpoint actionability, reliability and consistency of responses. A high AX score increases the probability that an AI agent will cite or select your brand over a less well-structured competitor.

Citation share The share of AI-generated responses (ChatGPT, Perplexity, Gemini) that actually cite your brand among sources or recommendations, for a set of target queries. It is the reference success metric in GEO, the equivalent of click-through rate in classic SEO.

AI-referred traffic / Dark traffic Traffic generated by users coming from an AI assistant (ChatGPT, Perplexity, Claude) to your site. Often called "dark traffic" because it doesn't always show up with an identifiable referrer in classic analytics tools, which complicates tracking and attribution.

Action Schema An extension of Schema.org vocabulary that no longer just describes "what a page is", but "what an agent can do on it" — book, buy, request a quote. Directly relevant to AEO, since it makes a page actionable, not just readable, by an AI agent. Source

Zero-click / Answer engine Concepts already familiar from classic SEO — a "zero-click" search is a query whose answer is given directly in the results, with no click to a site — that take on renewed importance in the agentic era: an "answer engine" like a generative AI answers the user directly, with or without citing its source.

LLM — Large Language Model (AEO perspective) From an AEO standpoint, an LLM is the "decision engine" of an AI agent. Making your content extractable and usable by an LLM — through clear structure, factual data and Schema.org markup — is the foundation of any GEO/AEO strategy. GPT-4, Claude, Gemini and Llama are LLMs. See section 2 for the technical definition →

llm.txt A text file placed at the root of a website, analogous to robots.txt but targeted at AI crawlers and LLMs. It indicates which pages may be read, summarized or cited by language models, and can point to a condensed site summary (llms-full.txt) to facilitate agent ingestion. Recommended by llmstxt.org and adopted by a growing number of sites. See Busony's llms.txt →