AI & AI Agent Glossary

The terms that come up most often around AI, AI Agents and conversation technology, gathered in one place with their counterparts on the CBOT platform.

A

Agentic AICore Concepts: An AI system that does more than answer a single question: it plans and carries out several steps on its own to reach a goal, retrieving information, making decisions, calling a tool, judging the result and changing its plan when it needs to. This category defines CBOT's platform approach. Architecture in detail: Agentic AI.
AI AgentCore Concepts: Software that can run a task end to end, make decisions and connect to systems to complete an action. What separates it from a classic chatbot is that it does not stop at answering questions; it can carry a whole process through, such as identity verification, a balance enquiry or booking an appointment. At CBOT the term is used at two levels: channel based AI Agents (voice, messaging, live chat) and role based Digital Employees, named agents such as TAHSİLDAR, GÜVEN, ADİL and KALYA that each own a specific job. Platform overview: Explore the Platform.
AI GovernanceGovernance: The framework that sets out how an organisation develops, audits and protects its AI systems against risk. It covers risk assessment, human oversight, explainability and model lifecycle management. In detail: AI Governance and Safety.
AI Infrastructure and InferenceInfrastructure: The stage where a model runs in real time after training, producing an answer for every user request. The speed and cost of this stage depend as much on hardware and architectural choices as on the model itself. In detail: AI Infrastructure and Inference, CBOT Research.
Attention MechanismInfrastructure: The core mechanism a language model uses to work out which other words it should pay attention to while processing each word in a sentence. In "I closed my bank account but I still have the card", it is what tries to connect "the card" to the right account. It is the foundation of today's large language models, the transformer architecture.
Autonomous AIAI Agent Design: The degree to which a system can make and act on its own decisions without consulting a person at every step. Autonomy is a scale rather than a binary state. An AI Agent can be designed as a mixture: fully autonomous on certain decisions and human consulted on riskier ones. See Human in the Loop.

B

BiasGovernance: A model's tendency, inherited from its training data, to produce systematically different and unfair results for particular groups or situations. If a collections model tends to use a different tone depending on a name or a region, that is a sign of bias and it needs to be tested regularly.

C

Chain of ThoughtCore Concepts: A language model working through its own reasoning steps in sequence rather than jumping straight to a conclusion. In situations that need a multi step decision, such as judging whether an exception request qualifies against several rules, this approach reduces the chance of error.
ChatbotCore Concepts: Software that holds a question and answer exchange based on predefined rules or a simple language model, usually limited to single step replies. It gets confused with an AI Agent, and the difference matters: a chatbot gives information, an AI Agent completes the task. Asked "what is my card limit", a chatbot returns an explanation; an AI Agent can verify identity, look up the real limit, state it and start an increase request if needed.
Context WindowInfrastructure: The amount of text a language model can hold in mind at once. Past conversation, instructions and knowledge sources all have to fit inside it. As the window fills the model can forget details from the start of the call, and on long voice calls that is a limit to account for at design time.
Continuous BatchingInfrastructure: An inference server processing several requests that arrive at different moments together rather than one after another, so the hardware is used more efficiently. Without it, at peak hours every new call behaves as though it were waiting for the previous one to finish. In today's inference stacks it is an optimisation that directly determines how many concurrent calls you can carry.
Conversation DesignAI Agent Design: The discipline of planning how an AI Agent asks, listens and steers. It is like interface design, except it works through dialogue rather than a screen. The basics: Foundations of Conversation Design.

D

Decode StageInfrastructure: The stage where a language model produces its answer word by word, or more precisely token by token. Unlike the prefill stage this one runs in sequence: producing the next token requires the previous one to be finished, which is why its speed is bound by memory bandwidth rather than raw compute. See Prefill Stage.
Digital EmployeeAI Agent Design: CBOT's name for an AI Agent that takes on a specific business role end to end, defined by a name and a job: TAHSİLDAR for collections, GÜVEN for claims, ADİL for emergencies, KALYA for quality. What separates it from the general term AI Agent is that it does not describe a vague capability. It takes over the work a real employee does, within defined boundaries. All roles: AI Agent Portfolio.
DistillationInfrastructure: Teaching a much smaller and faster model what a large, slow one knows. The aim is a version that comes close to the original's capability while being far cheaper and quicker to run, which makes it a common choice in voice AI where latency is critical.
DriftGovernance: A model quietly moving away from the performance it had when it was trained. User behaviour changes, new products appear, language patterns evolve, and the model may not keep up. Left unnoticed, answer quality falls even though nobody changed anything. Managing it: Managing Drift.

E

EmbeddingInfrastructure: Turning the meaning of a word, a sentence or a document into a list of numbers, a vector, positioned so that content with similar meaning sits close together mathematically. The retrieval step of a RAG system works by measuring similarity between these vectors.
EscalationAI Agent Design: An AI Agent noticing that a situation falls outside its authority or confidence and handing the conversation to a person or a more authorised process. A well designed escalation is not the agent failing; it is part of how the system stays reliable. Framework: Edge Cases and Escalation.

F

Fine-TuningAI Agent Design: Adapting a general purpose language model to a particular organisation's terminology, tone or sector with additional training. Rather than training a model from scratch for every organisation, specialising a strong existing one this way is far faster and easier to measure.
Function CallingAI Agent Design: The same capability as tool calling: the model deciding for itself when to call an external system or function and with which parameters. See Tool Calling.

G

GPU and Inference ChipsInfrastructure: The specialised hardware a language model needs both during training and when running in real time, able to carry out thousands of calculations in parallel. Most of the cost of an AI Agent comes down to how efficiently this hardware is used.
GQA and MQAInfrastructure: Architectural techniques that reduce the intermediate calculations a language model keeps in memory, the KV cache. Instead of every attention head keeping its own memory, several heads share it. The result is far more concurrent calls at low latency on the same hardware.
GroundingAI Agent Design: Making sure an AI Agent's answer comes from a verified source it was given, a document, a database record, a case study, rather than from its own guess. When grounding is weak the model can produce plausible information with no real source behind it. See Hallucination.
GuardrailsAI Agent Design: Fixed boundaries drawn around a model's behaviour that define what an AI Agent must never do or say, for example never citing an unverified customer name or approving an action beyond its authority. A good guardrail is the last layer of control that stops a model making a well meant but wrong decision.

H

HallucinationAI Agent Design: A language model producing false or invented information with confidence, without a verified source behind it. An AI Agent might state an unverified customer reference as though it knew it. The way to prevent this is rules and architecture that force the agent to speak only from verified sources. See Grounding, Guardrails.
Human in the LoopGovernance: An AI Agent putting certain decisions in front of a person for approval instead of taking them automatically, for instance a high value commitment or a sensitive exception. It is a design principle that accepts not every process suits full automation.
HyperparameterInfrastructure: Settings the developer chooses in advance rather than the model learning them, such as the temperature value that decides how creative an answer will be. See Sampling.

I

Intent RecognitionAI Agent Design: Working out the real purpose behind what a user says or writes. Behind "I lost my card" there is both alarm and a clear request for an action, blocking the card, and intent recognition separates the two.
ISO/IEC 42001Governance: An international standard for AI management systems. It frames how an organisation develops its AI systems, controls them against risk and audits them. In regulated sectors such as banking and the public sector it is becoming a supplier requirement. In detail: ISO 42001 and AI Governance.

K

KV CacheInfrastructure: The structure where a language model keeps intermediate results in memory so it does not recompute every earlier token while producing an answer. The longer the conversation, the larger this cache grows and the more memory it uses. On long voice calls it is one of the main bottlenecks that decides system capacity. See PagedAttention, GQA and MQA.

L

Large Language ModelCore Concepts: An AI model trained on an enormous amount of text, which learned to produce language by predicting the next word. These models are the basis of an AI Agent's ability to understand and speak today, but an LLM on its own is not an AI Agent: an agent is built by adding decision making, tool calling and memory on top of it. How that layer is managed at CBOT: LLM Orchestration, CBOT LLM.
Latency and ThroughputInfrastructure: Latency is how long a single request takes to answer; throughput is how many requests the system can serve at once. The two usually pull against each other: tuning a server to carry more concurrent calls, raising throughput, can raise the latency of a single call. Finding the right balance is a critical engineering decision in voice AI.

M

MCP (Model Context Protocol)Infrastructure: An open protocol that lets an AI model connect to external tools, databases and systems in a standard way. Rather than writing bespoke code for every integration, you build on a shared interface, which directly affects integration speed in agentic architectures.
Multi Agent ArchitectureAI Agent Design: Several agents each specialised in a particular job working together instead of one large agent: one understands intent, one verifies, one carries out the transaction. In complex processes, splitting the work this way gives more accurate results that are easier to test than loading everything onto a single agent. The approach in detail: Multi AI Agent Architectures.

N

Natural Language ProcessingAI Agent Design: The wider field that turns human language into a form computers can work with, covering both understanding and generation. Natural language understanding is the subfield focused on the understanding half.
Natural Language UnderstandingAI Agent Design: A system's ability to extract the intent and the required details, a date, an amount, a product name, from what a user says or writes. Getting both a promise and a date out of "I'll pay by tomorrow" is the job of NLU.

O

OrchestrationInfrastructure: The management layer that decides which component runs when, in cases where an AI Agent works with several parts rather than a single model: different models, tools and rule engines. In a complex call it is what runs the order of understand the intent, then verify, then call the right tool.

P

PagedAttentionInfrastructure: A technique that manages the KV cache in small blocks, much like paging in an operating system, rather than as one contiguous memory block. In practice it wastes far less memory and lets the same hardware carry many more concurrent calls.
Parameter CountInfrastructure: The total number of numeric values a model learned during training, which determine its behaviour, usually written as 7B for seven billion. More parameters generally mean a stronger model, and also more compute and higher latency, which is why the largest model is not always the right choice.
PersonaAI Agent Design: Defining an AI Agent's tone of voice, word choice and character so that they are consistent with the identity of the organisation it represents. A collections agent and an insurance agent should not sound the same; the emotional state they are meeting is different. Framework: Voice and Persona.
Prefill StageInfrastructure: The stage where a language model first reads and processes the user's input, the question and the conversation history. It can run in parallel and is usually fast; the real bottleneck comes next, in the decode stage where the answer is produced word by word. See Decode Stage, Time to First Token.
Prefix and Semantic CachingInfrastructure: Both avoid repeated computation, at different layers. A prefix cache does not recompute the opening of conversations that begin with the same system instruction. A semantic cache reuses the answer given to a question that was very close in meaning, even when it is not worded identically. Together they cut both cost and latency under heavy traffic.
Prompt EngineeringCore Concepts: The discipline of designing the instructions given to a language model so that it behaves consistently and as intended. A good prompt makes clear not only what to do but what NOT to do in particular situations; every point left ambiguous is a point where the model decides on its own, sometimes wrongly.

Q

QuantizationInfrastructure: Making a model smaller and faster by lowering the precision of its calculations, for instance computing in 8 bit instead of 16 bit. Applied well it can cut both cost and latency substantially with no visible loss of answer quality; applied aggressively it risks losing accuracy.

R

RAG (Retrieval-Augmented Generation)AI Agent Design: A language model pulling relevant information from a current source, documents or a knowledge base, in real time before it answers, and including it in the answer. It lets the model speak from information that is valid now rather than only from what it learned during training and that has since aged. It is one of the most effective ways to reduce hallucination.
Roofline ModelInfrastructure: An engineering framework for understanding what limits the speed of an inference operation: the bottleneck is either processor power or the speed of moving data out of memory. In voice AI, producing an answer for a single user is usually bound by memory, which explains why buying a more powerful processor does not on its own solve latency. Framework in detail: CBOT Research.

S

SamplingInfrastructure: The settings that decide how predictable or how varied a language model is when it picks the next word, among them top-p, top-k and temperature. A low temperature makes the model more consistent and more cautious, which is usually what a collections or banking agent wants, because reliability matters more than creativity there.
Speculative DecodingInfrastructure: A speed up technique where a small, fast model guesses the next few words and a larger, stronger model verifies those guesses in one pass. When the guess is right the large model confirms several words at once instead of producing them one by one, which can shorten response time noticeably.
Speech to Text and Text to SpeechVoice AI Agent: STT is the component that turns spoken audio into text; TTS turns the written answer back into natural speech. In a voice AI agent, STT runs when the customer speaks and TTS runs once the model has produced its answer. The speed of these two components is a large part of total latency.
System PromptCore Concepts: The base set of instructions an AI Agent receives in the background before every conversation, invisible to the user: who it is, how it should behave and which boundaries it must never cross. Most guardrails are technically applied in this layer.

T

Time to First TokenVoice AI Agent: TTFT is the time between a model receiving a request and producing its first word; TTFA is the total time until that first word has been turned into audio and reached the user in a voice AI agent. The human ear reads silences longer than about half a second as odd, which is why this metric is managed to a much tighter budget in voice than in written channels.
TokenCore Concepts: The smallest unit a language model uses when it processes text. It can be a word, part of a word or a punctuation mark. Both the speed and the cost of a model relate directly to how many tokens it processes.
Tool CallingAI Agent Design: An AI Agent deciding for itself, during a conversation, that it needs a piece of information or an action, and reaching it by calling a system, an API or a database query. This is the core capability that turns an agent from something that only talks into something that gets work done. Setup framework: Backend Integrations.
Transformer ArchitectureInfrastructure: The architecture nearly every large language model is built on today. Its key innovation is processing the words in a sentence by weighing their relationships to one another, through the attention mechanism, all at once rather than in sequence. That allows both better understanding of context and faster training through parallel computation.
Turn TakingVoice AI Agent: A voice AI agent deciding when to start answering by correctly reading whether the customer has actually finished speaking, or has simply paused to think. Get it wrong and the agent either interrupts the customer or leaves an uncomfortable silence.

V

Vector DatabaseInfrastructure: A specialised kind of database where text can be searched by meaning rather than by exact word match. Behind a RAG system, this is usually what finds the most relevant information. See Embedding.
Voice AIVoice AI Agent: An AI Agent interacting over a phone call, by speech rather than in writing. Understanding what is said, answering in a natural voice and deciding in real time all happen at once. Its tolerance for latency is far lower than in written channels. Product page: Voice AI.

Z

Zero-shot and Few-shot LearningCore Concepts: A model's ability to take on a new task with no examples at all, zero-shot, or with only a handful, few-shot. Being able to introduce a new scenario to an AI Agent with a few well chosen examples, rather than collecting dozens, shortens the time to deployment.