AI Customer Support Glossary: 60 Terms Defined in Plain English

If you compare AI support tools, fill in security questionnaires or read support dashboards, these definitions are for you. Vendors name the same idea differently: Intercom bills per outcome, Zendesk per automated resolution and Tidio per Lyro conversation, and each unit counts a different set of chats.

Each entry opens with a definition that stands on its own, then notes where definitions split. Vendor terms were read on Intercom's, Zendesk's and Tidio's own pages on 19 October 2026, and entries that describe PepoChat, the AI support agent we build, say so.

AI and model terms, from LLM to tool call

The ten AI and model terms below describe the software that answers your customers and why it sometimes gets things wrong.

AI support agent

An AI support agent is software that answers customers in a chat window from your own content, takes actions such as an order lookup, and hands over to a person when it cannot help. What is an AI support agent? compares it with chatbots and copilots.

Chatbot

A chatbot is any program that holds a text conversation, either by following scripted buttons and keyword rules (a rule-based chatbot) or by generating replies with a language model. When vendors contrast chatbots with AI agents, they usually mean the rule-based kind.

AI agent

An AI agent is software built on a language model that chooses actions, runs them through tools and reads the results before deciding what to do next. Marketing stretches the word to any AI chatbot; agentic AI in customer service lists what agents really do.

Copilot

A copilot, or agent assist, is an AI assistant for your human agents rather than your customers, drafting replies and summarising conversations. Intercom, Zendesk and Tidio all offer one named Copilot; PepoChat's inbox has a narrower tool, an Enhance button that rewrites drafts.

Large language model (LLM)

A large language model (LLM) is a neural network trained on vast amounts of text to predict the next token, which lets it write replies, summarise and follow instructions. PepoChat's replies come from OpenAI's gpt-4o-mini.

Token

A token is the unit of text a language model reads and writes, usually a word or part of one. Model providers bill per token, which is why some chatbot vendors, Crisp among them, sell AI usage as token-based credits.

Context window

The context window is the most text, measured in tokens, that a model can consider at once, including its instructions, the conversation so far and retrieved passages. Keeping earlier turns of a chat in that window is session memory, which is different from meeting memory.

Hallucination

A hallucination is a confident answer that is false or unsupported by its source, such as an invented refund policy. NIST's generative AI profile (July 2024) calls it confabulation and treats hallucination as the colloquial name.

Fine-tuning

Fine-tuning is further training of an existing model on your own examples so that it writes or behaves differently. Fine-tuning suits style, while facts belong in retrieval, because facts change and a fine-tuned model cannot show where an answer came from.

Tool call

A tool call is the model's structured request to run one operation you defined, such as looking up an order, with typed inputs and a result it reads before replying. Vendors package tools as actions, functions, skills or data connectors.

Retrieval and knowledge base terms, from RAG to sitemap import

The nine retrieval terms below explain where an AI support agent's answers come from, and why a missing page means a missing answer.

Knowledge base

A knowledge base is the private collection of pages, files and snippets an AI support agent searches before every reply, although helpdesks also use the phrase for a public help centre. How answers work traces a PepoChat reply through one.

Knowledge source

A knowledge source is one item in a knowledge base, such as a crawled website, an uploaded file or a connected Notion workspace. In PepoChat a whole-site crawl counts as one source and snippets as none; the free plan allows five as of October 2026.

Retrieval-augmented generation (RAG)

Retrieval-augmented generation (RAG) is searching your content for passages relevant to a question and handing them to the model as it writes the answer. The name comes from a 2020 paper by Patrick Lewis and colleagues that paired a text generator with a retriever over a Wikipedia index.

Embedding

An embedding is a list of numbers that represents the meaning of a text, so that texts with similar meanings get similar numbers. Embeddings let "Can I get my money back?" find a page titled "Refund policy"; PepoChat uses OpenAI's text-embedding-3-small.

Vector search

Vector search, or semantic search, returns the stored passages whose embeddings sit closest to the question's, matching meaning rather than exact words. Many systems add keyword search, because product codes and error messages match better on exact text.

Chunk

A chunk is a passage of a document, typically a few paragraphs, stored and retrieved as one unit. A table split across two chunks can come back incomplete, which is why RAG for customer support treats chunking as a major influence on answer quality.

Grounding

Grounding means making the model answer from source text supplied as it replies, not from training memory, so every claim traces back to your content. A grounded agent says it does not know when retrieval finds nothing; stopping chatbot hallucinations covers the techniques.

Crawl

A crawl is an import that starts at one address and follows a site's links to collect pages. PepoChat's crawl takes up to 200 pages, 3 levels deep, respects robots.txt and re-fetches weekly; training a chatbot on your website walks through one.

Sitemap import

A sitemap import adds the pages listed in a site's sitemap file instead of following links, so it reaches pages that no menu links to. In PepoChat a sitemap import takes up to 200 pages with path filters and counts as one knowledge source.

Conversation design terms, from system prompt to canned reply

The seven conversation design terms below cover what the agent says and how you shape it.

System prompt

A system prompt is the hidden text an application places before every conversation, telling the model what it is, what to answer from and what never to do. Users can sometimes extract it, so keep secrets out; 12 instruction examples show what belongs in one.

Agent instructions

Agent instructions are the part of the system prompt you write yourself: persona, tone and refusals on top of the vendor's base rules. PepoChat's two fields, for tone and for refusals, hold 4,000 and 2,000 characters, and its base rules always win, so instructions cannot make the agent invent answers.

Welcome message

A welcome message, or greeting, is the first message a chat widget shows before the visitor types. The best ones say in a line or two what the agent can do and invite a question.

Suggested questions

Suggested questions are tappable buttons shown with the greeting, also called conversation starters or quick replies. They work best as questions your content answers reliably; PepoChat shows up to three.

Fallback message

A fallback message is what a chatbot says when it cannot give a grounded answer, because the content is missing, the question is unclear or a rule blocks it. A good one offers a next step; 20 fallback message examples show the wording.

Proactive message

A proactive message is one the widget sends before the visitor writes, triggered by a rule such as time on a pricing page. Zendesk calls them proactive messages, Tidio's Lyro has proactive roles, and Intercom files them under outbound messaging.

Canned reply

A canned reply is a saved answer a human agent inserts with a shortcut, often with actions attached such as tagging or closing the conversation. Intercom, Zendesk and Tidio all call them macros; other tools say saved replies.

Handoff and team inbox terms

The five handoff terms below describe how the AI passes a conversation to a person, and where that person works.

Human handoff

Human handoff is the transfer of a live conversation from an AI agent to a person, with the transcript and collected details, so the customer need not repeat anything. A handoff can be live or asynchronous, as human handoff for AI chatbots explains.

Escalation trigger

An escalation trigger is a condition that makes the AI stop and hand over, such as a request for a person, no relevant content or a failed action. Strictly, escalation is the decision and handoff the transfer, though most vendors use the words interchangeably.

Team inbox

A team inbox is the shared view where your people watch conversations live and take over the ones the AI hands off. Replies typed in PepoChat's team inbox appear in the widget but are not emailed to the visitor.

Away message

An away message is what the agent tells a visitor who asks for a person while nobody can reply live. PepoChat's holds up to 300 characters, and 25 offline message examples show wording that keeps the lead.

Ticket

A ticket is a support request recorded in a helpdesk, the software that stores and routes requests, with a number, status, owner and history. An AI agent can open one when a chat needs follow-up, as creating tickets from a chatbot shows.

Support metrics: deflection vs containment vs resolution rate

Deflection, containment and resolution measure different things: a ticket avoided, a chat that never reached a person, and a problem solved. Benchmarks rarely say which one they mean.

First response time (FRT)

First response time (FRT) is the time from a customer's first message to the first substantive reply, summarised per channel. Definitions split on whether an instant AI reply counts; first response time benchmarks lists published figures by channel.

Deflection rate

Deflection rate is the share of would-be tickets that never reach a person because the customer found an answer first, a figure that can only be estimated. Chatbot vendors often use it for any chat that never reached a person, which is containment; see deflection benchmarks.

Containment rate

Containment rate is the share of bot conversations that ended without escalating to a person: one minus escalated conversations divided by all conversations. A visitor who gave up counts as contained, so containment shows where chats ended, not whether anyone was helped.

Resolution rate

Resolution rate is the share of AI-handled conversations that ended with the problem solved, by each vendor's own definition. Intercom counts 24 hours of silence after Fin's last answer as an assumed resolution, while Zendesk bills only verified resolutions; metrics that matter shows a strict count.

Fallback rate

Fallback rate is the share of visitor messages, or of conversations, that end in a fallback message instead of an answer. State which denominator you use, and read a rising rate on one topic as a sign of a missing page.

Customer satisfaction score (CSAT)

Customer satisfaction score (CSAT) is the share of customers who rate an interaction positively in a one-question survey, such as 4 or 5 out of 5. Some tools report the average score instead; PepoChat shows the share of thumbs-up votes on resolved chats over 90 days.

Identity and security terms, from verified visitor to prompt injection

The five identity and security terms below decide who the agent believes it is talking to and what a stranger can make it do.

Verified visitor

A verified visitor has proved control of an email address in the current chat session, usually with a one-time code. In PepoChat verification unlocks meeting memory, pre-filled bookings and an emailed transcript, while order lookups do not require it, as the visitor verification guide explains.

One-time code

A one-time code, or one-time password (OTP), is a short secret sent to a channel the customer controls and accepted once, within minutes. PepoChat's six-digit codes expire after 10 minutes, allow 5 attempts and can be resent after 60 seconds; chatbot identity verification covers when to require one.

Meeting memory

Meeting memory is PepoChat's feature that lets the agent recall a returning customer's past meetings with your team, such as commitments and summaries, from PepoSmart notes. Meeting memory is off by default and needs a verified visitor, as the meeting memory documentation describes.

Allowed domains

Allowed domains is the list of websites on which your chat widget may start conversations, so a copied snippet elsewhere cannot spend your AI replies. PepoChat accepts up to 20 entries, subdomains included.

Prompt injection

Prompt injection is text written to make a model ignore its instructions, typed into the chat or planted in a page or file the model reads. The OWASP Top 10 for LLM applications ranks prompt injection first for 2025; the usual defences are treating retrieved text as data and limiting what tools can do.

Privacy and compliance terms, from AI disclosure to WCAG

The nine privacy and compliance terms below come up in security reviews and procurement forms. They are defined here, not claimed: PepoChat's privacy notes say it does not currently claim a hosting region, a signed DPA template or a SOC 2 report.

AI disclosure

AI disclosure is telling customers they are talking to an AI rather than a person, ideally in the first message. Laws in the EU and several US states now require it in some situations; AI chatbot disclosure laws lists them.

Data controller

A data controller is the organisation that decides why and how personal data is processed, normally the business running the website chat. The UK regulator's guide to controllers and processors gives the legal wording.

Data processor

A data processor handles personal data on a controller's behalf and instructions, such as a chat vendor or its AI model provider. PepoChat's privacy notes ask GDPR customers to list it as a processor and the model provider as a sub-processor; see GDPR and AI chatbots.

Data processing agreement (DPA)

A data processing agreement (DPA) is the contract GDPR Article 28 requires between a controller and each processor, covering purpose, data types, security, sub-processors and deletion. Ask every chat vendor for its DPA and sub-processor list before launch.

Data retention

Data retention is how long a vendor keeps conversations and contact details before deleting them. PepoChat's retention setting is off by default, and owners and admins can have resolved conversations deleted after 30, 90, 180 or 365 days.

Zero data retention (ZDR)

Zero data retention (ZDR) means an AI model provider leaves your prompts and responses out of the logs it would otherwise keep. OpenAI's documentation says ZDR needs its prior approval and that API abuse-monitoring logs are otherwise kept for up to 30 days.

SOC 2

A SOC 2 report is an independent CPA firm's examination of a service organisation's controls for security, availability, processing integrity, confidentiality or privacy, under AICPA guidance. Type 1 covers control design at one date; type 2 also tests operation over a period.

Business associate agreement (BAA)

A business associate agreement (BAA) is the contract HIPAA requires before a US healthcare provider or health plan lets a vendor handle protected health information on its behalf (45 CFR 164.504(e)). Never send patient health details to a chat vendor that will not sign one.

Accessibility (WCAG)

Accessibility means people with disabilities can use the chat widget, including by keyboard alone and with a screen reader. The usual yardstick is WCAG 2.2, a W3C Recommendation with levels A, AA and AAA, of which AA is the usual legal and procurement target. The chat widget accessibility checklist turns it into 20 tests.

Pricing units: what AI support vendors bill you for

The billing unit decides what a busy month costs. The vendor examples below were read on each vendor's own pages on 19 October 2026.

AI reply

An AI reply is one answer the AI agent generates in a visitor conversation, the unit PepoChat's plans count; greetings, team messages and test chats are not counted. The free plan includes 200 a month as of October 2026, and conversations per month helps you size yours.

Resolution

A resolution, as a billing unit, is a conversation the vendor counts as solved by its AI. Help Scout and HubSpot bill per AI resolution and Zendesk per automated resolution, with different rules on whether a customer who goes quiet counts.

Outcome

An outcome is Intercom's billing unit for Fin, charged at most once per conversation when Fin finishes a job, such as a resolution or a procedure that hands over to a person. Intercom listed $0.99 per outcome on 19 October 2026; Intercom Fin pricing explained works through a bill.

Conversation

A conversation, as a billing unit, is one chat thread, but vendors draw the line differently. As of 19 October 2026, Botpress counts one once the user has sent two messages, Tidio once its AI has replied, and My AskAI groups one visitor's messages within an hour.

Credit

A credit is a prepaid unit of usage that a vendor converts into AI replies at a rate set by the model or feature. Chatbase charged 1 to 5 credits per reply, by model, on 19 October 2026, so the same allowance buys different numbers of answers.

Seat

A seat is a paid login for a person on your team who answers customers; visitors are never seats. Intercom says full seat, Zendesk prices per agent and PepoChat counts members, with extra seats at $15 a month as of October 2026.

Session

A session is a bounded period of one visitor's activity, with three common meanings: a visit in web analytics, a billing window such as Freshdesk's 24-hour Freddy AI Agent session, and the anonymous visitor record in a chat widget, which in PepoChat lasts 24 hours.

Overage

Overage is usage beyond your plan's allowance, which vendors either bill or block until a top-up or the next month. PepoChat's paid plans bill $10 per 1,000 extra replies as of October 2026, and the free plan sends new conversations to the team inbox instead.

Flat plan

A flat plan charges a fixed monthly price for a set allowance, so the bill does not rise when the AI resolves more conversations. PepoChat's plans work this way, with every feature on the free plan except new-lead emails.

Which words do Intercom, Zendesk and Tidio use for the same thing?

Intercom, Zendesk and Tidio name the same parts differently. The table maps each vendor's word to the plain term used in this glossary.

Last verified 19 October 2026 on each vendor's pricing page and help centre.

Only macros and Copilot match across all three vendors, while the billing unit differs at each, so translate that row first.

What to do next

Pick the terms that decide your purchase, usually the billing unit, the handoff behaviour and the resolution definition, and ask each vendor for its exact wording before comparing prices. The AI customer support pricing index compares 30 tools by pricing model, and a free PepoChat workspace shows these terms working, from verified visitor to team inbox.