AI customer service
glossary.
This AI customer service glossary defines the words you meet when comparing AI agents and chatbots, from automated resolution to prompt injection. Each term starts with a one-sentence definition, then explains what it means in practice and how Dazza handles it.
AI customer service agent
An AI customer service agent is software that answers customers’ questions in conversation and completes simple jobs for them, such as checking an order or booking a call, handing over to a person when it can’t help.
It uses a large language model to understand questions however they’re written, and answers from the business’s own content rather than a script. The better ones are grounded in that content, cite it, and say when they don’t know.
What makes it an agent is that it can act: collect the details for a return, look something up in the business’s systems or pass a request to the team. Those actions should run within rules the business sets, checked outside the AI model.
Dazza is an AI customer service agent for any business. It answers on your website, by email, on WhatsApp and in your app, gets jobs done through workflows, and hands over to your team with a summary.
AI agent vs chatbot
A chatbot answers in conversation; an AI agent also does things for the customer, such as looking up an order or starting a return, within limits the business sets.
Early chatbots followed scripts of buttons, keywords and decision trees. Many of today’s AI chatbots use a language model to answer more naturally, but still can’t act. An AI agent does both: it understands the question, then completes the task or passes it on.
The names are used loosely, so ask two questions of any product: can it finish the job, and what stops it doing the wrong one?
In Dazza, workflows collect the details a job needs and finish it by sending it to your team, sharing a link or calling your system. Decisions, customer confirmation, team approval and verification are checked in code before anything runs.
Automated resolution
An automated resolution is a customer conversation that an AI agent settles without a person getting involved.
Some vendors bill by it, so how it’s counted matters. Zendesk bills its AI agents by automated resolutions: conversations the AI agent resolved without escalation to a person, checked by an AI model once the conversation ends, which by default is 2 hours after the first message on messaging and 72 hours after the first email.
Intercom charges from $0.99 per Fin outcome. An outcome is a resolution, where the customer confirms the answer helped or leaves without asking for more help, or a set procedure that Fin completes before handing over.
Dazza doesn’t charge per resolution. It counts conversations: one thread Dazza replies to, however it ends.
Prices checked 27 October 2026. Sources: About automated resolution tiers (Zendesk); Fin pricing: outcomes (Intercom); Intercom pricing.
Resolution rate
Resolution rate is the share of conversations that end with the customer’s problem solved; for an AI agent, it usually means the share solved without a person.
It’s also called automation rate, and is close to containment rate, which counts conversations that never reached a person, solved or not. Each vendor measures it differently, so compare the definitions before the numbers.
A high rate isn’t the goal on its own: an agent that never hands over scores well while customers give up. Read it beside handovers, unanswered questions and customer ratings.
Dazza’s Insights shows the conversations handled by Dazza, meaning it answered from your knowledge and never handed to the team, beside handovers by reason and the questions it couldn’t answer.
Handover (human handoff)
A handover is when an AI agent passes a conversation to a person, ideally with a summary so the customer doesn’t have to repeat themselves.
Good handovers happen at the right moment: when the customer asks for a person, is upset, needs something only staff can do, or the agent can’t answer. Bad ones loop the customer back to the bot, or drop them with no context.
Dazza hands over in those moments with a short summary of what was asked and tried. Your team takes over in one click from the shared inbox, replies, and can hand the conversation back when it’s done.
Grounded answers and citations
A grounded answer is taken from specific sources, such as a business’s website or documents, rather than the AI model’s general knowledge; a citation shows which source it came from.
Grounding is the main defence against made-up answers, often called hallucinations. Citations let customers check an answer, and let the business see which page was behind a wrong one.
Dazza answers only from your knowledge and what the customer has said, adds a source number after each fact, and says “I don’t have that information” rather than guessing. Your own Q&As and pasted text win over web pages and files when they disagree, and Correct this turns a wrong answer into one of your own.
Retrieval
Retrieval is the step where an AI agent searches its knowledge for the passages most relevant to a question before it writes the answer.
An answer is only as good as what’s retrieved: miss the right passage and the model either says it doesn’t know or answers from the wrong one. The whole approach is often called retrieval-augmented generation, or RAG. Some systems search by meaning, with vector embeddings; others search by words.
Dazza uses full-text search, weighted so your own answers rank higher, with exact matching for order numbers and product codes. Small knowledge bases are sent to the model in full. In Knowledge, Test a question shows what Dazza would find.
MCP (Model Context Protocol)
The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside tools and data, such as a store’s orders or a payment system, through one common interface.
Anthropic introduced it in November 2024. An MCP server describes the tools it offers, such as “find an order” or “create a refund”, and any AI application that supports MCP can call them, so one integration works with many applications.
Dazza connects to MCP servers: Stripe, Zapier, n8n, Make or your own, and has Shopify built in without one. Every tool starts off. Switch each one to Test or Live and choose its checks: customer confirmation, verified customers only and team approval.
Verified customer
A verified customer has proved who they are, for example with a code sent to their email or by being signed in to the business’s site, before an AI agent shares or changes anything personal.
Anyone can type someone else’s email or order number into a chat. Without verification, an agent that looks up orders or accounts can hand one customer’s details to another.
Dazza verifies customers with a six-digit code sent to their email, which expires after 10 minutes and is rate-limited; with a login token signed by your own server; or, by email, when the sender’s domain passes DMARC. Verification lasts 12 hours in that conversation. An email typed into the chat never counts, and your system is told how the customer was verified.
WhatsApp 24-hour reply window
The WhatsApp 24-hour reply window, which Meta calls the customer service window, is the 24 hours after a customer’s last message in which a business can reply freely; once it closes, only pre-approved template messages can be sent.
Each new message from the customer resets the timer to 24 hours. Inside the window a business can send any kind of message; outside it, it needs a template that WhatsApp has approved in advance.
Dazza and your team can reply on WhatsApp while the window is open. Once it closes, a reply waits until the customer writes again, and then you press Try again.
Per-conversation pricing
Per-conversation pricing charges for each customer conversation an AI agent handles, however many messages it takes, rather than for each seat or each resolution.
It makes the bill predictable: a long, difficult thread costs the same as a quick one, and adding people to the team costs nothing. Seat pricing charges for each person who uses the software. Resolution pricing charges each time the AI is judged to have solved something, which depends on how the vendor counts it.
Dazza counts one thread it replies to as one conversation: a chat, an email thread or a messaging thread. Follow-ups in the same thread don’t count again, and preview chats, threads only your team answers, and spam or auto-replies Dazza ignores are free. Plans are Starter £39 a month with 300, Growth £119 with 1,500 and Scale £349 with 6,000 conversations, all with unlimited seats.
CSAT
CSAT, or customer satisfaction score, is the share of customers who rate an interaction as good, usually from a short survey at the end of a conversation.
It’s usually worked out as good ratings, such as 4 or 5 out of 5, divided by all ratings. It’s quick to collect, but only some customers answer and they lean happy or angry, so read it beside what customers actually said.
When a chat on your website or in your app ends, Dazza asks “How was this chat?”: Good, Okay or Poor, with an optional comment. Insights shows the share rated Good as the customer rating, beside the thumbs up and down customers give single replies.
Proactive chat
Proactive chat is when the chat starts the conversation, for example with a message that opens on a pricing or checkout page, instead of waiting for the customer to ask.
Used sparingly, it catches people who are stuck or about to leave. Used everywhere, it’s a pop-up people learn to close. The best triggers are specific: one page, a time on the page or a return visit.
Dazza calls these nudges: a short message above the chat button on the pages you choose, after the wait you set, on every visit or only the first. Clicking it opens the chat with your message as Dazza’s first line, and a nudge only becomes a conversation once the visitor writes.
Knowledge base
A knowledge base is the collection of content an AI agent answers from: help articles, website pages, documents and written answers to common questions.
It decides the quality of every answer. Out-of-date pages give out-of-date answers, and gaps become handovers. The fix is ordinary housekeeping: one clear answer per question, kept current.
Dazza builds its knowledge from your website’s sitemap (up to 250 pages on Starter, 1,000 on Growth and 5,000 on Scale), files (PDF, Word, text, Markdown, CSV and HTML), pasted text and your own Q&As, and can import help centres such as Zendesk. It re-reads pages daily or weekly and adds new ones.
Prompt injection
Prompt injection is an attack in which text written to look like instructions, in a customer’s message or a page the AI reads, tries to make the AI model ignore its rules.
It can be direct, typed by the customer, or indirect, hidden in a web page, file or email the agent reads. OWASP lists it first in its 2025 Top 10 for applications built on large language models. No model is immune, so the safest systems don’t rely on the model to enforce what matters.
Dazza’s model is told that knowledge and customer messages are information, not instructions. More importantly, what it can do is limited in code: workflows and tools run only within the checks you set, such as verified customers, confirmation and team approval; decisions are evaluated by the system, not the model; secrets are never shown to the model; and your system receives the verified identity, not what the AI says about the customer.
Source: LLM01:2025 Prompt Injection (OWASP).