Glossary
The words behind AI-native support
29 terms · 4 groups
Plain definitions for the terms you will meet in Callario and in AI customer support generally — and, where a word means something narrower here, exactly what it means in the product.
01The platform
The parts of Callario a support team works in every day.
- AI agent
- Callario's customer-facing AI. It answers in web chat, on the hosted chat page and in messaging apps from your own knowledge, in the customer's language (English and Arabic, with right-to-left layout), and hands the conversation to a person when it should. It can use only the tools you allow: knowledge search, actions in apps you connect (such as Shopify, Salla or Cal.com), or a procedure — a workflow it runs, which can call an API you define. The customer confirms an action on a card in the chat, and irreversible ones wait for a teammate's approval.
- Copilot
- The AI beside your teammates in the inbox. It summarises a ticket, drafts a reply with inline [n] citations grounded in resolved tickets and internal articles, and suggests a priority, type and tags for a person to confirm. It reads the customer's governed CRM context.
- Inbox
- One queue for tickets and AI chats, with Mine, Unassigned and Mentions views, assignment, statuses, SLAs, merge and split, macros, and a live reply lock that shows when a teammate is already answering.
- Help centre
- Your public knowledge base, served as a customer portal where customers search articles and follow their own requests. Articles are published from the same knowledge base the AI answers from, each pinned to the revision that was reviewed.
- Hosted chat link
- A full-page chat Callario hosts with your AI agent on it — a link you can share anywhere, with a QR code and a share card, without adding anything to your website. It works in English and Arabic, right to left.
- Workflow
- An automation drawn on a canvas, or described in plain language for the AI to draft. Workflows start from 14+ triggers and use steps such as branches, waits, approvals, AI generation and classification, actions in connected apps, “Call an API”, chat steps and sub-workflows. Every workflow can be test-run without side effects, versioned and restored, and each run keeps step-by-step history.
- Automation rule
- A simpler, ticket-scoped automation read as one sentence — WHEN something happens, IF conditions hold, THEN take actions — running instantly on ticket events or on a schedule, with loop prevention built in.
- Customer record
- One lifelong record per customer in the CRM: identities merged automatically (and reversibly), custom fields, segments, notes, tasks, consent, the company they belong to, and an AI knowledge card that shows exactly what the AI may know about them.
02How the AI works
What happens between a customer’s question and the AI’s answer.
- RAG (retrieval-augmented generation)
- Retrieving the most relevant passages from your knowledge at answer time and giving them to the model, so answers stay current and stand on your content. Callario searches with vectors and keywords together, in English and Arabic, for both the AI agent and Copilot.
- Grounding
- Keeping the AI's answer to your own sources — published articles and, if you opt in, resolved tickets — rather than the model's general knowledge. A grounding check runs on every AI turn.
- Citation
- A reference to the source an answer came from. A reply the AI agent sends lists the titles of the articles it cited; a Copilot draft carries inline [n] markers a teammate can check before sending.
- Handoff
- The moment the AI passes a conversation to a person — because the customer asked, showed frustration, the conversation looped, confidence was low or a turn limit was reached. The teammate receives an AI briefing, so the customer does not have to repeat themselves. A handoff never counts as a resolution.
- Confidence threshold
- The point below which the AI agent hands a conversation to a person instead of answering. It is an internal signal: customers never see a confidence score.
- AI triage
- AI that reads an incoming conversation and labels its intent, sentiment and language, and suggests a priority. Run it from a button, or let it run automatically when a routing queue is keyed on those labels.
- Guardrails
- Checks on every AI turn: prompt-injection detection in English and Arabic, a grounding check and a toxicity block. An injection attempt is blocked and handed to a person. Personal data is detected and flagged on the same pass.
- PII scrubbing
- Detecting personally identifiable information in AI conversations, flagging it, and scrubbing it from the traces Callario stores. It is not removed from the live reply — the agent needs those details to help the customer.
03Quality and oversight
How you check the AI’s work, after the fact and before it goes live.
- Verified resolution
- A conversation the AI resolved on its own, confirmed rather than assumed. A language-model judge reads the whole conversation; deterministic checks make sure a bare "ok" or "thanks" never counts; when the judge is unsure it rules against; and a reopened conversation loses its verdict. Callario reports it as a quality measure.
- Resolution rate
- The share of AI conversations that ended in a verified resolution. Handoffs, silence and give-ups never count, so the number reflects customers who were actually helped.
- Deflection
- A conversation kept from reaching a person. On its own it is not success — a customer who gives up was deflected but not helped — which is why Callario reports verified resolutions and handoff rate alongside it.
- AutoQA
- Automatic scoring of conversations against a quality rubric. Every verified AI resolution and a 5% sample of human conversations (up to 25 a day) are scored. It is on by default.
- Eval gate
- A check before an AI agent goes live: the agent answers your golden test questions as itself, and a failed critical question blocks publishing.
- Run trace
- The step-by-step record of an AI conversation: what was retrieved, which guardrails ran and what they decided, which tools were called, and why the agent answered or handed off.
- Audit log
- A hash-chained, tamper-evident record of configuration changes, publishes, takeovers, people and access changes, and API-key and webhook changes. Admins can verify the chain and export it as CSV. Individual AI turns are not in the audit log; they live in run traces.
- CSAT (customer satisfaction)
- A rating the customer gives after their conversation. Each rating stays with the teammate who held the ticket and shows up in reports next to resolution outcomes.
04Operations and trust
The terms behind getting the right case to the right person, safely.
- Routing
- Deciding who takes a case: queues with least-busy, round-robin, longest-idle or manual strategies, skills, capacity and availability. Queues can route on what AI triage detected, and every assignment records why it was made.
- SLA (service-level agreement)
- A first-response or resolution target, measured in business hours, with a pause for time that should not count. Callario raises "about to breach" and "breached" events that workflows can act on.
- Omnichannel
- Handling every support channel on one customer record. In Callario that is web chat, the hosted chat page, the help centre, email, WhatsApp, Slack, Telegram, Facebook Messenger, Instagram DMs and SMS.
- Tenant isolation
- Keeping each workspace's data apart. Callario enforces it in the database with row-level security, so one workspace's queries cannot read another's rows.
- Role-based access control
- Permissions by role. Callario has five — owner, admin, lead, agent and viewer — and an admin can require two-factor authentication for everyone in the workspace.
Design-partner programme · now onboarding
See every term working in the product.
A walkthrough of the AI agent, Copilot, workflows and the run trace, on your own use case.