Rules let you write your own IF/THEN logic for how Arbyn handles conversations. Instead of describing every edge case in your instructions to Arbyn, you build a condition (a VIP customer writes in, an order is over a certain value, the language isn't English) and pick what should happen when it's true.
What it does
A rule you save is checked at five points while Arbyn is actually working a conversation, so it can change behavior mid-flow rather than only at the very start. When a rule's conditions match, its actions run: escalate to your team, tag the conversation, route it, require your approval before anything is sent, hold a saved quick reply ready, snooze it, lower or raise the confidence bar, or log a custom event for your own reporting.
Where to find it
Go to Rules in the main navigation, or from Settings, Guardrails via the Custom Rules card.
Rules vs. Escalation Rules vs. Guardrails
These are three separate, deliberately distinct systems, and it's worth knowing which one you actually want:
Custom Rules (this page) run general automation: tag, route, snooze, attach a quick reply, require approval, and more.
Escalation Rules, a separate page, handle one specific decision: whether a conversation should hand off to a human at all.
Guardrails control how much freedom Arbyn has to actually take actions like refunds or address changes, not when its behavior changes.
A banner on the Rules page links straight to Escalation Rules if that's what you're actually after.
Setup (step by step)
Tap New rule, or Browse templates for a ready-made starting point, or Generate with AI and describe what you want in plain English (Arbyn drafts a structured rule for you to review and adjust before saving).
Give it a name and optional description (internal only, customers never see either).
Choose a scope: Global (applies everywhere), or narrowed to a specific product, tag, collection, or customer tag.
Build your conditions. Pick a field, an operator, and a value, then combine multiple conditions with AND or OR. You can nest one extra layer of conditions inside a condition group for cases like "A, and either B or C."
Add one or more actions for what should happen when the conditions match.
Toggle Active, then save.
How it works
Conditions
Fields come from five categories: conversation (intent, channel, language), customer (total spent, number of orders, tags, whether they're a returning customer), order (financial status, fulfilment status, total price), product (tags), and AI signal (Arbyn's own confidence in its answer). The operator options change to match the field, numbers get greater-than/less-than/equals, text gets contains/starts with, and lists get "is one of."
Actions
A rule can combine more than one action. The available ones are: escalate to your team, tag the conversation, route it to a queue, set a minimum confidence threshold, attach a saved quick reply, snooze it, require your approval before Arbyn sends anything, or track a custom event.
Templates
Sixteen starter templates cover common situations without building from scratch, among them: flagging a spam-tagged customer as do-not-reply, always escalating VIP customers, escalating orders above a value threshold, flagging non-English conversations for translation, auto-tagging complaints, blocking automatic refunds on final-sale items, handing wholesale discount requests to your team, escalating a chat-only product asked about over email, keeping no-bundle products out of recommendations, escalating language that reads as legally risky, tagging VIPs for a personalized greeting, escalating refund requests on delivered (stale) orders, quoting your refund policy automatically through a quick reply, snoozing low-confidence weekend questions until Monday, requiring your approval on drafts for high-value orders, and escalating complaints immediately rather than waiting for Arbyn to decide it can't help.
Generate with AI
Describe what you want in a sentence or two, and Arbyn drafts a structured rule (conditions, actions, and scope) for you to review, edit, and save, rather than you having to translate it into the builder yourself.
Rule health
Once a rule has existed for 14 days, this page shows whether it's actually doing anything: Healthy (it's matched and fired recently), Silent (it hasn't matched anything lately), or Dead (it hasn't matched in a long time and may no longer apply). New rules under 14 days old show Healthy by default, since there isn't enough data yet to judge them, so a fresh rule showing Healthy isn't proof it's working, only that it's too new to have failed yet.
Good to know
Every save to an existing rule is versioned, so you have a record of what a rule looked like before your latest edit.
Pause instead of delete if you want to turn a rule off temporarily without losing it; the same Actions menu that edits and deletes a rule also lets you toggle it active or paused.
A rule's condition fields are pulled from real Arbyn data, not a hand-typed list, so the values you can match against (like intent or order status) always reflect what the system actually tracks.
An escalate action discards the draft, it doesn't skip drafting. Arbyn still generates a reply behind the scenes the same way it does for any message; when a matching rule's escalate action fires, that draft is thrown away rather than sent, and your team sees the handoff instead.
FAQ
What happens if two rules both match the same conversation? Both can fire. Rules aren't mutually exclusive; if two different rules each have conditions that are true, both sets of actions apply.
Can I test a rule before it affects real customers? Save it as inactive first, review the health status once you're confident in the logic, then switch it to Active. A newly-active rule only evaluates conversations going forward.
Does a rule that requires approval slow every reply down? No, only conversations that match that specific rule's conditions get held for approval; everything else keeps moving at normal speed.