how to set human escalation rules in an AI agent.
heysteff is an AI platform for customer support and sales across WhatsApp, Instagram, Messenger, Gmail and Shopify; Steff is the AI agent that runs it. This guide is for whoever decides which conversations an AI agent should never handle alone, and wants it written down so the system enforces it every time — not only when someone remembers.
◆ prerequisites
Admin access to the workspace, a clear idea of which case types should always go to a human (complaints, VIPs, large deals, whatever applies to your business), and at least one person on the team available to receive those alerts.
Step 1: List the cases that ALWAYS escalate
Before touching any setting, put together a simple list of the conversation types that should never be left to the agent alone: formal complaints, VIP customers, deals above a certain amount, legal or medical topics depending on your category. This list is the foundation for everything else — without it, it's easy for escalation to end up half-baked or to depend on the agent "noticing" on its own.
Step 2: Configure trigger keywords
Translate that list into concrete words and phrases that, when they show up in the customer's message, fire automatic escalation: "refund," "lawyer," "cancel my order," or the terms specific to your business. The context router checks the conversation's most recent messages against these rules; when it finds a match, the conversation is flagged to escalate before the agent even finishes replying.
Step 3: Configure the supervisor alert
Define how your team learns there's an escalated conversation waiting. The goal is that no one has to watch the inbox "just in case" — the alert lands only when it should, so escalation doesn't depend on someone catching it in time on their own.
Step 4: Decide what the bot does while it waits
When a conversation escalates, the bot is paused automatically: it doesn't keep answering new messages until a human takes it or reactivates it. The agent's last reply does go out (with the escalation context already accounted for), but from there the conversation waits for a person — so you avoid the bot pushing an automated reply while the customer is waiting for someone on the team.
Step 5: Review escalations in the inbox
In the unified inbox, escalated conversations show first — ahead of ones paused for other reasons and ahead of normal conversations — sorted within each group by the most recent. This keeps what's most urgent at the top, visible, without the team having to filter manually.
Rule-based escalation vs. manual pause
Keyword escalation is automatic and happens without anyone stepping in, but it lives alongside the manual pause: any human agent can take a conversation at any moment, even if no rule matched, and heysteff detects that manual reply to pause the bot automatically there too. They're two mechanisms that complement each other — rules cover the predictable, the manual pause covers human judgment for everything else. A good rule of thumb is to start conservative: escalate a little too often at first, watch which cases the human team resolves in seconds, and tighten the keywords over the following weeks so the agent keeps the volume it can genuinely handle.
FAQ
Does the bot pause only for that conversation or for the whole channel?
By default, only for the conversation that escalated. The pause can also be applied more broadly — by channel, by whole workspace or globally — but that's a separate decision, not automatic from the escalation.
Can I have different escalation rules per agent or channel?
Escalation rules are configured at the agent level, so if you have different agents per channel or per brand, each one can have its own set of keywords and cases.
What happens if no one on the team takes the escalated conversation?
It stays waiting in the inbox, at the top of the priority order, until someone takes it — the bot doesn't reply on its own again while it stays paused.