Customers do not hate bots for existing. They hate explaining the same order number three times. Escalation without context is that failure mode with a smile: the assistant asked clarifying questions, ruled out two options, then handed a person a name and an email as if the conversation never happened.
The fix is operational, not poetic. Decide when to hand over, where it lands, and what travels with it. The step-by-step lives in handle escalations; this post is the product reasoning behind those settings.
Two triggers, different manners
Escalate on refusal when the visitor still needs something the sources cannot provide. Escalate immediately on explicit request for a person. Attempting “one more helpful answer” after “I want a human” is how you teach people to type “agent” as their first message forever.
The transcript is the payload
Send the full conversation, the page the visitor was on, and any contact detail already collected. The agent should open able to reply, not interview. If your destination is Slack, email, or a ticket system, the requirement is the same: the human sees what was asked and what was already ruled out.
A “lead” that is only a name is not an escalation — it is a cold contact with extra steps. Wire delivery to wherever the team already works so the transcript is not trapped in a dashboard nobody opens after week two.
Honesty about hours
Offering live handover when nobody is online is a broken promise. Configure working hours so the assistant offers a person while staffed and captures a message when not. The AI can still answer from sources around the clock; hours change only what the human path claims.
Where it should land
- Slack or Teams for teams that already triage there
- Email for lower volume with clear ownership
- Your existing helpdesk when agents live in tickets
- Per-bot routing if you run departments or clients separately
Pick the channel someone will see within your response-time target. A perfect transcript in an unwatched inbox is still a failure.
What agents should do with context
Train the team to read before typing. The first human reply should reference what the visitor already said (“I can see you already checked the tracking link…”). That single habit is what makes the bot feel like part of support rather than a maze in front of it.
Also feed patterns back: if escalations cluster on one missing page, that is a content gap, not an agent skill issue. Escalation quality is measurable — count repeats of information, not only escalation volume.
Launch check
- Ask for a human; confirm immediate offer.
- Force a refusal; confirm escalation is offered.
- Complete a handover; confirm transcript arrives intact.
- Test outside office hours; confirm the promise matches staffing.
If any of those fail, fix escalation before you spend another hour on tone. Personality cannot rescue a handoff that throws away the work.
