The failure mode support leads recognize
Zendesk (and peers) make it easy to wire an LLM to draft or send public replies. Volume drops. CSAT ticks up — until the model picks the wrong macro, invents a policy exception, or soft-commits a refund that your finance team does not honor.
The gap is not “AI in the helpdesk.” The gap is treating a public customer reply like a private draft. Once it is on the ticket, it is a commitment. Prompt hygiene is not an approval chain.
What LoopLabs demonstrates
In the live control plane, ticket public replies are a governed action class: FAQ-class autos can run with a receipt; money and legal macros queue for a named owner before send.
Classify the reply
Shipping FAQ ≠ refund promise ≠ legal hold language. Only the first class should ever auto-send by default.
Queue the exceptions
Billing disputes, chargebacks, and goodwill credits land in approval with the draft attached — the agent did the triage; a human releases the words.
Receipt on every send
Auto or human: log ticket id, macro/draft hash, policy version, and releaser. That is what you open when the customer forwards the thread to counsel.
Easy escalate
Customers (and agents) need a one-click path out of the auto lane. Silent dead-ends create chargebacks.
The first policy to ship
- Public ticket replies that mention refund, credit, legal, or account closure never auto-send.
- FAQ autos require an allowlisted macro id and write a receipt before send.
- Every public reply records draft/macro, policy version, and whether a human released it.
- Agents can always escalate to a human queue without re-prompting the model.
- Weekly review: autos that get overridden often mean the allowlist is wrong.
Related guides
Ticket reply governance pairs with refund-above-limit approval and external-party send rules. For tool calls with no evidence bundle, see MCP tool call without receipt.