CUSTOMER SUPPORT GUIDE
When AI customer support should escalate to a human.
The dangerous version of AI support tries to answer everything. A dependable system is designed around boundaries: routine questions can move faster, while sensitive, expensive, or uncertain cases are handed to a person.
Escalate money decisions
Refunds, chargebacks, disputed invoices, licensing exceptions, and unusual billing requests can create financial and contractual consequences. AI can collect context or identify the topic, but a person should decide what the business will authorize.
Even a familiar refund question may depend on purchase date, plan, jurisdiction, prior promises, or account history. A confident-sounding generic reply is not a substitute for reviewing the actual case.
Escalate security and privacy concerns
Reports involving vulnerabilities, suspicious access, exposed information, privacy requests, account compromise, or customer data need controlled handling. Do not let an automated reply reveal internal details or ask the customer to send credentials.
The first response may acknowledge receipt and provide a safe next step, but investigation, disclosure decisions, and remediation belong with an authorized person.
Escalate possible data loss and destructive actions
Requests involving deletion, restoration, migration, overwriting, or failed updates can become irreversible. The support system should not invent recovery steps or instruct a customer to run destructive commands without understanding the environment.
Attachments and screenshots also deserve caution because they may contain personal information, license keys, email addresses, order details, or server paths that the ordinary automated workflow was not designed to process.
Escalate uncertainty, not just keywords
A message can be low risk in topic but still unsafe to answer when the knowledge is incomplete, contradictory, outdated, or a poor match. Confidence should come from approved local knowledge and clear applicability—not from how fluent the draft sounds.
When the system cannot point to a suitable answer or documentation source, escalation is the correct result. Guessing creates more support work because someone must later repair the customer’s trust and the original problem.
Make escalation useful
An escalation should carry the customer message, relevant product and plan, detected risk category, any matched knowledge, and a concise reason the automated workflow stopped. The human should not have to reconstruct the case from scratch.
Good automation is measured by safe routine work completed and important cases surfaced—not by the percentage of tickets answered without a person.