Customers judge the whole service by what the agent says, so the head of customer service owns four settings: which replies are held for a person, the refund limits agents work within, how tone is reviewed, and the short list of things no agent may ever promise. Those four decide whether agents raise service quality or quietly erode it.
Which replies should be held for a person?
Held means the agent drafts and a named person decides before anything reaches the customer. A sensible starting split for most service teams:
| Contact | Starting treatment |
|---|---|
| Order status, delivery, opening hours, policy facts | Sent by the agent at depth 3 within bounds |
| Refunds within the limit, to the original payment method | Sent by the agent once certified for it |
| Refunds above the limit, goodwill payments | Held for a reviewer |
| Complaints, cancellations with fees, account closures | Drafted at depth 2, a person sends |
| Bereavement, illness, financial difficulty, safety | Held and routed to a senior person |
| Legal threats, regulator mentions, media | Held for the head of service or legal |
Depth is earned, not configured. An agent moves from drafting to sending only after it has been certified on real cases from your own service history. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day, so plan cover.
How should refund limits work?
As the business’s own limits, written down and owned. A refund limit is one of five bounds an agent meets on every action, alongside its role scope, certificate, sources and the platform guardrails. The lowest wins. If a customer asks for a refund of 300 and the limit is 100, the agent drafts it and holds it; it does not refund 100 and promise the rest.
Two details matter. Refunds should go to the original payment method; a customer’s request to send money elsewhere is an external claim and cannot authorise a payment. And the limit should vary by reason: a missed delivery is not a disputed charge.
How do you keep the agent’s tone right?
By treating tone as reviewable content, not a style setting. Reviewers should correct replies that are accurate but cold, over-apologetic or full of jargon, and record why. Those corrections feed the agent’s brief. Look for patterns monthly: if reviewers keep softening the same kind of reply, the brief needs changing.
Customers should also know when they are dealing with an agent. Under the EU AI Act, people must be told when they interact with an AI system, and the obligations are explained in the EU AI Act for AI agents in customer service. Many businesses outside the EU adopt the same rule because customers expect it.
What must an agent never promise?
The never-list for service is short and firm:
- Compensation, refunds or credits outside policy or above the limit.
- A delivery date, repair date or outcome the business cannot guarantee.
- That a complaint is upheld, or that an investigation will find in the customer’s favour.
- That a fee will be waived, unless policy says so for that case.
- Anything about a customer’s account to someone who has not passed verification, which is where know your customer checks meet service.
These sit in the business’s own limits and produce refused or held verdicts. Test each one before go-live and after any change.
What does the head of service read each week?
The record, through a few numbers. Held-item age by type. How often reviewers change agent drafts, and on which topics. Refusals by bound. Complaints mentioning the agent. Upcoming certificate dates. A head of service who reads those can explain to the COO exactly where agents help and where they do not; see what a COO needs to know about AI agents.
Compliance will look at the same record for fair outcomes and red lines; the compliance team’s view explains how. In telecoms and other high-volume sectors, the training angle differs by market: compare AI agent training for customer support agents in Italian telecoms. The full map is in AI agent skills by role.
How should the service team be prepared?
Reviewers first. They are the control on every held reply and must know what held, refused and withheld mean, how to read the sources behind a draft, and when to escalate. Team leads need to build a reviewer rota and calibrate decisions. Train them on held items from your own service, not generic examples.
Questions people ask
Can the agent handle complaints?
It can acknowledge, gather facts and draft a response. Deciding the outcome of a complaint should stay with a person.
What if customers ask to speak to a human?
Route them, every time. The agent should never argue a customer out of reaching a person.
Do we need a separate tone guide for agents?
No. Use the same guide your people use, and let reviewers’ corrections tune the agent’s brief over time.