A customer support agent working beside an AI agent needs a short, practical grounding before the first shift: what the AI agent sends on its own, what it holds for a person, how to check the facts behind a draft, what is never approved, and when to take over a conversation. With that, review is quick and safe.
What should you know before your first shift?
A checklist, in the order you will need it.
- What the agent sends without you. Order status, delivery updates, policy questions and other routine replies, at depth 3, within the business’s limits.
- What it drafts for you to send. Complaints, cancellations, account changes and anything outside routine, at depth 2.
- What it holds for a decision. Refunds above the limit, goodwill payments, exceptions to policy.
- What it refuses. Requests outside its role, or blocked by the business’s limits or the platform guardrails.
- Where to find the reason. Each held or refused item names the bound that decided it: role scope, certificate, sources, the business’s limits or the platform guardrails.
What do held, refused and withheld mean on your screen?
- Held: the agent drafted it and it waits for a person. It is your queue.
- Refused: the action was outside the agent’s bounds. You may still be able to help the customer manually under your own authority.
- Withheld: the platform refuses it on purpose for every agent at every depth. Do not look for a way round it.
How do you check the facts behind a draft?
Every fact carries its source class. Look for it on anything that matters.
- Sealed: from your own systems, such as the order record. Trust it.
- Signed: a named party answers for it, such as a courier’s delivery confirmation.
- Declared: what the customer said, such as “it arrived broken”.
- External: from outside, such as a photo or an emailed receipt.
A refund draft that rests on a declared statement and an external photo may be fine for a small amount. For a large one, or a customer with several recent claims, look harder. And never send a refund to a different card or account because the customer asked by message; that is an external claim, and an external claim cannot authorise a payment.
What should you never approve?
Promises outside policy. Disclosure of account details to someone who has not passed verification. Changes to an account’s email or phone number without the full verification process. Any reply that pretends the agent is a person. Your team will have its own short list; learn it before your first shift.
When should you take over the conversation?
Whenever a customer asks for a person. Whenever there is any sign of vulnerability: bereavement, illness, money worries, distress. Whenever the customer is angry and the agent’s replies are making it worse. And whenever you are not sure the agent understands the problem. Taking over is never a failure; it is the job.
How do you review quickly without missing things?
Read the customer’s last message first, then the draft, then the facts behind any promise or payment. Approve if right, edit if nearly right, decline or escalate if wrong. When you edit, add a short note saying why. Those notes are how the agent improves before its next certification.
Take a support agent at an online retailer. A held refund for a damaged kettle is above the limit. The draft is polite and the order is sealed. The damage claim is declared, with an external photo. The customer has had two refunds this month. The support agent approves this one, because the photo is clear, but flags the pattern to their team lead.
What changes if the AI agent is not available?
Its authority expires. Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. On that day, routine replies come back to you. Your team lead should warn you when a re-certification is due.
How do other roles and markets compare?
Your team lead has a different list; see what support team leads should learn first. IT administrators see the same agents from the systems side; see what IT administrators should learn first. Market practice differs: compare support training in Australian e-commerce with Singapore telecoms. Why support agents are the safeguard is argued in why customer support agents are the real control. The overview is in AI agent skills by role.
Where the Academy fits
Pryme Intelligence Academy’s Foundations track takes about four hours and earns a free badge stamped with the product version. Support staff who review held items every day go on to the Operator track, with tests set in their own industry and role. Lessons can be listened to as audio between shifts.
Questions people ask
Will customers know they are talking to an AI agent?
They should be told. In the EU the AI Act requires it, and it is good practice everywhere.
Can I change what the agent is allowed to send?
No. Raise it with your team lead with examples from the record. Limits are changed by their owner through governance.
What if I disagree with a refusal?
Help the customer manually if you have the authority, and report the case so the limits can be reviewed.