Peak trading exposes the training gaps online retailers leave. Service teams are taught tone, not judgement, so refunds get approved to protect satisfaction scores. Temporary staff receive approval rights on day one. Merchandisers sign off agent-drafted listings and price changes without reading them, and nobody teaches how refund and account fraud look when an agent has prepared the case.
Why do satisfaction targets distort refund review?
Online retailers measure service teams on customer satisfaction and speed. When a refunds agent drafts a refund that falls outside its bounds, perhaps above the customer's limit or without a matching return scan, it holds the refund for a person. That person is measured on keeping the customer happy.
The predictable result is approval. A held refund is a complaint waiting to happen, and releasing it ends the conversation. Over weeks, the record fills with approvals that override the agent's holds, and the business has trained its people, without meaning to, to ignore the very checks it bought. Measure refund reviewers separately: sample their decisions for correctness, and reward the ones who catch abuse as well as the ones who resolve contacts quickly.
Why shouldn't peak temps get approval rights?
Retailers scale service teams for peak with temporary staff, trained in a few days on systems and scripts. Giving them approval rights over held refunds, address changes or goodwill credits puts the least-trained people at the most consequential points, during the weeks when fraud attempts rise.
Let temporary staff handle contacts and draft notes, and route held approvals to certified permanent staff. A retailer can adopt the practice of granting approval rights only to people with a current certificate, depth by depth. Size the certified team for peak volume in advance, using the volume of held actions the agent produced in quieter months as a guide. A plan with dates and owners is in the AI agent training plan template for online retailers.
What does fraud look like when an agent prepares the case?
Retail fraud often works through ordinary-looking requests: a refund for an item never returned, a delivery address changed just before dispatch, a claim that a parcel never arrived on an account with a pattern of such claims. An agent can read order history, return scans and courier events, and it will hold cases that do not add up.
The reviewer must understand why. A return scan from the retailer's own warehouse system is sealed. A courier tracking event from a named carrier is signed. The customer's statement that the parcel did not arrive is declared. A screenshot sent in support of it is external and cannot authorise a refund on its own. Reviewers who know this read a hold as information, not obstruction, and look for the pattern behind the single case.
Why do merchandisers need agent training too?
Agents increasingly draft product listings, descriptions and price changes at depth 2, with a merchandiser approving each one. Retailers treat this as content work and skip training.
But a listing is a statement to customers and a price change is a commitment. A misplaced decimal on a price, a promotion that stacks with another unexpectedly, or a product claim copied from a supplier's data sheet can each cause real loss or real harm. Supplier-provided product data is external until checked, especially for safety, allergen or age-restriction claims. Merchandisers need to know what the agent drew on and what they must verify before approving. Depth 2 means a person sends; it does not mean a person skims. Train merchandisers on real past listing errors, including near misses, so they know which fields to read first.
What about telling customers they are talking to an agent?
Where an agent replies to customers directly, people should know they are dealing with an AI system. In the EU, Article 50 of the AI Act requires it. Service staff should be trained to explain held and refused requests clearly, without blaming the system, and to pick up a conversation from the agent without making the customer repeat everything. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What happens after a lapse in the middle of peak?
An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. If that happens during peak, refunds the agent would have issued within bounds start arriving in the held queue, just when the team is busiest.
Retailers should plan certificate renewals away from peak where possible, and train team leads to recognise a lapse quickly and staff for the surge. The worst response, and the most common, is to let temporary staff clear the backlog.
What should an online retailer do instead?
- Measure refund reviewers on correct decisions, not only satisfaction and speed.
- Keep approval rights with certified permanent staff; let temps draft and route.
- Teach provenance on real refund, delivery and fraud cases.
- Train merchandisers to verify listings and prices before approving.
- Plan certificate renewals and reviewer capacity around peak.
The broader order is in the guide to training teams for AI agents. For a retailer's version, see how online retailers should train staff before deploying AI agents and what an AI agent training programme for online retailers should include. Universities face a similar peak problem with temporary staff, covered in what universities get wrong about AI agent training.
Questions people ask
Can a refunds agent issue refunds without a person?
Within the retailer's own limits and once certified at depth 3, routine refunds with a matching return scan can go through without a person. Anything outside those limits, or resting on weaker evidence, is held for a trained reviewer.
Should warehouse staff be trained on the agent?
They need to know that their scans become the evidence behind refunds and claims, so accuracy matters. They need reviewer training only if they approve held actions.
How do we train marketplace seller support teams?
Treat seller payouts like any other payment: held at depth 4 for certified staff, with bank detail changes verified through a known channel. Seller-provided data is external until checked.