"Money deducted but no order" is the message Kenyan e-commerce support staff see most, and an AI agent can resolve many such cases if people supervise it well. Training should cover each agent's depth, reading verdicts, matching mobile money records, and keeping refunds, fraud and angry customers with people, under the Data Protection Act 2019.
What does a support agent at a Kenyan online shop handle?
In this brief, "support staff" are the people and "the agent" is the AI agent. Support staff at a Kenyan online shop or marketplace answer customers by chat, phone, text and social media, in English, Swahili and often a mix of the two. The common contacts are payments taken by mobile money where the order did not confirm, delivery updates for riders and pickup stations, cash on delivery issues, returns of wrong or counterfeit-looking items, and refunds to mobile wallets.
Scams add load. Fraudsters call customers pretending to be the shop and ask for payment to a different number, and customers then contact support to ask why their order has not arrived. Support staff need to recognise these quickly, because the money went to someone else.
How does the day change when an AI agent arrives?
The agent has a depth, from 1 (read) to 4 (moves money or is hard to undo), earned through certification on the shop's own past contacts. Its role sets a ceiling; the lower of ceiling and evidence applies. What shifts in a support person's day is described in what changes in a customer support agent's day when AI agents arrive. In a Kenyan shop it looks like this:
- Payment not reflected, depth 3 once certified. The agent matches the customer's mobile money reference to the payment record and confirms the order. If no match exists, it drafts a reply and holds.
- Delivery and pickup updates, depth 3. Sent from the rider and station tracking data.
- Returns, depth 2. The agent drafts instructions and the remedy for a person to approve.
- Refunds to mobile wallets, depth 4, within the shop's limit. Above it, the agent may only draft, and a named team lead makes the refund.
- Suspected scam reports, depth 1. The agent reads and routes to a trained person straight away.
What should support staff learn?
They need to read a verdict and know which of five bounds decided it: role scope, certificate, sources, the shop's own limits or the platform guardrails. They need to know the difference between held (drafted and waiting for a named person) and refused (outside the agent's bounds).
They need provenance. The shop's payment record and the provider's confirmation are sealed. A customer's screenshot of a mobile money message is external, and fake confirmation messages are a known fraud; an external claim cannot authorise a refund or confirm an order.
Corrections should be banked, not repeated. When a new payment reference format keeps tripping the agent, update its sources once; teaching support staff to bank corrections shows the routine. Staff should also expect the agent's certificate to lapse after 90 days, or to drop to depth 1 the same day after a run that clears no depth, and know how replies change when it does.
What stays with people?
| Decision | Why it stays with a person |
|---|---|
| Releasing refunds above the limit | Mobile money moves instantly and is hard to recover |
| Handling scam victims | They need care and clear advice, and the shop needs the facts |
| Judging counterfeit or damage claims | Needs evidence and judgement |
| Responding to angry or threatening customers | Tone and de-escalation |
| Any change to a customer's phone number or delivery details after ordering | A common route for fraud |
Who releases what, and up to what amount, belongs in an approver rule for support leadership.
Which Kenyan rules apply?
The Data Protection Act 2019, enforced by the Office of the Data Protection Commissioner, requires personal data to be processed lawfully, fairly and transparently, for specified purposes, limited to what is necessary, and kept secure. Phone numbers, mobile money references, delivery addresses and order histories are all personal data. Support staff must not copy them into personal chat apps or share them with riders beyond what delivery needs. The guide to the Data Protection Act and training staff who use AI agents goes further.
Mobile money and other payment service providers are supervised by the Central Bank of Kenya, so refunds and reversals follow the provider's rules as well as the shop's. Consumer protection law requires fair dealing and accurate information about goods. If the shop serves customers in the EU, GDPR may also apply. The broader view for Kenyan employers is in what AI agent training organisations in Kenya need.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
How to train the team in 90 days
- Weeks 1 to 4. Literacy for all support staff, in English and Swahili: depth, verdicts, provenance, the record and data handling. Include a fake payment screenshot in the exercises.
- Weeks 5 to 8. Team leads review held refunds and unmatched payments daily, and the team calibrates weekly on returns drafts.
- Weeks 9 to 12. Refund release ties to a current certificate. Rehearse a certificate lapse on a busy evening, when orders peak. Pryme Intelligence Academy's Operator track is one way to certify the team leads who release refunds.
Riders and pickup station staff are not reviewers, but they create the tracking data the agent relies on; the guide to AI agent skills by role shows where every role sits.
Questions people ask
Can the agent confirm an order from a customer's payment screenshot?
No. It must match the payment in the shop's records. A screenshot is external and may be fake.
What should the agent do when a customer reports a scam call?
Route the conversation to a trained person immediately and stop any automated replies about that order.
Should reviewers check Swahili drafts?
Yes, and only reviewers who read Swahili well should release them. Calibrate in both languages.