AI deconstruction 01Hong Kong retail · public workflow study
What an AI-prepared return case could look like.
An AI assistant could prepare a return case for review: the affected item, the customer’s account, the evidence supplied and the information still missing. This independent design study uses HKTVmall’s public reporting guidance and fictional examples to test that step. The person reviewing the case retains the decision.
The published process
HKTVmall’s customer guidance already describes an automated reporting route. A customer selects the affected product, records the problem category, quantity and condition, adds photographs, and receives an immediate confirmation. The page says approval takes two to three working days.
This is useful credit, and a boundary. The public page shows what the customer submits and when the result is expected. It does not show how the case is prepared internally, how much time that takes or which tools support it.
Report the affected product through the order record.
Category, quantity, condition and photographs are named in the guidance.
The system sends an immediate email or message.
The guidance gives a two-to-three-working-day approval period.
Reconstructed from HKTVmall’s published customer guidance · checked 8 September 2026
The proposed preparation step
The candidate task sits between submission and review: assemble the case from the customer’s account, show the source behind each field, and stop where the evidence runs out. This is narrow enough to test and clear enough for a reviewer to judge.
What the assistant prepares
Item, quantities, reported issue, evidence status and a draft request for missing information. Every field remains linked to its source.
What the reviewer decides
Whether the evidence is reliable, whether the request meets the applicable policy, and what outcome is approved. The demonstration does not automate those decisions.
Interactive concept · prepared examples
Fictional order · no data is uploadedsource material
AI-prepared case brief
Please attach a photograph showing the damaged mug and keep the original packaging while the case is reviewed.
Prepared outputs, written in advance for this demonstration. A functioning pilot would still need unseen test cases, source-fidelity checks and measurement of total review effort.
What the demonstration establishes
It establishes a design proposition: evidence can be organised into a compact case while uncertainty remains explicit. It does not establish that HKTVmall needs this step, that a production system would perform it reliably or that it would save time.
The useful output is a reviewable case
A reviewer needs the material, the account and the gap in one place. The decision remains separate.
Missing information is part of the output
A system that completes the form by guessing has made the work look tidy and made the evidence worse.
The pilot must count checking
Compare accepted cases with the same task completed manually. Include reviewer time and correction effort in the result.
How the proposition would be tested
Run representative cases through the proposed system, then add cases it has not seen. Record whether each field traces to the source, whether missing and contradictory information is identified, and whether the draft adds unsupported statements. Measure human time through accepted output, including checking and correction.
A result is worth carrying forward only when the evidence survives review and the complete task improves against the baseline. The next step is one bounded workflow, one named owner and one person accountable for the decision.