COD and Returns Support Playbook for Emerging Markets

If you’re selling into Southeast Asia, the Middle East, or Latin America, cash on delivery (COD) is probably carrying a meaningful share of your orders. It converts well, but it also creates a steady stream of support tickets: buyers changing addresses mid-transit, couriers showing up to a closed door, packages arriving short or damaged with a refund request attached. These three scenarios happen daily, and most of them are stuck inside a narrow logistics window — respond too slowly and you lose the order or take the bad review. Here’s a playbook broken down by scenario, with a clear line drawn between what AI support can handle on its own and what always needs a human approval.

Why COD markets are harder on support than prepaid ones

With prepaid orders, the pressure sits mostly in post-purchase support. COD is different — the risk starts the moment the order is confirmed, not after it ships. A buyer might change their mind, request a new address, or simply not be home when the courier arrives, and courier policies themselves vary (some allow opening the package for inspection at the door, some don’t). That means support isn’t dealing with one clean “refund flow” — it’s a chain of moments between order and delivery where anything can interrupt, often across multiple languages, time zones, and channels (WhatsApp, Instagram, the website widget, all at once). Agents manually checking shipment status, address rules, and return policy for every single ticket can’t keep up with a buyer’s decision window at the doorstep.

Three high-frequency scenarios worth mapping out

• Address or schedule changes: the package is already in transit, and the buyer wants to switch the delivery address, recipient, or timing.
• Refusal at the door: the courier arrives and the buyer isn’t home, disputes the amount, or simply changes their mind.
• Damage or shortage claims: the package arrives, and the buyer reports missing items, damage, or a mismatch with the listing, and wants a refund or replacement.

All three share one thing in common: they happen after the shipment is already moving, the window to act is usually a few hours to a couple of days, and missing it usually means the order gets marked delivered regardless.

Address changes: AI can handle these directly when the conditions are clear

Address changes are the best fit for AI support to take on first, because the rules are relatively well-defined — whether the carrier allows mid-transit changes, the time window for making them, and delivery coverage for the new address can all be documented in the knowledge base ahead of time. When a buyer asks to change an address, AI support can:

1. Check the knowledge base to confirm whether the carrier assigned to that order supports mid-transit address changes;
2. Verify the request is still within the allowed time window;
3. If it’s within the window and the request is complete, tell the buyer how to proceed and pass the update to the relevant logistics step;
4. Hand off to a human agent if the change crosses a country or customs boundary, or if the carrier doesn’t support address changes at all.

Once this logic is running, most “still changeable” requests get an answer in seconds instead of an agent digging through chat history to find a tracking number. For more on how AI support draws that line between answering directly and handing off, see where AI-first, human-backed support draws the boundary.

Refusal at the door: reassure first, then decide whether the order can be saved

Refusals are messier than address changes because the reasons vary widely — genuinely not wanting the item anymore, disputing the charged amount, or just being unavailable at delivery time. AI support can do the first pass of triage:

AI support can resolve most of the straightforward “reschedule delivery” cases outright, leaving the conversations that need judgment and recovery skill for human agents — who can see exactly what AI already asked and how the buyer answered inside the shared workspace, instead of starting the conversation over.

Refunds and compensation: AI can start the conversation, but approval always stays with a human

This is where the line has to be drawn clearly. For damage, shortage, or refund requests, AI support can verify order details, walk the buyer through the return or claim process and required documentation per the knowledge base, and collect photos and descriptions. But any step that confirms an actual refund amount, changes a price, or decides on compensation always routes to a human for approval — AI never executes it automatically.

For more on how this boundary works in practice, see how AI gets smarter without acting on unconfirmed guesses — when AI hits something uncertain or outside its scope, it hands off rather than guessing.

Layer your knowledge base by market, not one blanket COD policy

COD rules differ meaningfully across Southeast Asia, the Middle East, and Latin America — some regions allow opening packages for inspection before accepting, others don’t; who covers return shipping after a refusal varies by carrier too. It’s worth structuring the knowledge base by country or region rather than writing one generic policy and hoping it fits everywhere. You can upload each carrier’s policy documents, crawl the shipping and returns pages from their websites, or have a support lead turn common judgment calls into manual Q&A entries. The more granular the knowledge base, the higher the share of address-change and refusal decisions AI support can handle on its own — leaving humans to focus on the cases that genuinely need judgment. See how the knowledge base feeds AI for more on setting this up.

Every resolved ticket should turn into a reusable skill, not a one-off

Agents who spend enough time on refusal recovery and damage claims build up their own instincts — which refusal reasons are usually recoverable, which carrier’s claim requirements are stricter in practice than the policy doc says. If that knowledge stays in one agent’s head, the next person handling the same case starts from scratch. In YundaDesk, when an agent answers on AI’s behalf or corrects an AI response inside the shared workspace, the system generates a suggested learning item for review. It only becomes part of AI support’s skills or knowledge base after a manager approves it — and every item stays traceable to its source, testable on its own, and reversible with one click. A single misjudged case never gets baked into AI’s behavior automatically.

Handling the surge during peak season

COD order volume commonly doubles during major sales events, and address changes, refusals, and claims tend to grow at the same rate or faster, since buyers make more impulsive purchases during promotions. This is where AI support handling the routine cases pays off most — simple address lookups and standard refusal triage keep response times manageable during the peak, so human agents aren’t buried under repetitive questions and can focus on the genuinely complicated compensation disputes. Reviewing your COD-related knowledge base before peak season, and checking whether last year’s approved learning items still apply, is worth doing ahead of time. See the peak season support playbook for more on getting ready.

The hard part of COD support was never whether to use AI — it’s drawing the line between what AI can resolve on its own and what always needs human judgment and approval. Address changes and routine refusal triage, where the rules are clear, are a good fit for AI to handle directly. Anything touching refund amounts or compensation decisions stays on the human approval track. Get that boundary right, and you keep both response speed and risk control intact.