Last updated: August 2026
By Khanh Nguyen, founder of SMK Vietnam — a decade in SEO, content and AI-assisted marketing, working with Vietnamese brands and foreign companies entering the Vietnam market.
Most Vietnamese brands treat delivery as a cost line. It is not. From the moment a customer taps “Đặt hàng” to the moment the shipper hands over the box, your brand is speaking — through tracking messages, packaging, the shipper’s tone, the COD handover, and the return process. That is five to eight branded touchpoints per order, delivered to a customer who is actively paying attention. Almost nobody markets into that window. This post is about how to, using AI to make it operationally possible for a small team.
AI for last-mile marketing in Vietnam means using language models and automation to turn delivery events into branded, personalised communication across GHN, Giao Hàng Tiết Kiệm (GHTK), J&T Express, Viettel Post, Ninja Van and Shopee Express. In practice: AI rewrites carrier tracking updates into on-brand Zalo and SMS messages, drafts proactive delay apologies before the customer complains, classifies delivery complaints out of Shopee and Facebook comment streams into fixable operational themes, and generates post-delivery review-request and re-order sequences. A Vietnamese SME can build this in about 30 days with a marketer, a Zalo OA, carrier webhooks and one automation tool — no data team required.
On this page
Why delivery is a marketing channel · The touchpoints AI can improve · The stack and the carriers · A 30-day workflow · Mining delivery complaints · Mistakes to avoid · What to measure · FAQ
Why does last-mile delivery matter as a marketing channel in Vietnam?
Because of cash on delivery. In most Western markets the transaction ends at checkout; in Vietnam a large share of e-commerce orders are still COD, which means the sale is not final until the shipper is standing at the door. Every hour of uncertainty between order and handover is an hour in which the customer can refuse the parcel. Delivery communication is therefore not customer service — it is closing.
There is a second reason: attention. Open rates on a “your order is out for delivery” message are far above anything you will ever achieve with a promotional broadcast. You are given a guaranteed-read channel, several times per order, and most brands fill it with the carrier’s default template.
Figures are directional planning estimates drawn from operator disclosures and our own client accounts, not audited market data. Use them to size the opportunity, not to build a forecast.
Which delivery touchpoints can AI actually improve?
Not all of them. AI is good at language, classification and volume — it is useless at moving a motorbike through District 7 at 5pm. Here is where it earns its keep.
On-brand tracking messages
Carrier webhooks fire dry status codes. A model rewrites each state into your brand register — warm and “anh/chị” for a beauty brand, terse and factual for B2B parts — and pushes it to Zalo OA or ZNS instead of a generic SMS.
Proactive delay apologies
When a parcel sits in a hub past your SLA, trigger a drafted apology with a concrete new window and a small gesture. Apologising before the customer complains is the single highest-ROI automation in this list.
Complaint classification at scale
Thousands of Vietnamese comments and reviews per month, tagged automatically into carrier, packaging, wrong-item, late, rude-shipper and refused-COD buckets — with the province attached so you can see where the problem lives.
Timed review requests
Ask for the review 18–36 hours after delivery, not at the moment of handover, and vary the ask by product category. AI writes the variants; the automation picks the window.
Return and exchange copy
Returns are a retention moment, not a loss. Drafted step-by-step Vietnamese instructions, plus an exchange offer generated from what the customer actually bought, recover a meaningful share of would-be refunds.
Packaging insert copy
The one piece of print media your customer definitely opens. Generate category-specific insert cards — QR to a Zalo group, a care guide, a re-order code — and rotate them monthly instead of printing one design for a year.
What does the AI last-mile stack look like in Vietnam?
Three layers: the carriers who generate the events, the automation layer that catches them, and the language layer that writes the message. Most teams already have the first, half of the second, and none of the third.
| Layer | Typical tools in Vietnam | What you need from it |
|---|---|---|
| Carrier / event | GHN, GHTK, J&T Express, Viettel Post, Ninja Van, SPX, VNPost | Webhook or API access to status changes, plus province and hub fields |
| Orchestration | Make, n8n, Zapier; Haravan, Sapo, KiotViet on the commerce side | Trigger on state change, hold state, respect quiet hours, log every send |
| Language / AI | Claude, ChatGPT or Gemini via API, called from the automation | Brand-voice prompt, Vietnamese register rules, strict output length |
| Delivery channel | Zalo OA and ZNS, SMS brandname, Messenger, email | Template approval lead time — plan two weeks for ZNS templates |
The practical constraint is not the model. It is ZNS template approval and the fact that carrier webhook payloads differ in shape between operators — you will normalise those fields once and then never think about it again. If you are still choosing tools, our complete AI marketing stack for Vietnam covers how the pieces fit together, and the Claude playbook goes deeper on getting consistent Vietnamese output from a model.
How do you build this in 30 days?
You do not need a platform migration. This is a five-step sequence a two-person team can run alongside normal work.
Days 1–5 — Map the real journey
Place five test orders through your own store across two carriers and two provinces. Screenshot every message the customer receives. Most teams discover three to four touchpoints they did not know existed and one carrier SMS that contradicts their brand entirely.
Days 6–10 — Write the brand-voice prompt
One prompt file that fixes register (bạn vs anh/chị vs quý khách), maximum length, banned phrases, and how to name your product categories in Vietnamese. Every downstream message is generated against this file. Version it.
Days 11–18 — Wire one carrier, one event
Pick your highest-volume carrier and the “out for delivery” event only. Webhook into Make or n8n, call the model, send to Zalo OA. Ship a single working loop before adding anything. Submit ZNS templates in parallel — approval is the long pole.
Days 19–25 — Add the delay rule and the review ask
Two more branches: parcel exceeds SLA in hub, and delivered plus 24 hours. The delay branch should require a human approval click for the first two weeks so you can watch what the model writes before trusting it.
Days 26–30 — Turn on complaint mining
Pipe last month’s Shopee reviews and Facebook comments through a classification prompt. Produce one page: top five delivery complaint themes, by carrier, by province. Take that page to your carrier account manager. It changes the conversation.
The brands that win last-mile in Vietnam are not the ones with the fastest carrier. They are the ones whose customer always knows what is happening — and hears it in the brand’s own voice, not the carrier’s.
How do you use AI to mine delivery complaints without losing the human touch?
The split is clean once you draw it. Classification, translation, summarisation and first drafts are machine work. Judgement, compensation decisions and anything a customer will screenshot are human work.
Automate confidently
✓ Tagging comments by complaint type and province
✓ Status-change messages within an approved template
✓ Weekly carrier performance summaries for internal use
✓ Drafting review-request and re-order copy variants
✓ Translating complaints for a foreign head office
Keep human
⚠ Any compensation, refund or goodwill amount
⚠ Replies to a public complaint with traction
⚠ Damaged, lost or wrong-item cases with photos
⚠ Anything involving a named shipper’s conduct
⚠ Escalation to a carrier account manager
What mistakes do teams make first?
| Mistake | Do this instead |
|---|---|
| Messaging on every single carrier status change | Three messages maximum per order: confirmed, out for delivery, delivered. More than that and customers mute the OA. |
| Letting the model write freely into a ZNS template | Generate only the variable fields, within a strict character budget. The approved template shape must not change. |
| One tone for all customers | Set register by product category and order value. A 200,000đ accessory and a 20,000,000đ appliance are not the same conversation. |
| Sending at any hour because the webhook fired | Hard quiet-hours rule, roughly 21:00 to 07:00 Vietnam time, with the queue released in the morning. |
| Blaming the carrier in customer-facing copy | Own it publicly, escalate privately. Customers bought from you, not from GHTK. |
| No logging of what the AI actually sent | Store every generated message against the order ID. You will need it the first time something goes wrong. |
What should you measure?
Skip open rates — transactional messages always look good. Measure the things delivery communication is genuinely supposed to move: COD refusal rate, delivery-related complaint volume per thousand orders, review rate within seven days of delivery, second-order rate within 60 days, and inbound “where is my order” contacts per hundred orders. That last one is the cleanest proof of value: if proactive messaging works, it should fall sharply within a month. Pair it with the measurement approach in our Gemini and Workspace playbook if your reporting lives in Sheets, and use Perplexity to keep an eye on how competitors are handling their own delivery communications.
- In a COD-heavy market, delivery communication is closing, not customer service.
- Three branded messages per order beats fifteen carrier status pings.
- The highest-ROI automation is the proactive delay apology, sent before the complaint.
- AI does classification, translation and first drafts; humans own compensation and public replies.
- ZNS template approval, not the AI model, is the real timeline constraint — start it in week one.
- Measure COD refusal rate and “where is my order” contacts, not open rates.
Related reading
The Complete AI Marketing Stack for Vietnam
Claude AI for Marketing in Vietnam
ChatGPT for Marketing in Vietnam
Gemini in Google Workspace for Marketing in Vietnam
Perplexity AI for Marketing in Vietnam
About SMK Vietnam. SMK Vietnam (smkvietnam.com) is a marketing hub based in Vietnam. We help Vietnamese companies and foreign brands from the US, Japan, Korea, Thailand and Europe build AI-assisted marketing operations for the Vietnamese market — search, content, paid media, messaging channels such as Zalo, and the operational plumbing that connects them.
Want your delivery journey to sound like your brand?
We map the touchpoints, write the brand-voice prompt file and wire the first working automation loop — usually inside a month.
Talk to SMK VietnamFrequently asked questions
Do I need carrier API access to start?
No. GHN, GHTK, J&T Express and Viettel Post all expose tracking APIs or webhooks, but most Vietnamese SMEs start from their commerce platform instead — Haravan, Sapo, KiotViet and the marketplaces already surface fulfilment status. Start with whatever event stream you can reach today and add direct carrier webhooks later.
Should these messages go on Zalo or SMS?
Zalo where you can, SMS as fallback. Zalo OA and ZNS give you richer formatting, buttons and a persistent thread the customer can scroll back through, and they cost less per message at volume. SMS brandname still matters for customers who are not on Zalo or whose OA follow lapsed — run both and let the automation choose.
Is it safe to let AI write messages customers actually receive?
Within constraints, yes. Let the model fill defined variable fields inside an approved template with a hard character limit and a banned-phrase list, and log every send. Do not let it generate compensation offers, legal statements or replies to public complaints. Keep a human approval step on the delay-apology branch for the first two weeks.
How does this work for a foreign brand with no Vietnamese team?
The language layer is the easy part — a well-built brand-voice prompt plus one native reviewer produces reliable Vietnamese. The hard part is register and expectation-setting, which differ from Japanese, Korean or US norms. Have a Vietnamese speaker sign off the prompt file and the first fifty generated messages, then move to spot checks.
What does this cost to run?
The AI is the cheapest line: short transactional messages cost a fraction of a cent each to generate. Your real costs are ZNS message fees, the automation tool subscription, and roughly two weeks of one person’s time to build it. For most SMEs shipping a few thousand orders a month, the reduction in “where is my order” support contacts alone covers it.
Does better delivery communication actually reduce COD refusals?
In our client work it moves in the right direction, most visibly on delayed parcels — a customer who has been told why it is late and when it will arrive is far less likely to refuse at the door. Treat it as a measurable hypothesis rather than a guarantee: baseline your refusal rate for a month before switching the messaging on, then compare.
Leave a Reply