WhatsApp AI agents for e-commerce in the UAE and Saudi Arabia: start with the order, then earn the conversation
In June 2026, Trendyol's CEO told Meta's Conversations audience that its WhatsApp programme had started "in Saudi Arabia two weeks ago at scale". The launch was not the interesting part. How Trendyol built it was: basic one-way messages first, such as transaction received and delivery today, and only then a conversation. This page sets out that approach, the published numbers and the two recurring problems Trendyol found in failed AI projects. It ends with how Learnmind builds this for UAE retailers, including those selling into Saudi Arabia.
The short answer
Trendyol's older channels were fading. Push, SMS and email performance was "declining ... very fast", in its CEO's words.
WhatsApp reached people the other channels did not. It brought in "a different set of users", distinct from the CRM audiences Trendyol already had.
It started with the most basic messages. Order received, delivery today, a suggestion to go with the purchase. Conversation was added on top.
Meta's case study: more than 123,000 customers reached in one week through marketing messages and the agent, product recommendations in 13% of conversations, and a median response under seven seconds. Self-reported.
Failed AI projects shared two recurring problems: data quality and bad process. Neither is fixed by a better model.
Why a retailer with 40 million customers moved to WhatsApp
Trendyol is Turkey's leading e-commerce company. Derya Matras, who leads Meta's business in Europe, the Middle East and Africa, described it as one of the fastest-growing marketplaces in the Middle East. Its CEO, Erdem Inan, explained the problem. An average customer shops about 25 times a year, so the business constantly needs new reasons to be in touch.
The traditional channels kept declining. Trendyol tried more creative journeys and more interesting content on push, SMS and email, and "the performance was just dropping down."
WhatsApp worked differently because customers were "already willing to talk with us". Trendyol was also careful with it. In Inan's words: "It's an intimate channel, you know, we don't want to abuse it too much."
That caution is worth copying. A channel people use for family and friends tolerates far less noise than an inbox full of promotions.
The build order: basic messages first, conversation on top
Inan described the sequence plainly: "We started from the most basic, no-brainer things, and then added on top of each of them."
- Basic one-way messages"Your transaction is received" and "your delivery is coming today". Expected, useful, and the first messages in the thread.
- A suggestion tied to the purchaseInan listed it with the basic messages: along with the shoes, consider this and this product.
- The conversational layer, added on topThe customer replies and an agent answers, in seconds, from the catalogue.
- Recommendations with a checkout linkMeta's case study describes shoppable carousels, each with a one-tap link to checkout.
Each layer sits on messages the customer already wanted. By the time a recommendation arrives, the thread has a history of being worth opening.
The trap in copying step two
A delivery update is a utility message. A delivery update with a product suggestion in it is not.
Meta's rule: templates with mixed content, such as "an order update with a promo", are marketing templates. If you file one as utility and WhatsApp decides it is marketing, it is approved as marketing and billed as marketing.
UAE consumer law adds a second layer. Suppliers may not use consumer data for marketing purposes, so a customer who gave you a number to receive a parcel has not agreed to promotions.
The clean version: keep order and delivery updates free of promotion. Make a suggestion only when the customer asks for one, and send promotions only to customers who opted in to them.
Our page on WhatsApp for UAE e-commerce covers the full order message set, cash on delivery and returns. This page is about what an agent adds on top.
What the published numbers show, and what they do not
Meta's July 2026 report files Trendyol's case under retail and e-commerce in EMEA. It quotes Erk Kalender from Trendyol's CRM team: the intelligence of an agent connected to Trendyol's e-commerce systems "now lives inside the conversation".
| Reported | Figure | What it tells a Gulf retailer |
|---|---|---|
| Customers reached in one week, by marketing messages and the agent | More than 123,000 | This is reach, not agent performance |
| Conversations with personalised recommendations | 13% | About one conversation in eight produced a recommendation |
| Median response time | Under seven seconds | A benchmark for reply speed |
Two cautions. The 123,000 reach figure counts marketing messages and the agent together, so it does not isolate the agent's effect. And Meta marks every case in the report as self-reported and "not identifiably repeatable". The Saudi launch was two weeks old when Inan spoke, and Inan said it plainly: there is "a lot to test, a lot to fail."
Discovery is the other job an agent does
Deliveroo, which operates in the UAE, sat on the same panel. Its VP of Engineering, Vaughn Washington, described the difficulty of a big catalogue. Open the app and "you would see hundreds and hundreds of options". One of the hardest things, according to Washington, is not getting a hungry customer to open the app. It is "helping them navigate all that content".
Deliveroo is building towards an agentic experience in WhatsApp that lets customers "really quickly zone in" on what they want. For a Gulf retailer the same job looks like narrowing questions.
| Retailer (illustrative) | What the customer says | What the agent narrows on |
|---|---|---|
| Fragrance house with 400 scents | "Something for my husband, not too sweet" | Scent family, strength, budget, gift wrapping, delivery date |
| Modest fashion boutique | "An abaya for a wedding next week" | Size, colour, fabric, budget, delivery in time |
| Electronics store | A photo of a socket: "Please give me a cable that's gonna fit in this thing" | The socket in the photo, the device, what is in stock |
The last row is not hypothetical. Motorway's Tanya Cordrey described wanting to photograph a socket and WhatsApp a seller for the right cable. Meta's Alex Schultz replied that he already does exactly that, for camera cables.
Merchants need agents too
Trendyol's other agent works for its sellers. About 90% of its nearly 300,000 merchants are SMEs, often one or two people doing everything. Inan said an average merchant spends 40 to 45 hours a week on tasks such as answering questions, joining promotions, creating coupons and running advertising.
The project was called Ortak, which means partner in Turkish. It suggests actions, and "if the merchant says go, it just executes on behalf of" the merchant. Inan described it as a full-time employee working around the clock. Inside the company, Trendyol runs about 1,000 agents on different tasks.
For a UAE brand, the parallel is the back office. Trendyol said most of what its 75,000 influencers ask "is very trivial" and could be handled by an agent. The same is often true of the routine questions a brand gets from couriers and wholesale buyers.
Why AI projects fail, in Trendyol's words
Asked for one piece of advice, Inan described a pattern in Trendyol's failed AI projects. The team was not "AI-ready".
- Data quality. "The data should be there, consistent with a good time horizon." An agent recommending from a catalogue with missing sizes, stale stock or inconsistent names will recommend badly.
- Bad process. "In most cases, the process was bad." Automating a broken returns flow gives you a faster broken returns flow.
Before a retail agent goes live, check both, item by item.
- The product feedEvery live product has a name in Arabic and English, a price, stock, sizes or variants, and an image.
- The order status sourceOne system the agent can ask where an order is, with courier updates flowing into it.
- The written policiesDelivery times by emirate and country, returns, exchanges and cash on delivery rules.
- The handoverA named person for complaints, damaged items, refunds and anything the policy does not cover.
- The consent recordWho has opted in to promotions, kept separately from who has simply bought something.
What changes for the UAE and Saudi Arabia
- Two languages in one thread. Customers switch between Arabic and English, sometimes mid-sentence. Product names need both.
- Cross-border delivery questions. A UAE retailer selling into Saudi Arabia gets asked about delivery times and duties for each market. Those answers must come from written rules.
- Different data regimes. The UAE has Federal Decree-Law No. 45 of 2021. Saudi Arabia has its own Personal Data Protection Law, fully enforceable by SDAIA since September 2024. Selling into both means meeting both.
- Consent is not transferable. A purchase is not permission to market, and permission in one channel is not a list for another.
What Learnmind builds for Gulf retailers
We build WhatsApp agents for retailers in the UAE, including those selling into Saudi Arabia and the wider Gulf, on the official WhatsApp Business Platform and your own verified number.
- Order and delivery updates that invite a reply, kept clean of promotion so they stay utility messages.
- An agent that answers the reply in seconds, in Arabic or English, from your catalogue and order data once they are connected.
- Narrowing questions and recommendations when the customer asks, with a link to buy.
- A handover to your team for returns decisions, complaints and anything with a cost attached.
- Every conversation recorded in your AI CRM, with promotional consent tracked separately.
Our recommended order is Trendyol's: the messages customers already want first, conversation on top.
Want order updates that turn into conversations?
We build WhatsApp agents for UAE retailers selling across the Gulf. They answer from your catalogue and order data, in Arabic and English, and keep promotion and service where they belong.
Frequently asked questions
Can a WhatsApp agent recommend products from my catalogue?
Yes, if it is connected to a clean product feed with stock and prices. Recommendations are only as good
as the feed behind them.
Should I add product suggestions to my delivery updates?
No. Meta treats an order update with a promotion as a marketing template, and UAE law restricts using
purchase data for marketing. Make suggestions when a customer asks for them, and send promotions only to customers who opted in.
Does the agent work in Arabic?
Yes. Arabic and English on one number, including customers who mix the two.
What should I automate first?
Order confirmation and delivery updates that invite a reply. Trendyol started with basic messages like
these and added conversation on top.
What does WhatsApp messaging cost in the UAE?
It depends on the message category and your volumes. Our UAE
pricing calculator works it out from your own numbers.
How we checked this, and what we could not settle
Checked: Trendyol's and Deliveroo's remarks come from the transcript of Meta's panel "Building the Customer Journey of the Future", recorded at Conversations in London in June 2026. Tanya Cordrey's cable example comes from the session "The AI Shift: The New Era of Customer Engagement". Trendyol's 123,000, 13% and seven-second figures come from Meta's July 2026 report. The template rule comes from Meta's template categorisation documentation, read 22 September 2026. Deliveroo's UAE presence was checked against current reporting.
Not settled: Meta's Trendyol figures are self-reported, and the report does not say which countries they cover. We have no verified Gulf benchmark for how many WhatsApp conversations end in a sale, and we will not quote one. Measure your own before and after.
Sources
- PlatformMeta, AI Agents and the Future of Customer Engagement, including the panel "Building the Customer Journey of the Future"
- PlatformMeta, Beyond chatbots: The agentic economy is here, July 2026, Trendyol case study
- PlatformMeta, template categorisation, utility and marketing, and mixed content
- LawUAE Government portal, consumer protection, Federal Law 15/2020
- LawUAE Government portal, data protection laws, Federal Decree-Law 45/2021
- ReferenceMorgan Lewis, Saudi Arabia Personal Data Protection Law: transition period ends 14 September, 2024
Written by Edmund Gay, Learnmind.ai, Dubai. This page is commercial information about a market we operate in, not legal advice. Platform rules and figures carry the date we checked them, and both change.