AI Agent for an Ecommerce Store: 5 Chats and What Powers Them

It is 11:40 on a Thursday night. A message lands in the store's Instagram inbox: "Assalam o alaikum, mera order kab tak aayega? Monday ko kiya tha." Nobody replies, so ten minutes later the same customer asks the same thing on WhatsApp. By the time the team opens the phone at 10 am there is a third message: "Cancel kar dein."

Every online store we work with, whether it sells lawn suits in Lahore, perfume in Dubai or trainers in Manchester, gets the same few questions all day and half the night. Where is my parcel. Is this in medium. Can I exchange it. And the angry one: you sent the wrong thing.

An AI agent for an ecommerce store can handle most of these, on one condition: it reads your real order and product data before it answers. An agent that says "aap ka parcel kal tak pohanch jayega" without checking is worse than no agent, because now the customer holds a promise you never made.

So here are five conversations we see in store inboxes constantly. After each: what the agent needs behind the scenes, which rule it follows, and where it stops and calls a person.

The difference between a store agent and a FAQ bot

A FAQ bot has a document: delivery in 3 to 5 days, exchanges within 7 days. It repeats that politely. Fine for "do you deliver to Quetta?", useless for "where is my order?"

An agent has tools. A tool is simply permission to look something up or do something in another system. For a store, the useful ones are:

• Find an order by order number or phone number, from Shopify, WooCommerce or whatever you sell on.
• Check tracking from your courier (TCS, Leopards or PostEx in Pakistan, Aramex in the UAE, Royal Mail in the UK), or from the status your team updates.
• Check stock for a specific product, colour and size.
• Read your policies, written as rules rather than a long paragraph.
• Create a request, like an exchange or a back-in-stock alert.
• Hand over the chat to a named person with a summary.

Our habit is to start agents on read-only tools, and add actions like creating an exchange only after we have read enough of their chats to trust them. Connecting these systems is ordinary API integration work, and it is where most of the effort goes. The AI part is the easy bit.

Chat 1: "Mera parcel kahan hai?"

Customer: Order 4821 kahan hai? 5 din ho gaye.

Agent: Ji, order 4821 (2 kurtas, size M) 9 October ko Leopards ke saath dispatch hua tha. Tracking ke mutabiq parcel kal Faisalabad pohanch gaya aur aaj delivery ke liye nikla hai. Tracking link ye hai. Agar aaj na pohanchay to yahin bata dein, team courier se khud baat karegi.

Behind the scenes: the agent looked up order 4821, checked that the WhatsApp number matches the phone on the order, and pulled the latest courier scan. That number check matters: someone writing from another number with your order ID can get the status, but not the address or items.

The rule: the agent only repeats what tracking says. If tracking has not moved for, say, three days, it does not invent "Eid rush" or "weather". It promises to check with the courier and flags the order for a person.

If you sell on cash on delivery, the confirmation before dispatch is a separate flow, covered in COD order confirmation on WhatsApp. This agent handles what comes after.

Chat 2: "Bhai ye medium mein available hai?"

Customer: bhai ye medium mein available hai? (sends a screenshot of an Instagram post)

Agent: Ji, ye Olive Lawn 3-piece hai na? Medium mein abhi available hai. Medium ka chest 40 inch aur length 42 inch hai. Order ka link ye hai.

Customer: black mein?

Agent: Black mein medium khatam ho gaya hai, sirf large bacha hai. Medium dobara kab aayega ye abhi confirm nahi. Agar aap chahein to stock aate hi main aap ko yahin message kar doon?

Behind the scenes: it matched the screenshot to a product, then asked "ye wala hai na?" because image matching can be wrong when you post five similar prints a week. It checked stock by colour and size, not just "product in stock". And it read the size chart for that exact product. One size chart for the whole store is how customers end up exchanging.

The rule: never promise a restock date that is not in your data. Offering a back-in-stock message is better, and the customer saying "haan" is a clear opt-in you can use later.

The mistake we see most: stock that syncs once a day while the same pieces also sell in a physical shop. The agent says medium is available, the customer pays, and the last piece went at 4 pm. If stock is not live, the agent should say "abhi available dikh raha hai, team confirm karegi" instead of a flat yes.

Chat 3: an exchange request on day six, as a voice note

Customer: (a 40 second voice note) Shoes 42 size tight hain, 43 chahiye. 6 din pehle mile thay.

Agent: Ji, aap ka order 5130 hai, jo 8 October ko deliver hua. Exchange 7 din ke andar hota hai, to aap ke paas kal tak ka time hai. Size 43 is design mein available hai. Shoes unworn hon aur box ke saath hon. Kya main exchange request bana doon?

Behind the scenes: the agent counted from the delivered date in the courier data, not the order date. It checked that size 43 exists before offering it. And it read your exchange policy as rules: days, condition, who pays return delivery, and excluded items like sale pieces or undergarments.

The voice note is the part people forget to test. On one of our setups, speech-to-text wrote Urdu voice notes in Hindi script, so the AI assumed Hindi and replied in Hindi. Forcing Urdu transcription fixed it. Test with real voice notes, background noise and regional accents included, before launch.

Where it hands over: day nine, a worn item, or "pehle bhi exchange kiya tha, ab bhi karo". The agent explains the policy once, kindly, then passes the chat on. Exceptions are a business decision, never the agent's.

Chat 4: the cart left at 1 am

Store (approved template, sent at 11 the next morning): Hi Sana, aap ke cart mein Olive Lawn 3-piece (M) abhi bhi saved hai. Checkout yahan se complete karein. Messages band karne ke liye STOP likhein.

Customer: Karachi delivery kitne ki hai?

Agent: Karachi ke liye delivery 250 rupay hai, aur 5,000 se upar free. Aap ka cart 4,800 ka hai, ek aur item add karein to delivery free ho jayegi. Size M abhi available hai.

Behind the scenes: this one has the most rules, and most of them are Meta's, not yours.

• Opt-in first. The customer must have agreed to WhatsApp messages, usually a checkbox at checkout. A phone number typed for delivery is not permission to market to it.
• The 24-hour rule. If the customer has not messaged you in the last 24 hours, your first message must be a paid template that Meta has approved. Once she replies, a 24-hour window opens and the agent can answer freely. Store owners are often surprised by this the first time they open the dashboard and try to message a list of old customers.
• Category and cost. Meta usually treats cart reminders as marketing messages, which cost more than order updates. Meta sets and changes these prices, so check the current rate for your country.
• One reminder, maybe two. Five reminders for one cart is how customers start pressing "report", and that hurts your number. More on that in WhatsApp automation without bans.

Notice the agent did not invent a discount. It pointed out the free delivery threshold, which is your real policy. Want a code in the reminder? Write it into the template yourself.

Chat 5: "Wrong item bheja hai, paise wapis karo"

Customer: Ye kya bheja hai aap ne? Maine blue order kiya tha, green aaya hai. Mujhe refund chahiye.

Agent: Maafi chahta hoon, ye theek nahi hua. Order 5214 mein blue (L) hi likha hai, to ghalti hamari taraf se lagti hai. Kya aap parcel ki ek photo bhej sakti hain jis mein label nazar aaye? Main ye chat abhi Hamza ko de raha hoon, refund ya exchange woh aap ke saath khud final karenge. Aaj 6 baje tak woh aap ko reply kar denge.

Behind the scenes: the agent spotted the triggers (refund, wrong item, angry tone) and switched from answering to collecting. It confirmed what was ordered, asked for proof, and passed the chat to a person with a summary: order number, what was ordered, what arrived, what the customer wants.

The rule: the agent never promises a refund, an amount or a date for the money. It can apologise and say who will handle it and when. That "6 baje tak" must come from your real working hours, not hope.

Honestly, this is where stores keep a customer or lose one in public. A bot that replies "please read our return policy" to an angry customer ends up as a screenshot on Instagram. Damaged items, wrong items, missing parcels, double payments: those go to a human, every time, fast.

Test it on last month's chats before a customer meets it

Everything above depends on correct data and written rules. Before any agent goes live, this is what we do, and you can start most of it today:

1. Pull 50 real chats from last month, across WhatsApp and Instagram, and sort them into the five types above. Whatever is left over tells you what else the agent needs.
2. Write your exchange policy as short rules: days, condition, who pays delivery, exclusions. If two staff members answer the same question differently today, settle that first.
3. Check that every product has stock by colour and size, and its own size chart.
4. Test with messy Roman Urdu ("kitne ka", "ktny ka", "price?"), screenshots and real voice notes.
5. Try to break it: ask for a discount, ask about an order from a number that did not place it, ask about a parcel stuck for a week.
6. Go live on one channel and read every conversation for the first week. When an answer is wrong, fix the data first.

What we connect for stores, and an easy first step

When we set up an AI agent for a store, it runs on the official WhatsApp Business Platform through Meta's Cloud API, connected to your store and courier data. With WhatsApp Coexistence you keep the WhatsApp Business App on your phone and can jump into any chat, while the agent replies around the clock in Urdu, Roman Urdu or English and hands over when the rules say so. Plans start from PKR 12,000/month, Meta's own message charges are separate, and the Voice Replies add-on (PKR 2,500/month for 400 voice replies) helps if your customers live on voice notes.

An easy start: write down the five questions your team answered most this week and send them to us on WhatsApp at 0339 4285556, or through our contact page. We will tell you which ones an agent can safely take, and what data it needs to answer without guessing.

Further reading

See how this looks in practice across our AI automation, custom CRM and lead generation services, or tell us the outcome you need.