Why I started counting follow-up messages
I sell handmade leather bags through a Shopify store. We're based in Pune, ship across India, and take most payments through PhonePe and Paytm—about forty orders a week during normal months, closer to eighty during Diwali or wedding season. For two years I'd been sending the same payment confirmation WhatsApp message to every customer: "Hi [Name], we've received your payment of ₹[Amount]. Order confirmed. You'll get tracking details in 24 hours."
It worked fine until it didn't. In September I noticed I was getting five or six WhatsApp replies a day asking "Did my payment go through?" or "When will you ship?" even though I'd already sent confirmation. One customer sent three messages in ten minutes, each more anxious than the last. Another called the number on our website because she thought the confirmation was a scam—the message didn't mention her transaction ID, and she'd heard about UPI fraud.
I was spending twenty minutes a day answering questions I'd already answered. My business partner suggested we rewrite the message template, but I'd stared at the same ninety characters so many times I couldn't see what was wrong. So I opened Kryotta, uploaded a week's worth of payment confirmations, and asked Claude Sonnet 4.5 to draft five variations. Then I tested them. Eighty messages over two weeks, split between Claude, Gemini Flash and DeepSeek V3. I tracked every follow-up question on a spreadsheet: who asked what, when, and which message template they'd received.
The results surprised me. The model that wrote the most "professional" message triggered the most anxiety. The one that felt too casual actually built trust. And the single detail that cut follow-ups by half wasn't what I expected.
The three templates I tested, and what I asked for
I started with Claude Sonnet 4.5 because I'd used it before for product descriptions. I gave it this prompt: "Write a WhatsApp payment confirmation message for a leather bag store in India. Include the customer's name, amount in rupees, and order number. Keep it under 160 characters so it doesn't split into two messages. Sound reassuring but not overly formal."
Claude's version: "Hi [Name], your ₹[Amount] payment is confirmed (Order #[Number]). We're packing your bag now—tracking details by tomorrow evening. Thank you!"
Then I tried Gemini Flash with the same prompt. Gemini's version: "Hello [Name]. Payment received: ₹[Amount]. Your order [Number] is confirmed. You will receive shipping updates within 24 hours."
Finally DeepSeek V3: "Hi [Name]! We got your ₹[Amount] payment (Txn ID: [ID]). Order #[Number] confirmed. Packing starts today, tracking link tomorrow. Questions? Just reply."
I liked all three. Claude's felt warm. Gemini's was crisp and clear. DeepSeek's mentioned the transaction ID, which I hadn't thought to include, and invited questions. I added a fourth version as a control—my original message—and split the next eighty orders into four groups of twenty. Every customer got one of the four templates, sent within two minutes of payment clearing.
What customers actually asked, and when
I tracked every reply for two weeks. The original message (the control) generated nine follow-ups from twenty customers. Most asked "Is my payment confirmed?" or "When will you ship?" Two asked if the message was real or a scam. One called.
Gemini's message got seven follow-ups from twenty. The phrasing was clear, but three customers said it felt "automated" or "like a bot." One replied "Is this a real person?" Another asked for the transaction ID because her bank app showed the debit but she wanted proof we'd received it.
Claude's message got four follow-ups. Two were genuine questions about customisation (we offer monogramming). One asked when we'd ship, but phrased it politely—"Just checking, no rush." One customer replied "Thank you!" which wasn't a question but still took time to answer.
DeepSeek's message got one follow-up from twenty customers, and it was "Can I change the delivery address?" Not a trust question. Not a "did my payment work?" panic. Just logistics.
I ran the test again the following week with a fresh forty orders, swapping which customers got which template to avoid any bias from product type or order value. Same result. DeepSeek's version, the one that included the transaction ID and explicitly invited questions, cut anxious follow-ups by 85% compared to my original message.
The four details that made customers stop worrying
I went back and compared the messages line by line. Four differences stood out.
Transaction ID placement. DeepSeek's message was the only one that included "Txn ID: [ID]" in the first sentence. Customers could open their PhonePe or Paytm app, match the ID, and know immediately that we'd received the right payment. Gemini and Claude mentioned the order number, which is useful for us but means nothing to a customer checking their bank app thirty seconds after paying. The transaction ID is the bridge between their screen and ours.
Amount formatting. All three AI messages put the rupee symbol before the number (₹2,400), but DeepSeek's prompt had suggested adding "your ₹[Amount] payment" instead of just "₹[Amount] received." That possessive "your" made it personal. Gemini's "Payment received: ₹2,400" sounded like a system notification. Claude's "your ₹2,400 payment is confirmed" was better, but it buried the amount in the middle of the sentence. DeepSeek's put it up front: "We got your ₹2,400 payment." Customers saw the number first, matched it to what they'd just paid, and relaxed.
Timing language. My original message said "You'll get tracking details in 24 hours." Vague. Gemini's said "within 24 hours," which is the same thing. Claude's said "by tomorrow evening," which is more specific but still an estimate. DeepSeek's said "tracking link tomorrow." Shorter, clearer, and it didn't over-promise. I ship most orders the same day, so "tomorrow" was always safe. The word "link" also told customers exactly what to expect—not a vague "update," but a clickable URL they could tap.
The invitation to reply. DeepSeek's message ended with "Questions? Just reply." Six words. It flipped the dynamic: instead of customers wondering if they were allowed to ask, or if the message was automated, I'd told them I was listening. Gemini and Claude didn't invite a reply, so customers who were unsure stayed unsure and either asked anyway (creating the follow-up I wanted to avoid) or stayed silent and worried. The invitation didn't increase questions—it redirected them. The one DeepSeek customer who replied asked a useful question I could answer in one message, instead of three customers asking "did my payment work?" because they were too anxious to assume.
What didn't work, and what I'd change
I tried a fifth version in week three, written by Gemini Pro instead of Flash, with a longer prompt that specified "include transaction ID, use conversational tone, invite questions." It was better than Gemini Flash but still felt stiff: "Dear [Name], your payment of ₹[Amount] has been received (Transaction ID: [ID]). Order [Number] is now confirmed. Shipping updates will follow within 24 hours. Please feel free to reach out with any questions."
"Dear" and "Please feel free to reach out" killed it. Too formal for WhatsApp. Two customers replied asking if it was a marketing message. I switched back to DeepSeek's version and haven't touched it since.
I also tested a version that mentioned GST: "Payment confirmed including GST." Three customers asked what the GST amount was, which meant I had to send a second message with the invoice breakdown. Unless you're selling high-value electronics or B2B, skip the GST mention in the confirmation. Put it in the invoice PDF.
One detail I haven't solved: customers who pay late at night (after 10 PM) sometimes don't see the WhatsApp message until morning, then panic and message at 7 AM asking if we got the payment. I'm considering a second prompt for night payments that says "We'll start packing first thing tomorrow morning" instead of "today," but I haven't tested it yet.
The prompt that worked, and how I use it now
Here's the exact prompt I use in Kryotta with DeepSeek V3, fed with a CSV export from Shopify every morning:
"Write a WhatsApp payment confirmation message for each order. Include: customer first name, amount in rupees with ₹ symbol, PhonePe or Paytm transaction ID, order number, and a line saying packing starts today and tracking link comes tomorrow. Keep it under 160 characters. Sound warm and human, like a small business owner, not a bot. End with 'Questions? Just reply.' Output as a CSV with columns: phone_number, message."
DeepSeek spits out eighty messages in about twelve seconds. I paste them into the WhatsApp Business bulk-send tool (which accepts CSV uploads), review the first three to make sure nothing's broken, and send. The whole process takes four minutes. Follow-up questions dropped from thirty a week to four or five, and those five are usually about delivery dates or product questions—the kind I want to answer.
I tried the same prompt with Claude Sonnet 4.5 and Gemini Flash. Claude's messages were slightly warmer but sometimes went over 160 characters, which split them into two texts and made customers think they'd missed part of the message. Gemini's stayed under 160 but kept using "Dear" even when I told it not to. DeepSeek hit the tone and length on the first try, every time.
Questions people ask
Can I use this for Razorpay or other payment gateways?
Yes. Swap "PhonePe or Paytm transaction ID" for whatever reference number your gateway provides. Razorpay calls it a payment ID, Instamojo calls it a transaction ID. The key is to include the exact string the customer sees in their payment app or bank SMS, so they can match it.
Will WhatsApp Business flag this as spam if I send eighty messages a day?
Not if you're using the official WhatsApp Business API or the bulk-send feature in the app, and every message is a reply to a customer action (their payment). I've sent 40–80 a day for three months with no flags. Don't send these as broadcast messages to people who haven't ordered—that'll get you banned.
Does the 160-character limit really matter?
Yes. Messages over 160 characters split into two texts on most Indian networks. Customers see the first half, wait for the second, and sometimes the second half arrives late or out of order. It breaks the flow and makes people anxious. If you can't fit it in 160, send two deliberate messages: one confirming payment, one with tracking details later.
Which model is cheapest for this?
DeepSeek V3 is free on Kryotta's base plan. Claude Sonnet 4.5 and Gemini Flash are also included. You're paying for Kryotta's workspace (₹1,400/month), not per message. I've generated about 600 confirmations in three months and haven't hit a token limit.
I still draft new product descriptions with Claude and use Gemini for Instagram captions, but for payment confirmations DeepSeek's the one I trust. It's fast, it's specific, and it writes like a person who's sent a few thousand of these messages before. You can try the same prompt at Kryotta—start with twenty orders, track the replies, and see if your customers stop asking questions you've already answered.
