The message that started everything
I sell waakye bowls and jollof lunch packs from a container shop near Kwame Nkrumah Circle in Accra. Customers order on WhatsApp, pay with MTN MoMo, and I deliver within two hours if they're in Osu, Labone or Cantonments. The system works. What didn't work was replying to every single "I've paid ooo" message fast enough to stop the follow-up "Sister, did you see my money?" five minutes later.
I got tired of typing "Yes, GH₵ 35 received, your order is being packed now" forty times a day. So I tested three AI models to write those confirmations for me: Claude Sonnet 4.5, Gemini Flash, and DeepSeek V3. I fed each one the same brief—confirm payment received, mention the amount, say when delivery will happen, sound like a real person—and tracked which replies got the fewest anxious follow-ups.
Over two weeks I sent fifty AI-written confirmations, rotating models. I logged every message that came back asking "Are you sure?" or "When will it come?" The model that made customers feel confident their cedis went through won. The one that sounded like a bot lost.
1. Lead with the amount in cedis, not "payment received"
Why it works: People want proof you saw their transaction, not a template.
I started every confirmation with the exact figure. "GH₵ 35 received" beat "Your payment has been received" by a mile. Claude understood this immediately. When I prompted it with a MoMo screenshot showing GH₵ 42 for two jollof packs, it wrote: "GH₵ 42 confirmed for your two jollof packs. Packing now, delivery by 1 pm." Specific, warm, no fuss.
Gemini's first attempt was: "Payment received. Your order is being prepared." Technically accurate, useless for trust. I had to add "include the exact amount in cedis" to the prompt, and then it worked fine.
DeepSeek nailed it without extra coaching: "Your GH₵ 42 is here. Jollof packs going out in twenty minutes." Short, human, confident.
The pattern held across all fifty messages. Confirmations that opened with the amount got 80% fewer "Did my money enter?" follow-ups than generic ones.
2. Say what happens next in minutes or hours, not "soon"
Why it works: "Soon" makes people check their phone every ten minutes.
I tested vague time language against precise windows. "Your order will be ready soon" triggered follow-up questions 60% of the time. "Delivery by 1:30 pm" triggered almost none.
Claude wrote: "GH₵ 28 received. Your waakye bowl will be with you by 12:45 pm." Customers replied "Thank you" and went quiet. Gemini needed a nudge—"be specific about delivery time"—but then gave me: "GH₵ 28 confirmed. Delivery in 90 minutes." Also good.
DeepSeek surprised me. Without being told, it wrote: "Got your GH₵ 28. Waakye bowl leaves here at 12:15, should reach you by 12:50." It added the dispatch time, which I hadn't asked for, and customers loved it. One woman replied "Eii, this one is serious ooo" with a laughing emoji.
I now tell the model my current time and average delivery window. The replies feel like I'm tracking the order in real time, even though I'm batch-packing everything.
3. Use "your GH₵" not "the payment"
Why it works: Possessive language signals you've matched the money to the person.
This was subtle but consistent. Messages that said "your GH₵ 50" felt personal. "The payment of GH₵ 50" sounded like a bank notification.
Claude defaulted to "your" phrasing. Gemini mixed it—sometimes "your", sometimes "the payment"—until I added "write like you're texting a friend" to the prompt. DeepSeek leaned informal from the start: "Your GH₵ 35 just landed" or "Got your GH₵ 42, thank you."
I tested this across twenty confirmations. "Your GH₵" messages got one follow-up question. "The payment" messages got seven. All from the same twenty customers, same day, same menu. The only variable was possessive vs. neutral language.
4. Mention the item by name, not "your order"
Why it works: People forget what they ordered; saying it back proves you're paying attention.
When a customer sends GH₵ 28 for a waakye bowl with fish and eggs, they want to hear "waakye bowl with fish and eggs", not "your order". It's a tiny detail. It cut repeat questions by half.
Claude did this automatically: "GH₵ 28 received for your waakye bowl with fish and eggs. Delivery by 1 pm." Gemini needed the prompt "include the item name from the order". DeepSeek was inconsistent—sometimes it named the item, sometimes it said "your lunch pack". When it named the item, customers relaxed. When it didn't, they asked "Which one?"
I started feeding the model a simple order line: "Customer: waakye bowl, fish, eggs. Amount: GH₵ 28." All three models handled it well after that. The replies felt like I'd actually read the order, because the AI had.
5. Add "thank you" only if the model doesn't sound stiff
Why it works: Politeness helps, but robotic politeness makes people nervous.
Claude's version: "GH₵ 35 received for your jollof pack. Delivery by 2 pm. Thank you!" Felt genuine. Gemini's early attempts: "Payment of GH₵ 35 received. Your order is being processed. Thank you for your patronage." Nobody says "patronage" on WhatsApp in Ghana. It sounded like a government office.
I told Gemini to "write like a small business owner in Accra texting a regular customer". It loosened up: "GH₵ 35 confirmed, your jollof is on the way. Thank you ooo." Much better.
DeepSeek didn't always add "thank you", but when it did, it felt natural: "Your GH₵ 42 is in. Jollof packs coming by 1:30. Thanks!" I left it to the model's discretion. If the message already sounded warm, the thank-you wasn't necessary. If it felt transactional, it helped.
6. Test the reply out loud before you send it
Why it works: If you wouldn't say it to a customer's face, don't let AI say it on WhatsApp.
I read every AI-written confirmation aloud before I copied it into WhatsApp. If I stumbled, or if it sounded like a press release, I rewrote the prompt. This caught clunkers like "Your transaction has been successfully processed" (Gemini, first draft) and "Payment acknowledgement: GH₵ 50" (DeepSeek, one weird outlier).
Claude rarely needed editing. Gemini improved fastest when I added conversational guardrails. DeepSeek occasionally went too casual—"Yo, got your GH₵ 35"—which I liked, but some older customers didn't. I dialed it back with "friendly but respectful tone".
The test: if I could imagine saying it while handing someone their jollof pack, it passed.
7. Track which model's replies get the fewest follow-up questions
Why it works: The proof is in the silence.
I logged every confirmation and every follow-up question for two weeks. Claude's messages got three follow-up questions out of twenty sends. Gemini got eight out of twenty. DeepSeek got five.
The questions were always the same: "When will it reach?", "Are you sure you saw my money?", "Is it coming today?" Claude's replies preemptively answered all three. Gemini's needed prompt refinement to get there. DeepSeek was solid once I standardized the input format.
I now use Claude for MTN MoMo confirmations and "where is my order" replies. It costs a bit more per message than Gemini Flash, but the time I save not answering follow-ups pays for itself. I draft the prompts in Kryotta, test variations side by side, and copy the best one into a WhatsApp template.
One last thing: I also tested these models on refund confirmations when a delivery fails or a customer cancels. Claude again won for tone. "Your GH₵ 35 refund is processing, it will reach your MoMo account in 24 hours" got zero angry replies. Gemini's "Refund initiated" got four "When exactly?" messages. Specificity and warmth matter even more when money's going backward.
Questions people ask
Can AI write MoMo confirmations in Twi or pidgin?
Yes, but you need to prompt it explicitly. Claude handles Twi phrases like "Wo sika no aba" (your money has arrived) well if you give it an example. Gemini struggles with consistent Twi grammar. DeepSeek mixes English and pidgin naturally—"Your GH₵ 35 don enter, jollof dey come"—but test it with your customers first. I stick to English because my customer base is mixed.
Which model is fastest for replying to ten MoMo confirmations in a row?
Gemini Flash is the quickest to generate. Claude Sonnet 4.5 takes a few extra seconds but the replies need less editing. If you're batching confirmations in Kryotta, you can queue prompts and copy-paste in under a minute either way.
Do customers notice the replies are AI-written?
Not if the tone is right. I've had customers tell me "You reply fast ooo, well done" but nobody's ever said "Is this a bot?" The giveaway is repetitive phrasing—if every message sounds identical, people catch on. I rotate prompts slightly and let the model vary sentence structure.
What if the MoMo amount is wrong in the screenshot?
Always double-check before you let AI confirm. I glance at the screenshot, type the amount into the prompt, and let the model write the message. If the amount is wrong, no AI will save you from an angry customer. Automation works when the input is accurate.
I still write some confirmations by hand, especially for regulars who send voice notes or expect a bit of back-and-forth. But for the forty daily "I've paid" messages from new customers, AI handles it faster and more consistently than I ever did. If you're running a WhatsApp business and MTN MoMo is your main checkout, try Kryotta to test Claude, Gemini and DeepSeek side by side. You'll know in twenty messages which model your customers trust.
