The week I had 487 unread DMs and no intern
I run an Instagram shop selling womenswear: ankara dresses, ready-to-wear skirts, a few bags. About 3,000 followers, mostly Lagos and Abuja. In March I posted a flash sale—15% off everything, 48 hours only—and by hour three I had 487 unread DMs. Questions about sizes, bank details, whether I deliver to Ajah, whether the green dress was still in stock, whether I take payment on delivery. I was typing the same answers over and over: "Yes, we deliver to Lekki but Mainland takes two extra days", "Account number is 0123456789, Guaranty Trust Bank", "₦18,700 after discount."
I'd heard people talk about using AI to reply to customer messages, but I'd assumed it would sound robotic or mess up the details. Still, I was drowning. So I opened Kryotta, tested three models—Claude Sonnet 4.5, Gemini Flash, and Llama 3.3 70B—and spent one week letting AI draft replies while I approved and sent them. I handled 120 DMs that week. Some models understood Nigerian commerce. Others hallucinated prices or suggested courier services that don't exist here.
Here's what actually worked, what failed, and which model I'm still using two months later.
The test: same questions, three models, real customer expectations
I picked ten common DM types: price inquiries, delivery questions (Lekki vs Mainland vs outside Lagos), bank transfer details, stock checks, size requests, "I'll pay on delivery" negotiations, WhatsApp handoff requests, bulk discounts, return policy, and the classic "Is this available?" with no other context.
For each model I wrote a simple instruction: "You're a customer service assistant for an Instagram fashion vendor in Lagos. Reply to this DM professionally but warmly. Use naira for prices, mention our GTBank account for transfers, and offer WhatsApp contact for complex orders. Delivery to Lekki takes 1–2 days, Mainland 2–3 days, outside Lagos 4–5 days. We don't do payment on delivery." Then I pasted a real customer message and compared the outputs.
I wanted to see which model formatted naira correctly (₦18,700 not N18700 or 18,700 naira), understood Lagos logistics without inventing courier names, and kept the tone right: helpful, not stiff, not overly casual.
Claude Sonnet 4.5: got the context, occasionally too formal
Claude understood Nigerian commerce better than I expected. When someone asked "How much is the blue dress in naira?", Claude replied "₦18,700 after the 15% discount" and formatted the currency symbol correctly every time. It knew to mention bank transfer first, card second. It never suggested cash on delivery when I'd told it we don't offer that.
The delivery answers were solid. A customer asked if we deliver to Ajah; Claude said "Yes, Ajah is within our Lekki delivery zone—1 to 2 days once payment is confirmed." Accurate, specific, no invented logistics companies. When someone asked about Ibadan, Claude said "We deliver outside Lagos in 4–5 days; shipping is ₦2,500 extra"—I hadn't given it that shipping fee, so I added it to the prompt and it stayed consistent after that.
The weakness: tone. Claude's replies were polite but sometimes sounded like a corporate email. A customer wrote "Sis abeg the dress still dey?" and Claude replied "Yes, the green ankara dress is still available. Would you like to proceed with your order?" Correct, but a bit stiff. I wanted something closer to "Yes sis, it's still available! Should I send you the account details?" I edited about 30% of Claude's replies to make them warmer, but the facts were always right.
Gemini Flash: fast, but invented a courier service that doesn't exist
Gemini was noticeably faster—replies appeared in two or three seconds—and the tone was more conversational. When someone asked about sizes, Gemini said "We have UK 10, 12, and 14 left in that style—let me know which one works for you!" Natural, friendly, no complaints there.
But Gemini hallucinated logistics. A customer asked how we deliver to Ikorodu, and Gemini replied "We use GIG Logistics for Mainland deliveries, usually 2–3 days." We don't use GIG Logistics. I use a local dispatch rider and occasionally a bike courier. I went back to the prompt, added "Do not mention specific courier names; just say 'our dispatch rider' or 'local courier,'" and Gemini stopped inventing companies. But I had to catch that mistake before it went to a customer who might have called GIG and been confused.
Currency formatting was inconsistent. Sometimes Gemini wrote ₦18,700, sometimes "18,700 naira," once "NGN 18700" with no comma. I couldn't find a pattern. I added "Always format prices as ₦18,700 with the naira symbol and comma" to the prompt, which helped but didn't fix it completely. About one in every eight replies still dropped the symbol.
Gemini also struggled with conditional logic. A customer wrote "I want two dresses, do I get a discount?" and Gemini said "Yes, we offer 10% off for orders over ₦30,000"—a discount structure I'd never mentioned. I had to add "We do not offer bulk discounts beyond the current sale price" to stop it promising things I couldn't deliver.
Llama 3.3 70B: understood the vibe, terrible with numbers
Llama surprised me with tone. It sounded the most human: casual, warm, exactly how I'd reply to a friend. A customer asked "You dey deliver Surulere?" and Llama said "Yes now! Surulere is Mainland, so 2–3 days once you pay. I go send account details." Perfect. I barely edited those replies.
But Llama couldn't handle numbers reliably. A customer asked for the price of a dress originally ₦22,000 with 15% off. Llama said ₦18,500. The correct answer is ₦18,700. Not a huge error, but enough to confuse someone who'd done the maths themselves. I tested it on three other price calculations and it got two wrong—once rounding down, once forgetting the discount entirely and quoting the original price.
Delivery fees were also inconsistent. I'd told Llama that outside-Lagos shipping is ₦2,500, and it repeated that correctly four times, then told a customer in Port Harcourt the fee was ₦3,000. I couldn't work out why. I added "Shipping outside Lagos is always ₦2,500, no exceptions" and it improved, but I still checked every reply that mentioned a number.
Llama also didn't understand "I'll pay on delivery" negotiations. A customer wrote "I fit pay when e reach?" and Llama said "Sure, we can arrange that"—directly contradicting my policy. I had to add "We only accept payment before dispatch: bank transfer or card via Paystack" and repeat it twice in the prompt before Llama stopped offering payment on delivery.
What I actually use now: Claude for facts, Llama for tone, me for the final read
I don't use one model for everything. If the question is about price, stock, or delivery logistics, I use Claude Sonnet 4.5. It gets the details right and I only edit for warmth. If the question is casual or emotional—someone asking if a dress will suit them, or apologising for a delayed payment—I use Llama 3.3 70B because the tone is perfect. I stopped using Gemini after the second week; the hallucinations weren't worth the speed.
I still read every reply before I send it. That takes about 15 seconds per message, but it's faster than typing from scratch and I catch the occasional mistake—Llama quoting the wrong price, Claude sounding too formal, a customer using slang the model didn't parse. Over the week I handled 120 DMs in about four hours total. Before AI, that same volume would have taken me twelve hours and I'd have missed some messages entirely.
One unexpected benefit: consistency. I used to give slightly different delivery estimates depending on my mood or how busy I was. Now every customer in Lekki hears "1–2 days once payment is confirmed" and every Mainland customer hears "2–3 days." That's reduced follow-up questions and complaints.
The prompts that worked (and the ones that didn't)
The instruction I started with was too vague: "Reply to this customer professionally." That got me generic, overly polite answers that didn't sound like my brand. The version that worked was much more specific:
"You're the owner of an Instagram fashion shop in Lagos. Reply to this DM warmly and conversationally, like you're chatting to a friend but still professional. Use Nigerian English and Pidgin where natural. Always format prices as ₦18,700 with the symbol and comma. For delivery, say: Lekki 1–2 days, Mainland 2–3 days, outside Lagos 4–5 days with ₦2,500 shipping. Payment by bank transfer (GTBank 0123456789, Account Name: [Your Name]) or card via Paystack. No payment on delivery. If the question is complex, offer to continue on WhatsApp: 0801 234 5678."
I also added a line that helped with tone: "If the customer uses Pidgin or informal language, match their energy. If they're formal, stay polite but friendly." That stopped Claude from sounding like a bank and stopped Llama from being too casual with customers who clearly wanted a professional reply.
What didn't work: telling the model to "be creative" or "personalise the reply." That led to Gemini inventing discounts and Llama making assumptions about what the customer wanted. I got better results with strict instructions and room for tone variation within those limits.
The mistakes I'm glad I caught before customers saw them
Gemini told a customer we accept Opay transfers. We don't; I only listed GTBank and Paystack. The customer would have sent money to the wrong account and then messaged me confused.
Llama once replied to a price inquiry with "₦18,700 or $15 if you're paying in dollars"—I have no idea where it got that. I don't take dollar payments and the exchange rate was wrong anyway.
Claude told a customer in Abuja that delivery would take 4–5 days, which is correct, but added "via our Abuja dispatch partner"—implying I have a dedicated partner there, which I don't. It's just a courier service I book per order. Small detail, but it set an expectation I couldn't meet.
All three models occasionally misread "Is this available?" as a stock question when the customer was actually asking if we deliver to their area. I learned to check the context of the previous message before approving a reply.
Questions people ask
Can AI actually understand Nigerian English and Pidgin?
Llama handled it best, Claude understood it but replied more formally, and Gemini sometimes missed the meaning entirely. You'll still need to review replies if your customers use a lot of slang, but it's faster than starting from scratch.
Will customers know I'm using AI?
Not if you edit for tone. The facts come from AI, but the warmth comes from you reading it and tweaking anything that sounds off. I've had zero complaints or comments about robotic replies.
Does this work for WhatsApp Business too?
Yes, same approach. I paste the WhatsApp message into Kryotta, get a draft reply, edit it, and paste it back. WhatsApp doesn't have an API integration in Kryotta yet, so it's manual, but still much faster than typing everything yourself.
Which model is cheapest for high-volume DMs?
Gemini Flash is fastest and uses the least budget if you're on a paid plan, but you'll spend more time fixing mistakes. Claude costs a bit more per reply but needs less editing. I'd rather pay an extra few kobo per message and save the time.
I'm not saying AI will run your Instagram shop for you. You still need to check stock, process payments, and make sure the reply actually answers the question. But if you're spending two hours a day typing the same answers about delivery times and bank details, this will get you most of the way there. I've gone from dreading DMs to handling them in one focused session each evening, and I haven't missed a sale because I replied too slowly.
Try it with ten messages and see if it saves you time. If it does, you'll know within an hour. If it doesn't, you've lost nothing but the time you'd have spent typing those replies anyway. Start with Kryotta's free tier and test Claude or Llama on your next batch of DMs—you'll see which one understands your customers faster than I can explain it.
