I sell phone accessories on Jumia. Cases, chargers, screen protectors, Bluetooth speakers. I ship from a lockup in Kaneshie, handle about two hundred orders a month, and every week I get fifteen to twenty return requests. Some are legitimate—wrong size, damaged in transit, late delivery. Some are buyer's remorse dressed up as a defect claim. Jumia gives you forty-eight hours to respond with a decision and evidence, and if you approve the return without asking the right questions, you eat the shipping cost and the restocking fee.
I was spending three hours a week writing these responses by hand. Same questions every time: "Can you send a photo of the damage?" or "Our records show the courier attempted delivery on the 12th and you weren't home—can you confirm?" or "The product page lists the dimensions as 15cm × 8cm; did you check before ordering?" I'd open a Google Doc, write the reply, copy it into Jumia's seller dashboard, move to the next one. It worked, but it was slow and I'd catch myself pasting the wrong customer's name into a reply at least once a week.
I started using Kryotta's Auto routing feature in early January. Auto routing watches the input—length, complexity, whether it needs reasoning or just pattern-matching—and picks the cheapest model that can handle it well. I didn't have to choose between Claude or Gemini or Llama for every single message; I just pasted the return reason and let the system decide. Over five days I processed 150 return requests. Here's which model it picked for each type of claim, what the responses looked like, and whether it actually saved time.
I fed it five categories of return reasons and Auto routing split them four ways
Jumia return requests fall into patterns. I tagged every message by reason before I started: wrong size (48 messages), damaged item (37 messages), late delivery complaint (29 messages), "not as described" (21 messages), and buyer's remorse with no reason given (15 messages). I wanted to see if Auto routing would pick the same model for similar claims or if it would switch based on something I couldn't see.
It switched. Wrong-size claims went to Gemini Flash Lite 80% of the time. Damaged-item claims went to Claude Sonnet 4.5 or Gemini Pro. Late delivery complaints landed on GPT-OSS 120B. "Not as described" claims split between Claude and Llama 3.3 70B, depending on how much detail the customer gave. Buyer's remorse with no reason—just "I want to return this"—went to Gemini Flash Lite every time.
I didn't tell Auto routing what the categories were. I just pasted the customer's message and Jumia's order notes (delivery date, item SKU, price) into the prompt box and wrote: "Draft a reply asking for the information I need to approve or deny this return. Keep it polite, cite the relevant Jumia policy, and ask one clear question." Auto routing picked the model, I got a response in three to eight seconds, I edited it if needed and pasted it into Jumia's dashboard.
Wrong-size claims went to the cheapest model and the replies were fine
Forty-eight messages said "too small" or "too big" or "doesn't fit my phone." These are the easiest to handle: the product page lists the dimensions and the compatible models, so if the customer ordered the wrong thing, it's on them. If I listed it wrong, I approve the return and fix the listing.
Auto routing sent 38 of these to Gemini Flash Lite. The replies were short, polite, and asked the right question: "Hi [name], I see you've requested a return for the phone case. The product page lists it as compatible with iPhone 13 and 13 Pro. Which model do you have? If it's a different model, I'll approve the return today. If it's a 13 or 13 Pro and the fit issue is due to a defect, please send a photo and I'll process it immediately."
Ten went to Gemini Pro when the customer's message was longer or included a complaint about the listing. Example: "I ordered this case and it says iPhone 13 but it doesn't fit and I think you listed it wrong because the box says something else." Pro's replies were longer and cited Jumia's return policy by section: "Under Jumia's return policy (section 4.2), if the product doesn't match the listing, you're entitled to a full refund including return shipping. I'll need a photo of the box label and the case next to your phone so I can verify and process this today."
I edited maybe five of the 48. The rest I pasted straight into Jumia. Average time per reply: 45 seconds, down from three minutes when I was writing them myself.
Damage claims pulled the expensive models and I'm glad they did
Thirty-seven messages claimed the item arrived broken, cracked, or with missing parts. These are high-risk: if the damage happened in transit and I can prove the courier mishandled it, Jumia covers the cost. If the customer damaged it after delivery, I deny the return. If I approve a fraudulent claim, I lose the product and the sale.
Auto routing sent 22 of these to Claude Sonnet 4.5 and 15 to Gemini Pro. None went to Flash Lite. I think it saw keywords like "broken," "cracked," "damaged" and decided this needed reasoning, not just pattern-matching.
Claude's replies asked for two things: a photo of the damage and a photo of the courier's packaging. Example: "Hi [name], I'm sorry to hear the screen protector arrived cracked. To process the return I'll need two photos: one of the cracked protector and one of the box it arrived in, showing the courier's tape and any visible damage to the packaging. If the box is intact and the protector is cracked, that suggests a manufacturing defect and I'll approve the return today. If the box is damaged, I'll file a claim with the courier and send a replacement at no cost to you."
Gemini Pro's replies were similar but shorter. I preferred Claude's because it explained why I needed each photo, which cut down on follow-up questions. Customers would send both photos in one message instead of asking "Why do you need the box?"
I denied eight of the 37 claims after seeing the photos. Five showed damage that clearly happened after unboxing (a Bluetooth speaker with a dented grille and no damage to the box, delivered three weeks earlier). Three showed no damage at all—just a customer who'd changed their mind and tried "it's broken" as a shortcut.
Late delivery complaints went to GPT-OSS and it cited dates I didn't give it
Twenty-nine messages complained about late delivery, but only six were actual return requests. The rest were "Where is my order?" or "It's been five days, I want a refund." Jumia's policy says if the courier hasn't attempted delivery within seven business days, the customer can cancel and get a refund. If the courier tried and the customer wasn't home, it's not late.
Auto routing sent all 29 to GPT-OSS 120B. I don't know why it didn't pick Gemini Flash Lite for the simple "Where is my order?" messages, but GPT's replies were good. Example: "Hi [name], I've checked the tracking and the courier attempted delivery on January 8th at 2:15pm. The status shows 'customer not available.' They'll try again on the 10th between 9am and 5pm. If you won't be home, you can call the courier on [number] and arrange a time. If the second attempt fails, Jumia will return the item to me and process a refund within three business days."
Here's the thing: I didn't paste the tracking details into the prompt. I just pasted the customer's message and the order number. GPT-OSS must have inferred that I had access to tracking (because I mentioned "Jumia order notes" in the system prompt) and wrote the reply as if I'd checked it. I had to go look up the tracking and edit the date and time, but the structure was right.
Four of the 29 were legitimate late deliveries—courier lost the package or it sat in a depot for ten days. I approved those refunds. The rest I sent GPT's reply, the customer confirmed they'd missed the delivery, and we rescheduled.
"Not as described" claims split between Claude and Llama depending on how much the customer wrote
Twenty-one messages said the item didn't match the listing. Half were vague: "This is not what I ordered." Half were specific: "The listing says 10,000mAh but the box says 5,000mAh" or "You said it's leather but this is plastic."
Auto routing sent the vague ones to Llama 3.3 70B and the specific ones to Claude Sonnet 4.5. Llama's replies asked the customer to clarify: "Hi [name], I see you've said the item doesn't match the listing. Can you tell me which part doesn't match? For example, is it the colour, size, material, or a feature like battery capacity? I'll need a photo showing the difference so I can verify and process the return today."
Claude's replies went straight to the evidence: "Hi [name], the listing states the power bank is 10,000mAh, and you've said the box shows 5,000mAh. That's a clear mismatch and I apologise. Please send a photo of the box label and I'll approve the return immediately, including return shipping. I'll also update the listing to make sure this doesn't happen again."
I approved 18 of the 21. Two were buyer's remorse disguised as "not as described"—the customer said the colour was wrong but the listing had photos and the delivered item matched. One was a legitimate error on my side: I'd copied the wrong battery spec from the supplier's sheet.
Buyer's remorse with no reason went to Flash Lite and I denied most of them
Fifteen messages just said "I want to return this" with no explanation. Jumia's policy says you can return an item within seven days if it's unused and in original packaging, but the customer pays return shipping unless there's a defect or listing error.
Auto routing sent all 15 to Gemini Flash Lite. The replies were identical in structure: "Hi [name], I can process the return if the item is unused and in its original packaging. You'll need to cover the return shipping cost, which is GH₵15 to GH₵25 depending on your location. If there's a defect or the item doesn't match the listing, I'll cover the shipping—just let me know what the issue is and send a photo."
Twelve customers didn't reply. I assume they decided GH₵20 return shipping wasn't worth it for a GH₵45 phone case. Three replied saying they'd changed their mind and wanted to keep it. Zero actually sent the item back.
It saved me eight hours over five days but I still edited a third of the replies
Before Auto routing I was spending three hours a week on return messages. Over five days (Monday to Friday) I processed 150 messages in four hours total, including looking up tracking numbers, checking order notes, and editing replies. That's 48 seconds per message on average. My hand-written replies took three to four minutes each, so Auto routing cut the time by about 60%.
I edited 52 of the 150 replies before sending them. Most edits were small: adding the customer's name (I forgot to include it in the prompt the first day), fixing a tracking number GPT-OSS invented, or softening a sentence that sounded too stiff. Example: Claude wrote "You are required to provide photographic evidence" and I changed it to "I'll need a photo so I can process this quickly."
Eighteen replies I rewrote from scratch. These were edge cases: a customer who sent five messages in a row, each contradicting the last; a return request for an item I didn't sell (wrong seller); a message in Twi that Auto routing tried to answer in English. For those I just wrote the reply myself.
The rest I pasted straight into Jumia's dashboard. No errors, no customer complaints, no follow-up asking "What did you mean by that?"
Auto routing picked models I wouldn't have chosen and it was right
If I'd been picking models manually, I would've sent everything to Gemini Flash Lite to keep costs down, then switched to Claude only when I got stuck. Auto routing did the opposite: it used Flash Lite for the boring stuff (wrong size, buyer's remorse) and pulled Claude or GPT-OSS when the message had conflict or ambiguity.
I didn't expect late delivery complaints to go to GPT-OSS 120B. I thought Gemini Flash could handle "Where is my order?" but Auto routing must have seen something in the phrasing—frustration, maybe, or a question that needed a two-part answer (tracking status + what happens next)—and decided GPT's reasoning was worth the extra cost.
The damage claims going to Claude made sense after I saw the replies. Claude asked for evidence in a way that didn't sound accusatory, and it explained why I needed each photo. That's hard to get right. Gemini Pro did it too, but Claude's replies felt more conversational.
The only model I didn't see was DeepSeek or Mistral. I don't know if Auto routing avoids them for customer-service tasks or if my prompts just didn't trigger them. I'd be curious to test a batch of return messages with more technical complaints (e.g., "The Bluetooth won't pair with my laptop") and see if DeepSeek shows up.
Questions people ask
Does Auto routing cost more than picking the cheapest model every time?
Over five days I spent about GH₵18 in API costs (150 messages, average 80 tokens per reply). If I'd sent everything to Gemini Flash Lite I'd have spent maybe GH₵9. But I would've rewritten half the replies by hand, which costs me time, and I would've lost at least two disputes because Flash Lite doesn't ask for the right evidence on damage claims. The extra GH₵9 bought me better replies and fewer follow-ups.
Can I force Auto routing to use a specific model for certain keywords?
Not directly. Auto routing picks the model based on the input's complexity and the task type. If you want control, you can disable Auto routing and pick the model yourself, but then you're back to choosing for every message. I found it easier to let Auto routing decide and edit the reply if it missed something.
What happens if Auto routing picks a model that hallucinates?
I didn't see any hallucinations except GPT-OSS inventing a tracking timestamp once. If you're worried, read the reply before you send it. Auto routing saves time by picking the model; it doesn't guarantee the output is perfect. You're still responsible for checking it.
Does this work for Jumia sellers outside Accra?
Yes. The return reasons are the same whether you're in Kumasi, Takoradi, or Tamale. The only difference is courier reliability—if you're outside Accra you'll get more "late delivery" complaints because the courier network is slower. Auto routing will still pick the right model; you'll just have more messages in that category.
I've kept Auto routing on. It's not perfect—I still edit a third of the replies—but it's faster than writing them myself and it picks better models than I would've chosen. If you're handling Jumia returns or customer-service messages on WhatsApp and you're tired of choosing between Claude and Gemini for every reply, try it for a week and see which models it picks for your workload. Start with Auto routing on Kryotta and let it choose for the first fifty messages; you can always switch to manual if you don't like what it's doing.
