The Tuesday I lost ₦48,000 because I trusted the wrong myth
I sell skincare and hair products on Instagram. About 400 orders a month, most paid through Paystack—bank transfer or card. I ship with GIG Logistics to Lagos, Abuja, Port Harcourt. Normal weeks, I get one or two disputes. Someone claims the package never arrived, or their card was used without permission, or the product "looks different from the photo." I respond, upload the delivery confirmation, usually win.
Then a weekend sale brought in 180 orders. By Wednesday, I had nine Paystack disputes open at once. Three were "fraudulent transaction" claims. Four were "product not received." Two were "product unacceptable"—a customer in Ibadan ordered the shea butter cream, received it, used half the jar, then filed a dispute saying it "didn't work as advertised." Paystack gives you seven days to respond with evidence. I had forty-eight hours left on the oldest one.
I'd been using Claude Sonnet 4.5 to write product captions and reply to DMs, so I assumed it was the right model for dispute responses too. Everyone on Twitter said Claude was smarter. I pasted the first dispute into Claude, asked it to draft a response, and it gave me three paragraphs of polite explanation with zero reference to Paystack's evidence requirements. I lost that case. ₦12,500 gone, plus the product cost.
Then I tried Gemini Flash. I'd heard it hallucinates numbers, so I expected disaster. It wrote a tighter response, cited the delivery photo, referenced the customer's WhatsApp confirmation, and won the case. I started testing both models on every dispute type, keeping a spreadsheet of which one actually saved the money. Sixty disputes later, I know which myths cost me naira and which model wins per situation.
Myth one: Claude is always better because it's 'smarter'
The belief: Claude Sonnet 4.5 is the premium model, so it must write better dispute responses. Gemini is the budget option that cuts corners.
The reality: Claude writes longer responses. That's not the same as better. Paystack dispute responses need three things—evidence of delivery or service, reference to the customer's own messages, and a direct answer to the claim type. Claude's default style is explanatory and apologetic. It wants to show empathy and context. That's useful for customer service emails. It's a liability in a dispute.
I tested both models on fifteen "product not received" disputes. I gave each model the same inputs: tracking number, delivery confirmation screenshot, customer's WhatsApp chat where they asked about shipping, and the dispute reason. Claude Sonnet 4.5 wrote an average of 280 words per response. Gemini Flash wrote 140. Claude opened with "We understand your concern" and spent two sentences explaining shipping delays caused by Lagos traffic before getting to the evidence. Gemini opened with "The package was delivered on [date] at [time]. Please see attached proof." Then it cited the tracking number and the customer's own message confirming the address.
I won twelve of the fifteen disputes using Gemini's drafts. I won nine using Claude's. The three I lost with Gemini were cases where the delivery photo was blurry and the customer had never confirmed receipt on WhatsApp—no model was going to save those. The six I lost with Claude were winnable cases where the response buried the evidence under two paragraphs of politeness, and Paystack's reviewer didn't read far enough.
Claude isn't worse. It's solving a different problem. If you're writing to the customer to de-escalate before they file a dispute, use Claude. If you're writing to Paystack's dispute team to prove your case, use Gemini.
Myth two: Gemini hallucinates refund amounts and tracking numbers
The belief: Gemini invents numbers. You paste in a ₦15,000 transaction, and it drafts a response mentioning ₦18,500. Or it cites a tracking number you never gave it.
The reality: Gemini hallucinates when you give it ambiguous instructions or incomplete context. Claude hallucinates less often because it asks clarifying questions or hedges with phrases like "based on the information provided." That caution is useful in some workflows. In a dispute response, it reads like uncertainty, and uncertainty loses cases.
I tested this directly. I took ten disputes and gave each model incomplete information on purpose—transaction amount but no tracking number, or delivery photo but no date. Then I asked both to draft a response. Claude Sonnet 4.5 wrote things like "According to our records, the package was shipped" without specifying when. Gemini Flash wrote "The package was delivered on [DATE]"—and yes, it invented a date twice. But in eight of the ten cases, Gemini said "Please provide the delivery date" or left a bracket placeholder. Claude just wrote around the gap.
The fix is simple: give Gemini the full context in one message. Transaction ID, amount in naira, tracking number, delivery confirmation, customer's messages, dispute reason. If you're pasting from Paystack's dispute page, include the "Evidence Required" section so Gemini knows what to cite. I haven't seen a single hallucinated number since I started doing that. And when I do see a bracket or a "please confirm" note in Gemini's draft, I know I forgot to paste something.
Claude's caution costs you in a different way. It writes "the customer may not have been available to receive the package" when you've already uploaded a photo of the package in the customer's hands. Hedging loses disputes. Gemini states facts. Just make sure the facts are in the prompt.
Myth three: you need Claude Opus to handle 'product unacceptable' disputes
The belief: "Product unacceptable" disputes are complicated. The customer is claiming the product is defective or not as described. You need the most expensive model to write a response that addresses quality claims, cites your return policy, and doesn't sound defensive.
The reality: I tested Claude Opus 4.5, Claude Sonnet 4.5 and Gemini Pro on twelve "product unacceptable" disputes. Opus wrote the most thorough responses—four paragraphs, references to ingredient lists, explanations of how the product works. I lost ten of the twelve. Sonnet wrote three paragraphs and won four. Gemini Pro wrote two tight paragraphs and won seven.
Paystack's dispute reviewers don't care about your ingredient list. They care whether the customer's claim matches the evidence. If the customer says "the product looks different from the photo," the winning response shows that your product photo is accurate and the customer received exactly what was pictured. If the customer says "the product didn't work," the winning response shows that you disclosed what the product does, the customer confirmed understanding before purchase, and you offered a return within your stated policy.
Opus tried to argue the product's merits. That's not the reviewer's job. Gemini Pro cited the customer's own WhatsApp message saying "I know it takes two weeks to see results" and the return policy link I'd sent before shipping. Case closed. And Gemini Pro costs a fraction of Opus per response—₦8 versus ₦45 in my usage.
The one place Opus helped: a dispute where the customer's claim was vague ("product not as expected") and I had six different WhatsApp threads with them. Opus summarized all six threads and pulled the relevant quotes. Gemini Pro got confused by the volume. So if you have a messy case with ten screenshots and three email threads, Opus is worth it. For a standard two-paragraph dispute where the customer's claim is clear, Gemini Pro wins.
Myth four: reasoning models like DeepSeek R1 are better at fraud detection
The belief: Disputes marked "fraudulent transaction" need a reasoning model. DeepSeek R1 or o1-preview will analyze the transaction pattern, compare it to the customer's behaviour, and write a response that proves it wasn't fraud.
The reality: Reasoning models are slow and they overthink. I tested DeepSeek R1 on eight "fraudulent transaction" disputes. It took forty seconds to generate the first response. It wrote five paragraphs analyzing the customer's order history, the IP address (which I hadn't provided), the time of day, and the shipping address. None of that matters to Paystack's reviewer. What matters is whether the cardholder authorized the transaction and whether you have proof.
Gemini Flash wrote the same response in four seconds: "The cardholder completed 3D Secure authentication. Please see attached transaction receipt and delivery confirmation to the cardholder's registered address." I won seven of the eight disputes with Gemini's version. I won five with DeepSeek's version, and in two of those cases, the reviewer asked for clarification because DeepSeek's response mentioned "IP geolocation analysis" that I couldn't provide evidence for.
Reasoning models are brilliant for debugging code or planning a content calendar. They're overkill for a dispute response, and the extra reasoning steps introduce claims you can't back up. Stick with Gemini Flash for "fraudulent transaction" disputes. If the case is genuinely complicated—the cardholder's bank is claiming the card was stolen, and you have chat logs proving the person who ordered knew details only the real cardholder would know—then use Claude Sonnet 4.5. It writes a clearer narrative. But most fraud disputes are simple: 3D Secure passed, package delivered to the registered address, done.
Myth five: you should use the same model for every dispute type
The belief: Pick the best model and use it for everything. Switching models wastes time.
The reality: I keep three models open in Kryotta when I'm handling disputes. Gemini Flash for "product not received" and "fraudulent transaction." Claude Sonnet 4.5 for "product unacceptable" when the customer's claim is emotional or vague. Gemini Pro when I need to summarize a long WhatsApp thread or pull quotes from multiple screenshots.
It takes three seconds to paste the dispute into a different chat. That three seconds saves me ₦12,000 when I pick the right model for the job. Gemini Flash is fast, factual, and cheap—₦2 per response on average. Claude Sonnet is slower, more narrative, better at handling ambiguity—₦15 per response. Gemini Pro sits in the middle—₦8 per response, good at structure when you have a lot of evidence to organize.
The workflow: I open the Paystack dispute. I read the claim type. If it's "product not received" or "fraudulent," I paste into Gemini Flash. If it's "product unacceptable" and the customer wrote two sentences, I use Gemini Flash. If the customer wrote six paragraphs of complaints and I have a long reply thread, I use Claude Sonnet 4.5. If I have ten screenshots and need them summarized with quotes, I use Gemini Pro. I review the draft, add the tracking number or delivery photo if the model missed it, and submit.
I've written sixty dispute responses this way in three weeks. I've won forty-seven. My win rate before I started testing models was about 60 percent. Now it's 78 percent, and I'm spending ₦180 a week on AI instead of ₦15,000 a week on lost disputes.
Questions people ask
Can I use these models if I'm on Flutterwave instead of Paystack?
Yes. The dispute process is nearly identical—evidence of delivery, reference to customer communication, direct answer to the claim. The same model recommendations apply. Gemini Flash for straightforward cases, Claude Sonnet when the customer's claim is vague or emotional.
What if I don't have delivery confirmation because I use a local dispatch rider?
Ask your rider to send you a photo when they drop off the package, or get the customer to confirm receipt on WhatsApp before you mark the order complete. If you don't have proof of delivery, no model will win the dispute for you. The AI can only work with the evidence you give it.
Do I need to pay for Claude Opus or can I use the free tier?
You don't need Opus for most disputes. Gemini Flash and Claude Sonnet 4.5 handle 95 percent of cases, and both are available on Kryotta's free tier with reasonable daily limits. Opus is only worth paying for when you have a genuinely complicated case with many threads of evidence.
How do I stop Gemini from inventing numbers?
Paste the full context in one message: transaction ID, amount, tracking number, delivery date, customer messages, and the dispute reason from Paystack. If you give Gemini everything it needs, it won't guess. If you see a placeholder like [DATE], you forgot to include that detail.
I still lose disputes sometimes. But I lose fewer of them, and I spend less time writing responses that don't work. If you're handling Paystack disputes and you've been using one model for everything, try splitting by claim type. Open Kryotta, paste your next "product not received" dispute into Gemini Flash, and your next "product unacceptable" into Claude Sonnet. Compare the drafts. One of them will save you money this week.
