The month Klarna disputes swallowed my evenings
I run a small online shop selling outdoor gear—hiking boots, waterproof jackets, camping cookware, the sort of kit people order in March when they've decided this is the year they'll finally hike the Wicklow Way or the Camino. We're based in Dublin, ship across the EU, and about sixty per cent of our orders go through Klarna. Customers love the split-payment option; I tolerate it because it lifts conversion by a third.
In April I had 103 Klarna disputes. Not chargebacks—Klarna handles those separately—but the buyer-initiated claims that land in your merchant portal with a seven-day clock and a polite threat that if you don't respond adequately, Klarna will side with the customer and deduct the amount from your next settlement. The claims fell into three buckets: "item never arrived" (41 cases), "item damaged or not as described" (38 cases), and "I returned it within the window but you haven't refunded me" (24 cases).
Each dispute wanted the same structure—acknowledge the claim, present your evidence, explain the resolution—but the tone and the detail varied wildly depending on the type. A non-delivery claim needed tracking proof and a calm explanation of courier delays. A not-as-described claim needed photos, product specs, and a bit of pushback if the customer had clearly misread the listing. A return-window dispute needed your returns policy quoted verbatim and proof of when (or whether) the item came back.
I was spending two hours every evening drafting responses in the Klarna portal, and my accept rate—the percentage of disputes Klarna closed in my favour—was 68 per cent. That's below the 75 per cent threshold where Klarna starts asking pointed questions about your processes. I decided to test whether Claude or Gemini could write better responses, faster, and which model worked best for each dispute type.
The test: same evidence, two models, one month of real claims
I used Kryotta's Compare arena to run Claude Sonnet 4.5 and Gemini Pro side by side. For each dispute I fed both models the same information: the customer's claim (copied from Klarna's portal), our order record, the tracking number and delivery status, and any photos or return documentation we had. I gave each model a base prompt that laid out the structure Klarna expects—acknowledge, evidence, resolution—then tweaked the prompt slightly depending on the dispute type.
I drafted responses with both models, read them blind (Kryotta hides which model wrote what until you pick), chose the one that felt clearer and more defensible, submitted it to Klarna, and tracked the outcome. Over thirty days I processed all 103 disputes this way. I logged which model I picked, which dispute type, and whether Klarna accepted my response or sided with the customer.
The headline: Claude won 61 of the 103 head-to-heads, Gemini won 42. But the pattern mattered more than the overall score. Claude dominated non-delivery claims (31 wins vs 10 for Gemini). Gemini edged ahead on not-as-described disputes (22 vs 16). Return-window claims split almost evenly (14 Claude, 10 Gemini), and in that category the model choice mattered less than having your returns policy nailed down in the first place.
My Klarna accept rate climbed to 81 per cent by the end of April, and I was clearing the evening dispute queue in forty minutes instead of two hours.
Non-delivery claims: Claude's structured calm beats Gemini's verbosity
A non-delivery dispute is the simplest on paper—customer says the parcel didn't arrive, you show tracking that says it did—but it's also where emotions run highest. The customer genuinely believes they've been robbed, and Klarna's reviewers want to see that you've treated the claim seriously.
I used this prompt for both models:
You are responding to a Klarna dispute where the customer claims the item never arrived. Acknowledge their frustration, present the tracking evidence (include carrier name, tracking number, delivery date and signature or photo proof if available), explain what "delivered" means in this case, and offer a resolution (replacement, refund, or ask them to check with neighbors/building reception). Keep the tone calm and factual. 150–200 words.
Claude's responses were consistently tighter and more structured. Here's a sample (order #8472, hiking boots, delivered 12 March to a locker in Utrecht):
We understand how frustrating it is when a parcel doesn't reach you. Our records show that order #8472 was delivered by PostNL on 12 March at 14:22 to the Bruna parcel locker at Voorstraat 45, Utrecht (tracking number 3SABCD1234567890). PostNL's system confirms the parcel was collected from the locker on 13 March at 09:15 using the collection code sent to the mobile number ending in **68.
If you didn't collect it yourself, we recommend checking whether someone else in your household picked it up, or contacting PostNL directly at 088-2255555 to request locker access records. We're happy to open an investigation with the carrier if needed, but based on the delivery confirmation we're unable to issue a refund at this stage. Let us know how you'd like to proceed.
Gemini's version of the same dispute ran to 240 words and included two paragraphs explaining how parcel lockers work and a sentence about GDPR that had nothing to do with the claim. Klarna's reviewers don't want a lecture; they want evidence and a clear next step.
Claude won 31 of 41 non-delivery disputes in the blind test. When I submitted Claude's responses to Klarna, 28 were accepted. Gemini won 10, and 7 of those were accepted. The difference wasn't the facts—both models had the same tracking data—it was that Claude laid out the evidence in the order Klarna's process expects, then stopped.
Not-as-described claims: Gemini's detail and nuance pulled ahead
A not-as-described dispute is trickier. The customer received the item but says it doesn't match the listing—wrong colour, wrong size, missing a feature, or just "not what I expected." Sometimes they're right and you've mislabelled something. Sometimes they've misread the specs or changed their mind and are trying to force a return outside your policy.
I used this prompt:
You are responding to a Klarna dispute where the customer claims the item is damaged or not as described. Acknowledge their disappointment, quote the exact product description from the listing (include size, material, colour as listed), compare it to what they received (based on their claim or photos), and explain whether this is a quality issue (offer replacement/refund) or a misunderstanding (politely hold your ground). Attach proof if you have it—photos from the listing, manufacturer specs. 180–220 words.
Gemini handled these better. It was more careful about tone when pushing back, and it wove in the product details more naturally. Here's a sample (order #8591, waterproof jacket, customer claimed it wasn't "fully waterproof" after wearing it in heavy rain):
We're sorry the jacket didn't perform as you'd hoped. The product listing describes this as a "water-resistant softshell jacket (5,000 mm waterproof rating, DWR coating)," which is suitable for light to moderate rain but not designed for prolonged heavy downpours. Fully waterproof hardshell jackets—rated 10,000 mm or higher—are listed separately in our "Expedition" category and are priced accordingly.
The jacket you received matches the listing specification: 5,000 mm polyester outer, taped seams, DWR finish. We've attached a photo from the original listing and the manufacturer's spec sheet. If the DWR coating has failed (water soaking through rather than beading), that would be a warranty issue and we'll replace it. If the jacket performed as rated but wasn't suitable for the conditions you encountered, we can't offer a refund under our returns policy (14 days, unworn condition), but we're happy to discuss a partial credit toward a hardshell upgrade.
Claude's version of the same dispute was more defensive and less willing to meet the customer halfway. It quoted the specs but didn't explain why a 5,000 mm rating isn't the same as "fully waterproof," and it didn't offer the upgrade path.
Gemini won 22 of 38 not-as-described disputes. I submitted 20 of those to Klarna, and 17 were accepted. Claude won 16, and 13 were accepted. The gap came down to Gemini's ability to acknowledge the customer's expectation, explain the technical reality, and offer a face-saving resolution that Klarna's reviewers could see as fair.
Return-window disputes: either model works if your policy is airtight
Return-window disputes are binary: either the customer returned the item within your stated window and you owe them a refund, or they didn't and you don't. The model's job is to state your policy, show the timeline, and explain the outcome.
I used this prompt:
You are responding to a Klarna dispute where the customer claims they returned the item within the return window but haven't received a refund. State your return policy (14 days from delivery, item must be unworn/undamaged, customer pays return postage unless item is faulty). Show the delivery date, the return window deadline, and the date you received the return (or confirm you haven't received it). If they're outside the window or the item came back used, explain clearly but politely. If they're right, apologize and confirm the refund timeline. 150–180 words.
Both models handled these competently. Claude was slightly more concise; Gemini was slightly warmer. The outcome depended more on whether I had solid records—return tracking, photos of the returned item's condition—than on which model wrote the response.
Claude won 14 of 24 return-window disputes, Gemini won 10. Klarna accepted 11 of Claude's and 8 of Gemini's. The three cases where both models failed were ones where our records were incomplete—no return tracking, no photo of the returned item—and no amount of good writing was going to save that.
The prompt tweaks that mattered
Three changes lifted the quality across both models:
Include the exact word count. Klarna's portal has a 2,000-character limit, but responses shorter than 150 words look like you haven't engaged with the claim. I found 150–220 words hit the sweet spot—long enough to show you've investigated, short enough that Klarna's reviewer will read the whole thing. Both models respected the word count when I put it in the prompt; without it, Gemini often ran to 300 words.
Name the carrier and tracking number in the first paragraph. Klarna's reviewers skim. If they have to hunt for the tracking number, they're more likely to side with the customer by default. Putting "PostNL tracking 3SABCD1234567890, delivered 12 March" in the opening sentence made a visible difference to accept rates.
Offer a next step even when you're holding your ground. "We're unable to issue a refund" sounds final and unhelpful. "We're unable to issue a refund, but we're happy to open a carrier investigation or offer a replacement if the item is faulty" gives Klarna's reviewer a reason to close the dispute in your favour without feeling like they've left the customer stranded.
What didn't work: trying to make one prompt fit all three types
I tried a universal prompt in the first week—"You are responding to a Klarna dispute. Acknowledge the customer's concern, present the evidence, and explain the resolution"—and it produced responses that were too generic to be useful. Non-delivery claims need tracking details up front; not-as-described claims need product specs and a willingness to negotiate; return-window claims need your policy quoted verbatim. The models couldn't guess which type of dispute they were handling without being told.
I also tried asking both models to "write in a friendly, empathetic tone" and got worse results. Claude became overly apologetic ("We're so sorry for any confusion"); Gemini added filler ("We truly value your business and want to make this right"). Klarna's reviewers don't care about empathy theatre—they care about facts and fairness. "Calm and factual" in the prompt worked better than "friendly."
My current workflow: Claude for non-delivery, Gemini for not-as-described, either for returns
I now route disputes by type. Non-delivery claims go to Claude Sonnet 4.5 with the structured prompt above. Not-as-described claims go to Gemini Pro with the detail-heavy prompt. Return-window claims go to whichever model I'm already using that evening—the difference is too small to bother switching.
I draft the response in Kryotta, copy it into Klarna's portal, attach the evidence (tracking screenshot, product photos, returns log), and submit. The whole process takes four to six minutes per dispute instead of fifteen. My Klarna accept rate has held at 80–82 per cent since May, and I've cleared the dispute backlog that was sitting at thirty-plus cases in March.
I still read every response before submitting it—both models occasionally hallucinate a tracking number or misread a date—but I haven't had to rewrite one from scratch in two months.
Questions people ask
Can I use these prompts with other BNPL providers like Clearpay or PayPal Pay in 3?
Yes. The structure—acknowledge, evidence, resolution—works for any payment dispute process. You'll need to adjust the word count and tone slightly (PayPal's process is more formal, Clearpay's is closer to Klarna's), but the core prompts translate directly.
What if I don't have tracking proof for a non-delivery claim?
You're in trouble regardless of which model you use. Klarna will almost always side with the customer if you can't show delivery confirmation. The best you can do is explain your dispatch process, offer a replacement, and tighten your tracking procedures going forward.
Do I need to translate responses for customers in other EU countries?
Klarna's dispute portal is in English, so your response should be too—Klarna's reviewers work in English regardless of the customer's language. If you're replying directly to the customer outside the portal, yes, translate it (DeepL is faster and cheaper than asking Claude or Gemini to translate).
How do I know which model to use if the dispute type isn't obvious?
Read the customer's claim and your order notes, then decide: if it's about delivery, use Claude; if it's about product quality or expectations, use Gemini; if it's about return timelines, flip a coin. You'll know within two disputes whether you've picked wrong—the response will feel either too blunt or too wordy.
I've been using this two-model workflow since May, and it's cut my dispute-handling time by about seventy per cent while lifting my accept rate above the threshold where Klarna stops sending warning emails. If you're drowning in BNPL disputes and drafting every response from scratch, try splitting them by type and letting Kryotta's Compare arena show you which model writes the version that actually closes the case.
