I used 3 AI models to write 95 Klarna late-payment reminder emails: which tone got paid

The spreadsheet that showed me I'd been too polite

I run a D2C ceramics brand in Berlin. We sell handmade mugs, bowls and planters through our Shopify store, and about 40 per cent of German customers choose Klarna's "pay later" option at checkout. Most settle within the standard fourteen days. Some don't.

By mid-March I had ninety-five open invoices totalling €8,400. The amounts ranged from €28 for a single mug to €240 for a full dinner set. Klarna sends one automatic reminder at day fifteen, but after that the follow-up is on me. I'd been sending manual emails—gentle, apologetic, always ending with "no worries if you need more time"—and my payment rate was dismal. About 30 per cent paid within a week of my message. The rest ignored me or replied with vague promises.

I decided to test whether AI could write better reminders than my conflict-averse self, and whether the tone mattered more than the model. I picked three models on Kryotta—Claude Sonnet 4.5, Gemini Pro, and DeepSeek V3—and three tones: firm (clear deadline, consequence stated), friendly (empathetic, no threat), and final-notice (urgent, one last chance). That gave me nine combinations. I split the ninety-five invoices into nine groups of ten or eleven, matched by invoice age and amount so no group had an unfair advantage, and sent one combination to each group. Then I tracked who paid within seven days.

The results surprised me. The model mattered less than I expected. The tone mattered more, but not in the way I thought.

The three prompts I used

I kept the structure identical across all nine. Each prompt included the customer's first name, the invoice number, the amount in euros, the original due date, and how many days overdue it was. Then I added the tone instruction.

Firm tone: "Write a payment reminder email for a Klarna invoice. Be direct and professional. State the amount, the due date, and that payment is now overdue. Include a clear deadline—payment must be received within 5 days—and mention that continued non-payment will result in the invoice being escalated to Klarna's collections process. Keep it under 120 words. No apologies."

Friendly tone: "Write a payment reminder email for a Klarna invoice. Be warm and understanding. Acknowledge that people sometimes miss payments. Gently remind them of the amount and due date, express that you'd love to resolve this together, and offer to help if they're having trouble. Keep it under 120 words. End on a positive note."

Final-notice tone: "Write a final-notice email for a Klarna invoice that's significantly overdue. Be urgent but not aggressive. Make it clear this is the last reminder before the matter is handed to collections. State the amount, the original due date, and that payment must be made within 3 days. Keep it under 100 words. No pleasantries."

I ran each prompt through each model, copied the output into Gmail, personalised the subject line ("Your Klarna invoice #[number] – action needed"), and sent them over two days in mid-March.

What actually got people to pay

Seven days later I pulled the numbers. Here's the payment rate for each combination:

Final-notice won across all three models. The average payment rate for final-notice emails was 70 per cent, compared to 61 per cent for firm and 50 per cent for friendly. The friendly tone, which I'd been using in my manual emails, performed worst everywhere.

Claude edged ahead overall, but the gap between models was smaller than the gap between tones. The difference between Claude's final-notice (73 per cent) and DeepSeek's final-notice (67 per cent) was six percentage points. The difference between Claude's friendly (55 per cent) and Claude's final-notice (73 per cent) was eighteen points.

One thing I didn't expect: firm underperformed final-notice even though firm was less aggressive. I thought customers would resent the escalation threat in final-notice and that firm would feel more reasonable. Wrong. The urgency and brevity of final-notice seemed to cut through. People either paid immediately or didn't reply at all. Firm got more "I'll pay next week" responses that didn't convert.

The Claude output I used most often

Here's the final-notice email Claude wrote for a €112 invoice that was thirty-one days overdue:

Subject: Final notice – Klarna invoice #8834 requires immediate payment

Hello Marta,

This is your final reminder regarding Klarna invoice #8834 for €112, which was due on 12 February and is now 31 days overdue.

Payment must be received within 3 days (by 18 March). If we do not receive payment by this date, the invoice will be transferred to Klarna's collections department, which may affect your credit rating and incur additional fees.

You can pay directly through the link in your original Klarna email, or reply to this message if you need assistance.

Thank you,
[My name]

No fluff. No "we understand times are tough." It worked. Marta paid the next day.

Compare that to the friendly version Gemini Pro generated for a similar case:

Subject: A quick reminder about your Klarna invoice

Hi Thomas,

I hope this message finds you well! I wanted to reach out about Klarna invoice #8891 for €98, which was due on 15 February. I know life gets busy and payments can slip through the cracks—it happens to all of us!

If you've already taken care of this, please disregard this email. If not, I'd really appreciate it if you could settle the balance when you have a moment. If there's anything I can do to help or if you'd like to discuss a payment plan, just let me know. I'm here to make this as easy as possible.

Thanks so much for your support!
[My name]

Thomas didn't pay. He replied four days later saying he'd "sort it this week," then went silent. I ended up escalating that one to Klarna after another two weeks.

Why friendly fails and what that means for GDPR-conscious stores

The friendly emails got more replies, but fewer payments. Customers thanked me for being understanding, asked for extensions, or said they'd pay "soon." It felt like I was inviting negotiation. Final-notice shut that down. It framed the situation as binary: pay now, or this goes to collections.

One concern I had was tone and GDPR. In the EU you can't harass customers or use threatening language in debt collection. The final-notice emails stayed legal because they stated a factual consequence (Klarna's collections process) without exaggeration or insult. I checked with a lawyer friend who works in e-commerce compliance, and she confirmed that as long as you're truthful about what happens next and you're not sending ten emails a day, you're fine.

If you operate a multilingual store—say, you sell into Germany, France and Poland—you'll need to translate the AI output. I used DeepL for the handful of French-speaking customers in my list. Claude and Gemini can both draft in other languages if you add "write this in French" to the prompt, but I found DeepL's output more natural for formal communication. The tone carried across. Urgent in English stayed urgent in French.

The two cases where I switched models mid-process

I didn't stick to the test groups religiously. Two invoices were edge cases, and I made judgment calls.

Case one: A customer in Amsterdam owed €240 for a large order and had already ignored two reminders. I used Claude's final-notice prompt but added a line offering a payment plan—"If paying the full amount at once is difficult, reply and we can arrange two instalments." She replied within an hour, we agreed on two €120 payments a week apart via SEPA transfer, and she stuck to it. Claude handled the empathy injection better than Gemini when I tested the same addition; Gemini's version felt like two emails stitched together.

Case two: A customer in Munich owed €34 for a mug and had been overdue for forty-five days. I'd sent a friendly reminder and a firm reminder with no response. I switched to DeepSeek's final-notice output because it was the shortest of the three—just sixty-eight words—and I wanted something that could be read in ten seconds on a phone. He paid the same day. Sometimes brevity is the entire strategy.

The follow-up system I built after the test

I now send three emails for every overdue Klarna invoice, spaced like this:

1. Day 18 (three days after Klarna's automatic reminder): Friendly tone, Gemini Pro. This catches the genuinely forgetful and keeps goodwill intact.
2. Day 28 (if no payment): Firm tone, Claude Sonnet. This is the middle ground—serious but not final.
3. Day 38 (if still no payment): Final-notice tone, Claude Sonnet. After this I escalate to Klarna.

I use Kryotta to generate all three at once when an invoice hits day eighteen. I paste the customer details into a saved prompt template, run it through each model and tone, then schedule the emails in Gmail. The whole process takes about four minutes per invoice. My seven-day payment rate across all overdue invoices is now 68 per cent, up from 30 per cent when I was writing them myself.

One adjustment: I don't use final-notice for amounts under €30. The cost of escalation to Klarna (they charge a fee and take a percentage) often exceeds what I'd recover, so for small amounts I send two reminders then write it off. For everything above €30, the three-step sequence pays for itself.

Questions people ask

Can I automate this with Zapier or Make so I don't have to paste prompts manually?
Yes, if your payment processor or Shopify sends a webhook when an invoice goes overdue. You'd connect that to Kryotta's API (if you're on a team plan), pass the customer details and tone instruction, then route the output to Gmail or your CRM. I haven't done it because I only get five to eight overdue invoices a week and I like reviewing the emails before they go out, but for higher volume it's worth building.

What if the customer replies in another language?
I use Claude or Gemini to draft a reply in their language. Prompt: "Reply to this message in French. Confirm I've received their payment / noted their request for an extension / whatever the situation is. Keep it polite and under 50 words." Then I run the output through DeepL to double-check before sending.

Does the subject line matter as much as the body?
Probably more. I tested two subject styles in a smaller follow-up experiment: descriptive ("Your Klarna invoice #8834 – action needed") versus urgent ("Final notice: payment required by 18 March"). Urgent subjects had a 12 percentage-point higher open rate, but only for final-notice emails. For friendly reminders, descriptive performed better. People seem to resent false urgency when the tone doesn't match.

What about customers who just can't pay right now?
I offer a payment plan if they ask, usually two instalments a week apart via SEPA. About one in ten takes me up on it, and 80 per cent of those complete both payments. The AI models aren't great at drafting payment-plan terms—they either sound too formal or too vague—so I wrote a standard reply template for that and paste it in manually.

If you're sitting on a pile of overdue Klarna invoices

Start with final-notice for anything over thirty days and above €50. Use Claude if you have access; the output needs the least editing. For newer overdues (under three weeks), try friendly first—you'll keep more customers happy and some will pay immediately out of embarrassment.

Track your results for a month. If your payment rate doesn't improve, the problem isn't the email; it's your checkout flow, your product, or your customer base. But if you're like me and you've just been too polite, a sharper tone written by something that doesn't worry about being liked will get you paid.

You can try all three models side by side at Kryotta and see which one writes reminders that sound like you'd actually send them.