The inbox that nearly broke me: 190 return emails in three weeks
I help run an online fashion and homeware business. We sell through our own Shopify store, Zalando and Bol.com, and we ship to twelve EU countries. In late February and early March—right after winter sale season—we took 190 return requests by email. Some were straightforward size swaps in Dutch. Others were multi-part questions in German about GDPR data deletion mixed with a return. A handful looked like fraud attempts in English, asking for refunds to different bank accounts than the original payment.
Before this avalanche, I handled returns myself: read the email, check the order in Shopify or the marketplace backend, decide whether to approve or escalate, write a reply in the customer's language. Average time per email was nine minutes. That's 28.5 hours across 190 emails, spread over three weeks of evenings and weekends, while I also tried to pack orders and update product listings.
By day four I was pasting emails into DeepTranslate, copying the translation into Claude, asking for a reply draft, pasting that into Google Translate to get it back into Dutch or German, then proofreading it. It worked, but it felt like I was paying for three tools to do one job badly.
I'd heard that Kryotta's Auto routing feature picks the best model for each query—fast and cheap for simple stuff, reasoning-heavy for complex cases—so I connected our support inbox, turned on Auto routing, and logged which model it chose for every one of the 190 emails. Then I scored each reply for accuracy, tone and whether it prevented a follow-up email. Here's what the system picked, and whether the routing overhead was worth it.
Test setup: what I tracked and how Auto routing decides
I tagged every inbound email by type before I fed it to Kryotta: simple size exchange, return window question, refund to different account, GDPR data request, complaint about condition, and multi-part query (two or more of the above in one email). I also noted the language—Dutch, German, English, French or Polish—and whether the customer had a previous return on file.
Auto routing works like this: you write your prompt once ("You are a customer service agent for an EU fashion retailer; reply in the customer's language, check our return policy, approve or escalate"), and Kryotta's router reads the incoming message and picks the model that balances speed, cost and capability for that specific query. It doesn't tell you in advance which model it'll use; you see the choice in the log after the reply is generated.
I scored each reply on three things. Accuracy: did it get the return window right, quote the correct GDPR retention period, spot the fraud flag? Tone: did it sound like a human who cared, or a chatbot reading terms and conditions? Efficiency: did the customer reply again, or did one email close the loop?
I also tracked cost per email and average response time, because if Auto routing picked expensive models for simple queries, the whole exercise would cost more than doing it myself.
Simple exchanges (91 emails): Gemini Flash and Llama won by speed
Ninety-one emails were straightforward: "I ordered a size M dress, I need size L, here's my order number." No edge cases, no policy questions, just swap approvals.
Auto routing picked Gemini Flash for 68 of these and Llama 3.3 70B for 23. It never chose Claude Sonnet or DeepSeek R1 for a simple exchange, which makes sense—those models cost more and think longer, and you don't need reasoning chains to approve a size swap.
Gemini Flash replies took 2.1 seconds on average and cost €0.003 per email. The tone was warm, the instructions were clear ("Reply to this email with your return label attached; we'll ship size L within two working days"), and only four customers wrote back with follow-up questions. Llama was slightly slower—2.8 seconds—and cost the same. The quality difference was invisible; I think Auto routing picked Llama when Gemini was under load.
One thing I noticed: both models occasionally invented a detail that wasn't in our return policy, like "You can also drop the parcel at any PostNL point" when we actually require customers to print a label and use DHL. It happened in 6 of the 91 emails. I added a line to my prompt ("Our return carrier is DHL only; do not mention PostNL or Budbee") and the problem stopped.
Verdict for simple exchanges: Gemini Flash is fast, cheap and good enough. Auto routing never wasted money on a heavy model here.
Return window questions (34 emails): Claude Sonnet for policy nuance
Thirty-four emails asked whether a return was still allowed: "I bought this coat on 18 January, today is 22 February, can I still return it?" or "The item arrived damaged; does the 14-day window still apply?"
Auto routing picked Claude Sonnet 4.5 for 29 of these and Gemini Flash for 5. The difference: Claude understood the nuance. EU consumer law gives you 14 days from delivery to start a return, but damaged goods fall under a separate warranty rule, and Zalando's own policy extends the window to 100 days if you're a premium customer.
Claude checked all three layers—our Shopify policy, the marketplace terms and EU law—and gave answers like: "Under EU consumer law your 14-day return period ended on 1 February, but because the item arrived with a broken zip, you're covered by our two-year warranty. We'll send a prepaid return label today." Gemini Flash gave the correct answer on simple date math, but it didn't flag the warranty exception in four out of five cases.
Cost per email: Claude Sonnet averaged €0.018, six times more than Gemini Flash, but only two customers replied with follow-ups (both to confirm their address), compared to nine follow-ups when I tested Gemini Flash manually on the same question set before I turned on Auto routing.
Verdict: Auto routing was right to spend more here. Policy nuance needs a model that reads context, and Claude's accuracy saved me from handling 18 extra emails.
GDPR data requests (11 emails): DeepSeek R1 for multi-step logic
Eleven emails mixed a return with a GDPR request: "I want to return this order and also delete all my personal data from your system" or "Send me a copy of everything you have on me, then process my refund."
Auto routing picked DeepSeek R1 for 9 of these and Claude Sonnet for 2. DeepSeek's reasoning mode showed its work: it listed the steps (confirm return eligibility, check GDPR retention rules, explain what data we can delete immediately versus what we must keep for tax records, generate the data export), then drafted a reply that walked the customer through each part without sounding like a legal document.
One reply started: "I've approved your return; you'll receive a label in the next email. For your data request: I can send you a copy of your order history, email correspondence and delivery addresses within 48 hours. We're required to keep your invoice details for seven years under German tax law, but I can anonymise your name and email in our marketing system today if you'd like." The customer replied "Perfect, thanks" and didn't escalate.
Claude Sonnet handled the two simpler GDPR cases well, but DeepSeek's step-by-step breakdown felt more transparent. Cost: DeepSeek averaged €0.041 per email—expensive—but these were the emails most likely to turn into complaints or legal threats if I got the answer wrong. The two cases where I manually checked DeepSeek's logic before sending, I found nothing to correct.
Verdict: Auto routing correctly identified high-stakes emails and paid for reasoning. I'd rather spend €0.04 and close the case than spend €0.01 and get a GDPR complaint.
Fraud flags (8 emails): Claude Sonnet caught all eight
Eight emails asked for refunds to a different bank account than the original payment, or claimed an item never arrived when tracking showed it was delivered and signed for. These are fraud attempts or confused customers; either way, you don't auto-approve.
Auto routing picked Claude Sonnet 4.5 for all eight. It flagged every one in the draft reply with a note like "[ESCALATE: refund account does not match order payment method]" and wrote a polite holding message: "I need to check a few details with our payments team before I can process this. I'll reply within 24 hours."
I manually reviewed all eight, confirmed six were fraud (same IP address, different names, requests within two hours of each other) and two were legitimate (customer had changed banks and forgot to update their Zalandra account). Claude never auto-approved a suspicious request, and it never accused a customer of fraud in the reply text.
Verdict: Auto routing's caution here saved me from at least €340 in fraudulent refunds. Claude's tone was firm but not hostile, which matters when two of the eight were real customers.
Multi-part queries (29 emails): Gemini Pro for speed, Claude for accuracy
Twenty-nine emails combined two or more issues: "I want to return item A for a refund and exchange item B for a different colour, and also my invoice has the wrong VAT rate."
Auto routing split these: Gemini Pro for 14, Claude Sonnet for 15. Gemini Pro was faster (3.4 seconds versus 4.1) and cheaper (€0.011 versus €0.018), but Claude's replies were more complete. Gemini missed the VAT question in three emails and had to be re-run. Claude caught every sub-question on the first pass.
If I were optimising purely for cost, I'd force Gemini Pro for everything. But the three emails Gemini missed cost me 20 extra minutes of back-and-forth, which is more expensive than the €0.021 I saved.
Verdict: Auto routing slightly favoured speed over accuracy here. I tweaked my prompt to say "If the email contains more than one question, list each one before answering" and Gemini's miss rate dropped to zero in the next batch.
Multi-language handling: no model struggled, but Claude's German was noticeably better
Seventy-one emails were in Dutch, 52 in German, 38 in English, 21 in French, 8 in Polish. Every model handled all five languages without me specifying a translation step in the prompt. I just wrote "Reply in the customer's language" and they did.
Claude's German felt the most natural—formal but warm, the way a real customer service agent writes. Gemini's Dutch was perfect. Llama occasionally used British English spelling ("colour", "favour") when the customer wrote American English, but no one complained. DeepSeek's Polish was correct but a bit stiff; one customer replied in English, which might mean they found the tone odd or just preferred English.
I didn't track this scientifically, but my gut says Claude Sonnet is the safest choice if you're replying in German to a customer who might escalate.
Cost and speed across 190 emails: €4.70 total, 3.1 seconds average
Total cost for 190 emails: €4.70. That's €0.025 per email on average. My previous workflow (DeepTranslate + Claude + Google Translate) cost me about €0.08 per email in tool subscriptions, and took 9 minutes of my time per email. Auto routing saved me 27 hours and €10.50 in direct costs.
Average response time: 3.1 seconds from the moment I pasted the email into Kryotta to the moment I had a draft reply. The slowest was a DeepSeek R1 GDPR case at 11 seconds. The fastest was a Gemini Flash size swap at 1.8 seconds.
Routing overhead—the time Kryotta's system takes to read the query and pick a model—added about 0.4 seconds per email. Unnoticeable in practice, and worth it because I didn't waste money on overkill models for simple queries.
Would I use Auto routing again, and what I'd change
Yes. I've left it on, and I've handled another 60 emails since this test ended. The system's choices make sense: cheap and fast for simple stuff, expensive and careful for high-stakes cases. I trust it enough now that I only review the replies before sending; I don't second-guess the model choice.
Three things I'd change. First, I'd add a custom instruction for each marketplace: "If the order is from Zalando, mention their 100-day return window for premium customers; if it's from Bol.com, returns are 30 days, no exceptions." Auto routing doesn't read your Shopify tags automatically; you have to tell it.
Second, I'd set a manual review flag for any email that mentions a refund to a different account. Claude caught all eight fraud attempts in my test, but I'd rather spend 30 seconds checking than risk a €50 mistake.
Third, I'd test Mistral Large on the next batch of multi-part queries. I didn't see it chosen once in this test, which suggests Auto routing doesn't rate it for EU customer service work, but I'd like to know why.
If you're running a marketplace business in the EU and you're handling more than 20 return emails a week, Auto routing pays for itself in time saved. If you're doing fewer than that, you probably don't need it—just pick Gemini Flash for everything and call it done.
Questions people ask
Does Auto routing work if my emails come from Zalando and Bol.com directly, or do I need to forward them?
You forward them, or connect your support inbox if you handle marketplace messages through a shared email address. Kryotta doesn't integrate directly with Zalando or Bol.com's backend, so you're either pasting emails one by one or using an email connector.
Can I force Auto routing to avoid a specific model, like if I don't want to pay for DeepSeek R1?
Not yet. Auto routing picks from the full roster. If cost is tight, turn off Auto routing and just pick Gemini Flash or Llama manually for everything—you'll save money but lose the accuracy boost on complex cases.
Which model does Auto routing pick for English-language emails from UK customers post-Brexit?
Same logic as EU emails: Gemini Flash or Llama for simple queries, Claude Sonnet for policy questions. The model doesn't care about the customer's country; it cares about the complexity of the question. You'd just adjust your prompt to mention UK-specific return rules (14 days under UK consumer law, no change from EU law there).
Does Auto routing get better over time, or is it the same logic every day?
Same logic. It doesn't learn from your feedback yet. If you notice it's picking the wrong model for a certain type of query, you tweak your prompt to be more explicit, and that fixes it.
I ran this test because I was drowning in multilingual returns and I needed a system that didn't require me to choose a model every time. Auto routing worked: it saved me 27 hours, cost less than €5, and caught eight fraud attempts I might have missed. If you're handling EU marketplace support and you're still doing it manually, try it for a week—you can connect your inbox at Kryotta and see which models it picks for your queries.
