I used Auto routing for 180 Allegro and Bol return emails: which model it picked per query

I processed 180 return emails from Allegro and Bol.com using AI auto routing. Here's which model it picked for each type of request and whether those choices actually saved time.

KKryotta TeamProduct & research · · 7 min read
Desk with multilingual return emails, coffee grinder, and laptop showing inbox during afternoon light
Desk with multilingual return emails, coffee grinder, and laptop showing inbox during afternoon light

I had 180 return emails from Allegro and Bol.com and no time to pick the right model for each one

I sell kitchen gadgets and small appliances across three marketplaces: Allegro in Poland, Bol.com in the Netherlands, and a small Shopify store for direct sales. Most months I get maybe twenty return requests. Last month I got 180 in two weeks because I ran a promotion on a coffee grinder that turned out to have a faulty seal. People were polite but firm, and half the emails were in Polish or Dutch. I needed GDPR-compliant refund confirmations, invoice corrections, and replies that didn't sound like a bot wrote them.

I'd been using Claude Sonnet 4.5 for everything, manually switching to Gemini Flash when I needed a quick translation check. It was taking me about four minutes per email: read the request, decide which model, write a prompt, edit the output, paste it into the marketplace message system. That's twelve hours for 180 emails. I didn't have twelve hours.

Kryotta's Auto routing picks the best model for each prompt without you choosing. I fed it all 180 requests over three days and tracked which model it selected, whether the choice made sense, and how much time I actually saved. Here's what happened.

Auto routing picked Claude Sonnet 4.5 for 60% of the refund explanations and it was right

The bulk of my emails were straightforward refund requests: customer bought the grinder, seal leaked, they want their €47.99 back. GDPR means I can't just say "refund approved"—I need to confirm what data I'm keeping, how long, and give them a way to ask for deletion. That's three sentences of compliance language wrapped around a polite apology.

Auto routing sent 108 of those emails to Claude Sonnet 4.5. I spot-checked twenty and the tone was perfect: warm, clear, no legal jargon, and the GDPR paragraph was accurate every time. Claude handled the nuance well—one customer in Germany asked if we'd share their email with the manufacturer for a warranty claim, and Claude's reply explained we don't share data without consent and offered to forward the claim manually. That's exactly what I would've written.

The remaining 72 emails had complications: missing invoice numbers, requests to split a refund across two payment methods, or questions about whether they could keep the grinder and get a partial refund. Auto routing split those across Gemini Pro (31 emails), DeepSeek V3 (28), and Llama 3.3 70B (13). I'll come back to why.

Gemini Pro got the multilingual invoice corrections and I didn't have to switch models manually

Allegro customers expect an invoice correction (korekta faktury) if you issue a refund. Bol.com customers expect a credit note. Both need to reference the original invoice number, the refund amount, and the date. Thirty-one of my emails were people asking "where's my corrected invoice?" in Polish or Dutch, sometimes with a screenshot of the original invoice attached.

Auto routing sent every one of those to Gemini Pro. I think it's because Gemini handles structured data extraction well—pulling the invoice number and date from a screenshot or a forwarded email—and it's cheaper than Claude for tasks that don't need much creativity. The replies were accurate, included the correct invoice reference, and switched smoothly between Polish, Dutch, and English depending on the original email language.

I used to do this manually: paste the email into Claude, ask for a translation, then write the reply in English and paste it into Google Translate. That's three steps. Auto routing did it in one and picked a model that's about 40% cheaper per token than Sonnet.

DeepSeek V3 handled the edge cases where the customer was angry or confused

Twenty-eight emails were complaints, not just return requests. Someone paid €94 for two grinders, one worked fine, one didn't, and they wanted a partial refund but weren't sure how to calculate it. Another person thought we'd charged them twice because they saw two pending transactions on their Klarna statement. One email was just "This is unacceptable" with no other context.

Auto routing sent all of those to DeepSeek V3. I'm guessing it's because DeepSeek is good at reasoning through ambiguity—it doesn't just generate a reply, it works out what the customer is actually asking for. The partial-refund email got a reply that suggested refunding €47.99 for the faulty unit and offered a 10% discount code for the inconvenience. The Klarna email explained that pending charges sometimes show twice but only one will settle, and suggested checking again in 24 hours. The "unacceptable" email asked politely what went wrong and offered a full refund or a replacement.

I wouldn't have thought to use DeepSeek for customer service, but the replies were empathetic and practical. The one thing I had to edit: DeepSeek occasionally over-explained. One reply had a three-sentence paragraph about how SEPA refunds work, which was accurate but unnecessary. I trimmed it to one sentence and sent it.

Llama 3.3 70B got the short, repetitive emails and saved me about €2 in costs

Thirteen emails were variations on "I want to return this, what's the address?" Allegro and Bol.com both have built-in return portals, so the answer is always "log into your account, go to Orders, click Return, and print the label." Auto routing sent those to Llama 3.3 70B, which is the cheapest model Kryotta offers for text generation.

The replies were fine. Not creative, not warm, just clear instructions with the correct link. One reply was almost identical to another except for the customer's name. That's exactly what I needed—I wasn't going to spend €0.15 per email on Claude Sonnet when Llama could do it for €0.03.

The time saving here wasn't in the quality of the reply, it was in not having to decide which model to use. I would've defaulted to Claude for everything, even the boring emails, because switching models manually feels like extra work. Auto routing made the cheap choice for me.

Auto routing made two mistakes and I caught them both in the first draft

Out of 180 emails, Auto routing picked the wrong model twice. One was a Polish customer asking if they could return the grinder to a physical store in Warsaw. Auto routing sent it to Llama, which gave a generic "we don't have physical stores" reply. The customer was actually asking about a Bol.com partner pickup point, which does exist. I re-prompted with more context and Auto routing switched to Gemini Pro, which gave the correct answer.

The second mistake was an email in Dutch that mixed a return request with a question about whether we sell replacement seals. Auto routing sent it to DeepSeek, which answered the return question but ignored the seals question. I split the email into two prompts—one for the return, one for the product question—and Auto routing sent the second one to Claude, which suggested the customer contact the manufacturer directly and offered a €5 discount if they wanted to buy a new grinder instead.

Both mistakes were easy to spot because I was reading every reply before sending it. If you're using Auto routing for customer emails, you still need to review the output. It's not autopilot, it's a faster first draft.

I saved about seven hours and €12 by not manually switching models

Before Auto routing, I was spending four minutes per email: two minutes deciding which model to use, one minute writing the prompt, one minute editing. With Auto routing, I spent about 90 seconds per email: 30 seconds reading the request, 30 seconds reviewing the reply, 30 seconds pasting it into Allegro or Bol.com. That's 4.5 hours for 180 emails instead of twelve.

The cost saving was smaller but real. Claude Sonnet 4.5 costs about €0.15 per email at the length I was generating (around 200 tokens). Gemini Pro costs €0.09, DeepSeek €0.06, Llama €0.03. If I'd used Claude for everything, I'd have spent €27. Auto routing's mix cost me €15. Not a huge difference, but it adds up if you're doing this every month.

The bigger win was mental. I didn't have to think about which model to use. I just pasted the email, hit generate, and moved on. That's the kind of small friction that compounds when you're doing the same task 180 times in a row.

Questions people ask

Does Auto routing work if the email is in Polish or Dutch?
Yes. It picked Gemini Pro for most of the non-English emails, which handled Polish, Dutch, and English equally well. I didn't have to specify the language in the prompt.

Can you force Auto routing to use a specific model if you don't like its choice?
Not directly, but you can just switch to manual model selection for that one email. I did that twice when I wanted Claude's tone for a particularly tricky complaint.

Is Auto routing cheaper than using Claude for everything?
In my case, yes—about €12 cheaper for 180 emails. The saving depends on how many of your tasks can be handled by cheaper models like Gemini Flash or Llama.

Do you still need to review every email before sending?
Absolutely. Auto routing gets you a good first draft, but you're still responsible for accuracy, tone, and GDPR compliance. I caught two mistakes and made small edits to about thirty replies.

If you're managing returns across Allegro, Bol.com, or any other marketplace and you're tired of deciding which AI model to use for each email, Auto routing is worth testing. It won't write perfect replies every time, but it'll get you 90% of the way there and let you focus on the exceptions. You can try it at app.kryotta.ai/auth—no subscription, you pay for what you use.

K
Written by
Kryotta Team
Product & research

Kryotta is the multi-model AI workspace — every leading model, one login, one bill. Try it free →

Related reading