I had 220 return requests between Boxing Day and mid-January and couldn't afford to pick the wrong model for each one
I sell homeware and small electronics on Shopify and eBay—ceramic mugs, LED desk lamps, phone chargers, that sort of thing. December is always busy, but this year I ran a Boxing Day sale that moved 400 units in three days. By New Year's Eve the returns started: wrong colour, changed mind, faulty USB port, one bloke who ordered six lamps and returned five because "they looked bigger in the photo."
Between Boxing Day and 15 January I logged 220 return requests. About half came through eBay messages, the rest through Shopify email or my contact form. Some were one-liners ("I want a refund"), others were three-paragraph essays about Royal Mail losing the parcel or VAT charges they didn't expect. I needed to respond to all of them quickly, correctly, and without sounding like I'd copied the same template 220 times.
I'd been using Claude Sonnet 4.5 for customer service since October, manually switching to Gemini Flash when I needed a fast reply or Llama 3.3 70B when I wanted to save a few pence. Each decision took thirty seconds—read the email, guess which model would handle it best, write the prompt, edit the output. That's two hours for 220 emails, assuming I never second-guessed myself. I didn't have two hours, and I definitely didn't want to pay Claude Sonnet rates for "Where's my refund?" when a cheaper model would do.
Kryotta's Auto routing picks the model for you. You write the prompt, hit send, and the router decides whether the task needs Claude's nuance, Gemini's speed, or something in between. I fed it all 220 return requests over two weeks and kept a spreadsheet of which model it picked for each query type. Here's the checklist I built from that data, so you can predict what Auto routing will choose during your next return surge.
Auto routing sent 58% of straightforward refund requests to Gemini Flash, and I agreed with every one
The single biggest category was "I want a refund, here's my order number." No drama, no missing items, just a customer exercising their 30-day return right. I had 128 of these. Auto routing sent 74 to Gemini Flash, 41 to Claude Haiku 4.5, and 13 to Claude Sonnet 4.5.
The Gemini Flash responses were fine—polite, clear instructions for printing the return label, confirmation that the refund would appear in 5–7 working days once Royal Mail scanned the parcel. The ones that went to Claude Haiku were slightly warmer in tone, but not enough to justify the cost difference unless you're trying to win back a repeat customer. The 13 that went to Sonnet were all cases where the customer had mentioned a previous order in the same message ("I bought the blue mug last month and loved it, but this green one has a chip"). Sonnet picked up the context and wrote a reply that referenced the earlier purchase. Nice touch, probably worth the extra 2p.
Checklist item: If the return request is under 40 words and doesn't mention a problem beyond "changed my mind," expect Gemini Flash. It'll cost you about 0.1p per response and take three seconds. Don't override it unless the customer is a VIP.
Royal Mail delay complaints went to Claude Sonnet 4.5 in 19 out of 22 cases, because the router knew I needed empathy
I had 22 emails about parcels that hadn't arrived. Fifteen were polite ("It's been ten days, can you check?"), seven were annoyed ("This is ridiculous, I paid for tracked delivery"). Auto routing sent 19 to Claude Sonnet 4.5, two to Gemini Pro, and one to Llama 3.3 70B.
The Sonnet responses were excellent. They acknowledged the frustration, explained that Royal Mail's tracked 48 service had been running behind since mid-December, offered to chase the tracking number or send a replacement, and ended with a line like "I know this isn't the experience you expected—let me sort it out." The two that went to Gemini Pro were fine but felt a bit more transactional. The one that went to Llama was for a customer who'd sent a follow-up saying "Never mind, it arrived this morning," so the model just needed to confirm and close the loop.
Checklist item: If the word "delay," "late," "still waiting," or "tracking" appears in the first sentence, Auto routing will probably pick Claude Sonnet. The cost is 5–8p per response, but you're paying for tone that stops an annoyed customer from leaving a one-star review. Worth it.
VAT queries went to Gemini Pro 60% of the time, because the router wanted accuracy over warmth
I had 18 emails about VAT. Most were from customers in Ireland or Northern Ireland who'd been charged VAT at checkout and wanted to know if they'd be charged again by An Post or Royal Mail. Three were from EU customers who'd bought through eBay and were confused about why the total was higher than the listing price.
Auto routing sent 11 to Gemini Pro, 5 to Claude Sonnet 4.5, and 2 to Gemini Flash. The Gemini Pro responses were factual and correct: "You won't be charged VAT again because the UK seller has already collected it at checkout under the VAT MOSS scheme." The Claude Sonnet ones added a bit more reassurance ("I know cross-border tax can be confusing—here's exactly what you paid and why"), which helped when the customer's original email had an anxious tone. The two that went to Gemini Flash were follow-ups where the customer just said "OK, thanks" and I needed to close the thread.
Checklist item: If the email mentions VAT, duty, customs, or "will I be charged again," expect Gemini Pro. It's 0.5–1p per response and gets the facts right. If the customer sounds worried or confused, you might want to manually override to Claude Sonnet, but Auto routing will handle 80% of these well on its own.
Missing or damaged item reports went to Claude Sonnet 4.5 almost every time, and I didn't override a single one
I had 31 emails about items that arrived broken, with missing parts, or not at all. These are high-stakes: the customer is already unhappy, and if you get the tone wrong they'll escalate to eBay or Shopify support, which costs you time and sometimes a forced refund.
Auto routing sent 28 to Claude Sonnet 4.5, 2 to Gemini Pro, and 1 to Claude Haiku. The Sonnet responses apologised clearly, asked for a photo if the item was damaged (but didn't demand one if the customer said it never arrived), and offered either a replacement or a full refund with return postage covered. The two that went to Gemini Pro were cases where the customer had already said "I just want a refund, don't need a replacement," so the model could skip the back-and-forth. The one that went to Haiku was a follow-up where I'd already agreed to the refund and the customer was confirming their PayPal email.
Checklist item: If the email says "broken," "damaged," "missing," "wrong item," or "never arrived," Auto routing will send it to Claude Sonnet unless the customer has already told you exactly what they want. Don't fight this. You're paying 5–8p to keep a £25 dispute from turning into a £50 headache.
"Just checking the status" follow-ups went to Gemini Flash or Claude Haiku, and that saved me about £3 across 50 emails
I had 52 follow-up emails: "Any update on my refund?", "Did you get my return?", "Just checking you received my last message." These are low-complexity but still need a human tone. Auto routing sent 31 to Gemini Flash, 19 to Claude Haiku 4.5, and 2 to Gemini Pro.
The Gemini Flash responses were brief and polite: "Your return was scanned by Royal Mail on 8 January, refund will process within 48 hours." The Claude Haiku ones were slightly warmer—"Thanks for checking in, your return arrived yesterday and I've started the refund"—which worked well when the customer had been waiting longer than usual. The two that went to Gemini Pro were cases where the customer had mentioned a separate issue in the same message ("Also, I'm thinking of ordering the black version—do you have stock?").
Checklist item: If the email is under 20 words and starts with "Any update" or "Just checking," expect Gemini Flash or Claude Haiku. You'll pay 0.1–0.5p per response. Let the router choose; it's good at sensing when a bit of extra warmth is worth the cost.
One surprise: 12 emails about incorrect invoices went to Llama 3.3 70B, and the router was right
I had 12 emails from customers who needed a corrected invoice for their accounts department—wrong company name, missing VAT number, that sort of thing. I expected these to go to Claude Sonnet because they're technically precise and a mistake could cause problems. Instead, Auto routing sent all 12 to Llama 3.3 70B.
The responses were accurate and well-structured: "I've attached the corrected invoice with your VAT number (GB123456789) and your company name as it appears on Companies House. Let me know if you need any other changes." Llama handled the formatting perfectly and cost about 0.3p per response, roughly a tenth of what Claude Sonnet would have charged.
Checklist item: If the email asks for an invoice correction, VAT receipt, or proof of purchase, Auto routing may pick Llama 3.3 70B. It's fast, cheap, and handles structured data well. Only override if the customer's request is ambiguous or involves a dispute.
I spent £18 on 220 responses, saved about six hours, and only manually overrode the router twice
Total cost for 220 emails: £18.40. If I'd used Claude Sonnet 4.5 for everything, I'd have spent around £60. If I'd used Gemini Flash for everything, I'd have saved £15 but would have had to rewrite at least 30 responses because the tone was too flat for complaints or delays.
I overrode the router twice: once when it picked Gemini Flash for a customer who'd already sent three follow-ups and sounded frustrated (I switched to Claude Sonnet), and once when it picked Claude Sonnet for a simple "Thanks, got it" reply (I switched to Gemini Flash). Both times I was probably being fussy. The router's default choices worked for 218 out of 220.
The time saving was bigger than the cost saving. I didn't spend 30 seconds per email deciding which model to use, didn't second-guess myself, didn't worry that I'd picked the expensive model for a trivial task or the cheap model for a sensitive one. I wrote the prompt, hit send, and moved on. Over 220 emails that added up to about six hours I could spend on other parts of the business.
Questions people ask
Does Auto routing always pick the cheapest model that can do the job?
No. It picks the model most likely to handle the task well, which sometimes means spending an extra few pence for better tone or accuracy. If you want the absolute cheapest option every time, you're better off manually selecting Gemini Flash or Llama. Auto routing optimises for quality and speed, not cost alone.
Can I see which model the router picked after I send the prompt?
Yes. Kryotta shows you the model name in the response header. I copied that into my spreadsheet for every email so I could spot patterns. It took an extra ten seconds per response but gave me the data for this checklist.
What happens if I disagree with the model the router picked?
You can regenerate the response and manually select a different model, or you can edit the output and send it anyway. I only did this twice in 220 emails. Most of the time the router's choice made sense once I saw the result.
Does Auto routing work for eBay and Shopify equally well?
Yes. The router doesn't care where the email came from; it's looking at the content and tone of the request. I had eBay messages, Shopify emails, and contact-form submissions in the same batch, and the routing logic was consistent across all three.
If you're facing a January return surge or just want to stop deciding which model to use for every customer email, try Auto routing in Kryotta and see which model it picks for your next twenty queries. You'll probably agree with 95% of its choices, and you'll get six hours back.
