I started with 220 Boxing Day return messages, three policy edge cases, and no patience left to choose a model for each one
I sell homeware on eBay—ceramic mugs, linen tea towels, scented candles, that sort of thing. My workshop is in Bristol, I pack everything myself, and between Black Friday and Christmas I shifted about £18,000 worth of stock. On 27 December the returns started. By 2 January I had 220 messages sitting in my eBay inbox: refund requests, damaged item claims, people who'd ordered the wrong size candle holder, one person who said the mug "didn't spark joy" and wanted me to pay return postage.
Every message needed a different reply. Refund requests were straightforward—check the order date, confirm it's within 30 days, send the return label. Damaged items needed photos and a decision on whether to refund immediately or wait for the item to come back. Buyer's remorse sat in a grey area: eBay's policy says I have to accept returns, but I don't have to cover postage unless the item's faulty. Then there were the VAT questions. A business buyer in Dublin wanted to know if the refund included Irish VAT. Another buyer claimed I'd charged them VAT twice because they'd used a forwarding service in Northern Ireland.
I'd been using Claude Sonnet for most customer service emails. It's good—polite, clear, doesn't waffle. But it costs more per message than I wanted to spend on 220 returns, and some of the simpler queries didn't need that much reasoning. I'd tried Compare mode a few times, where you send the same prompt to three models and pick the best answer, but that felt like overkill when I just wanted one decent reply per message. Then I remembered Kryotta has Auto routing: you write the prompt, it picks the cheapest model that can handle it, and you only pay for what you use.
I decided to run all 220 messages through Auto routing and track which model it chose for each type of query. I wanted to see if it would save me money, whether it would catch the policy edge cases Compare would have missed, and what would go wrong.
I sorted the messages into four piles and wrote one prompt template for each
I exported my eBay messages into a spreadsheet—date, buyer name, order number, message text. Then I tagged each one by type: straightforward refund, damaged item, buyer's remorse, or VAT query. The breakdown was 112 straightforward refunds, 68 damaged items, 31 buyer's remorse, and 9 VAT queries.
For straightforward refunds I wrote this prompt:
"The buyer wants to return [item name] from order [number], purchased on [date]. Our policy allows returns within 30 days. Draft a reply confirming we'll issue a full refund once we receive the item, include a Royal Mail return label, and keep the tone warm but not overly apologetic."
For damaged items:
"The buyer says [item name] from order [number] arrived damaged. They've attached [photo or no photo]. If there's a photo, assess whether the damage looks like shipping or manufacturing, then decide whether to refund immediately or ask them to return it first. If no photo, ask for one politely. Explain our process and keep the tone reassuring."
For buyer's remorse:
"The buyer wants to return [item name] from order [number] because [reason]. Our policy accepts returns but doesn't cover postage unless the item is faulty. Draft a reply explaining they're welcome to return it at their own cost, provide return address, and keep the tone friendly but firm."
For VAT queries:
"The buyer is asking [specific VAT question] about order [number]. They're in [location]. Check if the original invoice included UK VAT at 20%, whether the buyer is VAT-registered, and whether the refund should include or exclude VAT. Draft a clear reply with the correct figures and explain the VAT treatment."
I pasted each prompt into Kryotta's Auto routing workspace one at a time. Auto routing picked the model, I copied the output into eBay, tweaked it if needed, and noted which model it had chosen.
Auto routing picked Gemini Flash Lite for 80% of straightforward refunds and I saved £11 in three days
Out of 112 straightforward refund requests, Auto routing sent 89 to Gemini Flash Lite, 18 to Llama 3.3 70B, and 5 to Claude Haiku 4.5. The Flash Lite replies were short, clear, and covered everything: confirmation of the refund, return label instructions, expected processing time. They didn't add any personality, but they didn't need to. The buyer just wanted to know I'd received their message and would sort it.
Llama 3.3 70B got picked when the message was slightly longer or included a follow-up question—"Can I return two items from the same order?" or "I've already posted it back, when will I get my money?" Those replies were a bit warmer, answered both parts of the question, and still came in under budget.
Claude Haiku 4.5 showed up five times when the buyer's message had an edge case. One person wanted a partial refund because they were keeping one mug but returning the other three. Another had ordered a candle as a gift, it arrived late, and they wanted a refund plus compensation for the "emotional distress" of not having a Christmas present. Haiku handled those without me needing to rewrite the prompt—it picked up on the complication and adjusted the tone.
I compared the cost. If I'd sent all 112 messages to Claude Sonnet, I'd have spent about £16. Auto routing cost me £5. That's £11 saved, and the replies were just as good for the task.
It sent every damaged item claim to Claude Sonnet and caught two policy edge cases I'd have missed
All 68 damaged item messages went to Claude Sonnet 4.5. I was surprised at first—Sonnet's the most expensive option—but when I looked at the outputs I understood why. Damaged items need judgement. Is the crack in the mug from rough handling or a manufacturing flaw? Should I refund immediately or wait for the item to come back? If the buyer's furious, how do I de-escalate without admitting liability?
Sonnet read the message, assessed the tone, checked if there was a photo attached, and made a call. When the buyer had included a clear photo of a smashed candle jar, Sonnet drafted a reply offering an immediate refund and a discount code for the next order. When there was no photo, it asked for one politely and explained we'd process the refund as soon as we'd seen the damage.
Two messages had edge cases I'd have missed if I'd been skimming. One buyer said the tea towel had arrived "stained" but the photo showed a print misalignment—our manufacturer's fault, not shipping damage. Sonnet flagged that and suggested I refund without asking for a return, because it wasn't worth the postage cost and the item couldn't be resold. The other was a buyer who claimed a mug was chipped, but the chip was tiny and in a spot that looked like normal wear. Sonnet drafted a reply offering a partial refund or a return, and explained our policy doesn't cover minor cosmetic issues that don't affect function. Both replies saved me from either losing money or starting an argument.
Compare mode would have shown me three answers, but I'd still have had to choose. Auto routing just picked Sonnet, handled the nuance, and moved on.
Buyer's remorse went to Llama and Haiku, and one refund calculation went wrong because I didn't specify VAT treatment
Thirty-one buyer's remorse messages split between Llama 3.3 70B (22 messages) and Claude Haiku 4.5 (9 messages). Llama handled the simple ones—"I ordered the wrong colour", "It doesn't match my kitchen"—with a friendly but firm reply explaining the return process and that they'd need to cover postage. Haiku got the trickier ones: a buyer who wanted a refund but hadn't actually opened the parcel yet, another who'd used the candle once and then decided they didn't like the scent.
The Haiku replies were better at holding the line without sounding cold. One message said, "You're welcome to return the candle, but since it's been used we can only offer a partial refund of 50% to cover the cost of disposal and restocking." That's not in my official policy—I hadn't thought to write one—but it felt fair and Haiku explained it clearly.
Then I hit the VAT problem. A buyer in Dublin had ordered £60 worth of tea towels, returned them, and asked for a refund. I sent the prompt to Auto routing, it picked Llama, and the reply confirmed a refund of £60. The buyer wrote back: "I'm VAT-registered in Ireland, shouldn't the refund be £50 ex-VAT?" She was right. I'd charged her £60 including UK VAT, but because she's a business buyer in the EU, the refund should have been £50 and she'd reclaim the Irish VAT separately.
I went back and rewrote the prompt to specify: "Check if the buyer is in the UK, Northern Ireland, or the EU. If they're VAT-registered and outside the UK, calculate the refund excluding UK VAT at 20%. If they're a consumer, refund the full amount including VAT." Auto routing picked Claude Haiku for the rewrite, and the second reply got it right.
That's on me, not the model. I should have been clearer in the first prompt. But it's a reminder that Auto routing is fast, not psychic.
The nine VAT queries all went to Claude Sonnet and one reply saved me from a £140 mistake
Nine messages asked about VAT. Three were from business buyers in Ireland wanting to confirm the VAT treatment. Two were from Northern Ireland buyers confused about whether they'd been charged UK or Irish VAT. One was from a buyer in the Netherlands who'd used a UK forwarding service and thought I'd charged her VAT twice. Three were from UK buyers who just wanted a VAT invoice.
Auto routing sent all nine to Claude Sonnet 4.5. The replies were detailed, accurate, and explained the rules without sounding like a tax guide. For the Irish business buyers, Sonnet confirmed the refund would be ex-VAT and they'd reclaim Irish VAT from Revenue. For the Northern Ireland buyers, it explained that under the protocol they're treated as UK for VAT purposes on goods, so the charge was correct. For the Netherlands buyer, it walked through the invoice line by line and showed that she'd only been charged UK VAT once; the forwarding service had added their own handling fee, which looked like a second VAT charge but wasn't.
The one that mattered most was a buyer in Dublin who'd ordered £180 worth of candles for her shop. She wanted to return half the order and asked for a refund of £90. Sonnet's reply calculated the refund as £75 ex-VAT, explained why, and noted that because the original order was over the UK's £135 distance-selling threshold, I should have charged Irish VAT at 23% instead of UK VAT at 20%, but since I hadn't, the refund should still be based on what she actually paid.
That reply saved me from issuing a £90 refund when the correct figure was £75. It also flagged a compliance issue I hadn't spotted: I should have registered for Irish VAT once my sales to Ireland hit €35,000. I'm not there yet, but I'm close, and now I know to watch it.
I processed 220 returns in three days, spent £22 instead of £50, and Auto routing picked the right model for the job every time
Auto routing didn't choose randomly. It read the prompt, assessed the complexity, and picked the cheapest model that could handle it. Straightforward refunds went to Flash Lite because they didn't need reasoning—just a polite confirmation and some instructions. Damaged items went to Sonnet because they needed judgement. Buyer's remorse split between Llama and Haiku depending on how awkward the message was. VAT queries went to Sonnet because getting the numbers wrong would cost me more than the API call.
I spent £22 across 220 messages. If I'd used Claude Sonnet for everything, it would have been £50. If I'd used Compare mode, I'd have spent even more and still had to choose the best answer for each one. Auto routing just worked.
The one thing it didn't do was catch my VAT mistake in the first Dublin refund. That's not a flaw in Auto routing—it's a reminder that the prompt matters. If you don't tell the model what to check, it won't check it. Once I rewrote the prompt to include VAT treatment, the replies were spot on.
I'm using Auto routing for all my eBay messages now, not just returns. It's faster than choosing a model myself, cheaper than defaulting to Sonnet, and I haven't had a single reply that needed a major rewrite.
Questions people ask
Does Auto routing work for messages that need attachments or photos?
It reads the prompt and the message text, but it doesn't analyse photos directly. If a buyer sends a photo of a damaged item, you'll need to describe what you see in the prompt—"the mug is cracked along the handle" or "the candle jar is shattered"—and Auto routing will draft the reply based on that. It won't look at the image and assess the damage for you.
Can I force Auto routing to pick a specific model if I don't trust its choice?
No, that's not how it works. Auto routing picks the model based on the prompt's complexity and cost. If you want control, use Compare mode or just select the model manually. But I found that once I trusted it for a few dozen messages, I stopped second-guessing the choice.
What happens if Auto routing picks a cheap model and the reply is rubbish?
You rewrite the prompt and send it again. I had that happen twice—both times the prompt was too vague and Flash Lite gave a generic answer. I added more detail, Auto routing picked Haiku the second time, and the reply was fine. It's faster to rewrite a prompt than to manually choose a model for every message.
Is Auto routing cheaper than just using Gemini Flash for everything?
Sometimes. Flash is cheap, but it's not always good enough. Auto routing will use Flash when it can and upgrade to Llama or Haiku when it needs to. Across 220 messages I saved about £28 compared to using Sonnet for everything, and about £6 compared to using Flash for everything and manually rewriting the failures. The saving depends on how many of your messages actually need the expensive models.
I processed 220 post-Boxing Day returns in three days without choosing a model once. If you're handling a January sales surge and you're tired of guessing which AI to use for each message, try Auto routing in Kryotta and let it pick for you.
