The setup: 165 return messages, three models, one routing experiment
I sell handmade candles and wax melts through Etsy. We do about $18,000 a month, but November and December account for nearly half our annual revenue. This year, Cyber Monday alone brought in $9,200—our best single day ever. Then the returns started.
Between December 3rd and December 18th, I dealt with 165 return requests. Some were straightforward ("this scent gives me a headache, can I swap it?"), some required careful handling ("the candle arrived shattered"), and a handful were just confusing ("I ordered three but only want two now, do I return all three or just one?"). I used to answer every message myself, which meant spending two to three hours a day in the Etsy Messages tab during the holiday surge.
This year I tried Auto routing in Kryotta. The idea: feed it a return message, let it decide whether the reply needs Claude's nuance, Gemini's speed, or Llama's straightforward logic, then generate a draft I could edit and send. I set up routing rules based on query complexity, tracked which model it picked for each of the 165 messages, and noted the three times it made the wrong call. I'll walk through the ruleset I built, what worked, what didn't, and the adjustments I made halfway through December.
How Auto routing decides which model to use
Auto routing looks at the query and scores it on three axes: emotional complexity (is the customer angry, confused, or just asking a factual question?), factual density (does the answer require pulling together multiple policies, dates, and tracking numbers?), and ambiguity (is the request clear or does it need interpretation?). Based on those scores, it routes to the model best suited to the task.
I set up three tiers. Claude Sonnet 4.5 for high-complexity cases: angry customers, multi-part questions, or anything involving a judgment call about our return policy. Gemini Pro for medium-complexity: straightforward returns that still needed a warm tone and a two-paragraph explanation. Llama 3.3 70B for simple factual replies: "here's our return address," "yes, you can exchange that scent," "your refund will process in three to five business days."
The routing rules looked like this:
- If the message contains words like "disappointed," "frustrated," "never ordering again," or describes damage → Claude.
- If the message asks about timelines, exchange options, or combines two questions ("can I return this AND get a different scent?") → Gemini.
- If the message is a single factual question with no emotional content → Llama.
I also added a manual override: if Auto routing picked Llama but the message felt off to me when I skimmed it, I could re-route to Gemini before generating the draft.
The first 60 messages: Auto routing worked better than I expected
In the first week after Cyber Monday, I processed 60 return messages. Auto routing sent 18 to Claude, 31 to Gemini, and 11 to Llama. I edited every draft before sending it, but the edits were minor—usually just adding the customer's name or adjusting one sentence to match our shop's voice.
Here's a typical Claude case. A customer ordered a set of six candles as gifts, and two arrived with cracked jars. She was polite but clearly annoyed, and her message ended with "I really wanted these for my holiday party on the 15th, so I'm not sure what to do now." Auto routing flagged the emotional weight and the time pressure, routed it to Claude, and generated a reply that apologized, offered to send replacements via Priority Mail at no charge, and included a 15% discount code for her next order. I changed one phrase ("we're so sorry" to "I'm so sorry") and sent it. She replied two hours later saying the replacement plan worked and thanked me for the quick response.
Gemini handled the bulk of the straightforward exchanges. A customer wanted to swap a lavender candle for a eucalyptus one; Gemini's draft explained our exchange process, confirmed we had eucalyptus in stock, and gave her the return shipping address. I added one line ("just drop the original candle in any box—doesn't need to be the one it came in") and hit send. Total time: ninety seconds.
Llama took care of the factual questions. "Do I need to include the receipt when I return this?" Llama's draft: "No receipt needed—just include your order number on a slip of paper, or write it on the outside of the box. We'll process your refund within three business days of receiving the return." I sent it as-is.
The three times Auto routing picked the wrong model
Around message 73, I hit the first routing mistake. A customer wrote: "I ordered the winter sampler set but I'm allergic to cinnamon. Can I return just the cinnamon ones and keep the rest?" Auto routing scored it as low-complexity—single question, no emotional language—and sent it to Llama. Llama's draft was factually correct ("yes, you can return individual items from a set") but missed the nuance: our return policy technically allows partial returns, but we lose money on shipping if someone returns a $6 candle from a $40 set. I wanted the reply to gently suggest an exchange instead, which required a judgment call Llama doesn't make.
I re-routed it to Claude and regenerated. Claude's draft acknowledged the allergy, confirmed she could return just the cinnamon candles, but added: "If you'd like, I can also send you a different scent to replace them—no need to return anything, and no extra charge. Just let me know which scent you'd prefer." She took the exchange. I updated my routing rules to flag any message containing "allergy," "sensitive," or "can I return just" as medium-to-high complexity.
The second mistake happened at message 108. A customer wrote three sentences: "The candle I got smells nothing like the description. I'm really disappointed. I want a refund." Auto routing saw "disappointed" and routed it to Claude, which was correct—except Claude's draft spent four sentences apologizing and explaining that scent perception varies by person. The customer didn't want an explanation; she wanted her $18 back. I shortened Claude's reply to two sentences ("I'm sorry the scent didn't work for you. I've processed your refund—it'll appear in your account in three to five business days") and noted that Claude sometimes over-explains when a customer just wants resolution. I added a rule: if the message is under 20 words and contains "I want a refund," route to Gemini instead. Gemini's drafts are warmer than Llama's but more concise than Claude's.
The third mistake was at message 134. A customer asked: "Can I return this if I already burned it for an hour?" Auto routing sent it to Llama, which replied: "Our return policy covers unused candles only. If you've burned the candle, we can't accept a return." Technically true, but tone-deaf—the customer was asking, not demanding, and Llama's reply felt curt. I re-routed to Gemini, which softened it: "Our return policy covers unused candles, but if there's an issue with the scent or burn quality, I'm happy to send you a replacement or a different scent instead. Let me know what works best for you." She accepted the replacement. I updated the routing rules to treat any question phrased as "Can I…?" as medium complexity, even if the answer is a simple yes or no.
The final 105 messages: adjusted rules, faster replies
After tweaking the routing rules on December 11th, I processed the remaining 105 messages with almost no manual re-routing. Auto routing sent 29 to Claude, 58 to Gemini, and 18 to Llama. My average time per reply dropped from three minutes (in the first 60 messages) to about ninety seconds. I still read every draft before sending it, but I only made edits to 11 of the 105—mostly adding a sentence or adjusting a word.
The revised ruleset looked like this:
- Claude Sonnet 4.5: any message over 50 words, any message containing emotional language ("disappointed," "frustrated," "upset," "never again"), any message describing damage or a shipping issue, any message that combines a complaint with a question, any message containing "allergy" or "sensitive," any message where the customer's tone suggests they're deciding whether to leave a negative review.
- Gemini Pro: any message asking about exchanges, timelines, or policy details, any message phrased as a question ("Can I…?", "Is it possible to…?"), any message under 20 words that requests a refund, any message that's polite but needs more than a one-sentence reply.
- Llama 3.3 70B: single factual questions with no emotional content, any message asking for the return address or refund timeline, any message that's clearly just confirming a detail.
I also added a catch-all: if I wasn't sure, I defaulted to Gemini. Gemini's drafts were almost always usable, even if Claude might have done slightly better on a few edge cases.
What I'd do differently next Cyber Monday
I'd start with the adjusted ruleset from day one instead of learning it on the fly. The three routing mistakes cost me maybe ten extra minutes total, but they also taught me that Auto routing gets better when you feed it specific triggers rather than asking it to infer complexity from scratch.
I'd also set up a shared inbox in Kryotta so my partner could help with replies during the peak days. Right now, all 165 messages went through my Etsy account, which meant I was the bottleneck. If we'd split the load, we could have processed the same volume in half the time.
One thing I won't change: I'll keep reading every draft before sending it. Auto routing saved me hours, but I caught a handful of small errors—wrong scent names, slightly off tone—that would have been embarrassing to send. The goal isn't to remove myself from the process; it's to handle the volume without burning out or letting reply times slip past 24 hours.
The routing rules I'm keeping for next year
Here's the ruleset I'll start with in November 2025, formatted as a checklist I can paste into Kryotta's Auto routing settings:
| Trigger | Model | Why |
|---|---|---|
| Message >50 words, or contains "disappointed" / "frustrated" / "upset" / "never again" | Claude Sonnet 4.5 | Needs empathy and judgment |
| Describes damage, shipping issue, or time pressure | Claude Sonnet 4.5 | Requires problem-solving and tone management |
| Contains "allergy" / "sensitive" / "can I return just" | Claude Sonnet 4.5 | Needs a judgment call, not just policy |
| Asks about exchange, timeline, or policy | Gemini Pro | Straightforward but needs warmth |
| Phrased as "Can I…?" or "Is it possible to…?" | Gemini Pro | Question implies uncertainty, needs reassurance |
| Under 20 words, requests refund | Gemini Pro | Needs to be concise but not curt |
| Single factual question, no emotional language | Llama 3.3 70B | Fast, accurate, no judgment needed |
| Asks for return address or refund timeline | Llama 3.3 70B | One-sentence reply is fine |
If I'm unsure, I'll route to Gemini and regenerate with Claude if the draft feels too flat.
Questions people ask
Does Auto routing work with Etsy Messages, or do I have to copy-paste?
You copy-paste the message into Kryotta, let Auto routing pick the model and generate a draft, then paste the reply back into Etsy. There's no direct Etsy integration yet. It adds about fifteen seconds per message, but it's still faster than writing from scratch.
What if Auto routing picks the wrong model and I don't notice until after I send the reply?
I caught the three mistakes by skimming the draft before sending. If you do send a reply that feels off, you can follow up with a second message—customers are usually fine with "I realized I didn't fully answer your question, here's a bit more detail." It's not ideal, but it's fixable.
Can I use Auto routing for messages that aren't returns?
Yes. I only tested it on return messages because that's where my volume spiked, but the same routing logic works for pre-sale questions, shipping inquiries, or custom order requests. You'd just adjust the triggers to match the types of questions you get.
How much does Auto routing cost if I'm processing 165 messages in two weeks?
Kryotta's pricing depends on which models you use and how many tokens each reply generates. For my 165 messages, I spent about $4.80 total—most of that on Claude, since it uses more tokens per reply than Gemini or Llama. If you route more messages to Llama, the cost drops.
I'll run the same experiment next Cyber Monday and see if the adjusted ruleset holds up at higher volume. If you're dealing with a post-holiday return surge and want to try Auto routing without building the rules from scratch, you can start with the table above and tweak it as you go. Set up Auto routing in Kryotta and test it on your next ten messages—you'll know within an hour whether it fits your workflow.



