I ran the test the week after Cyber Monday, when the inbox was already a mess
Two hundred and six return requests came in between November 30 and December 6. I run a small Shopify store selling kitchen gadgets—silicone baking mats, bench scrapers, that sort of thing—and I'd pushed hard on Black Friday bundles. The sale went well. The returns started the moment people opened the boxes and realised they'd bought three spatula sets they didn't need.
I'd connected Kryotta's Shopify connector the week before, mostly to see if Auto routing could handle the volume without me writing a different prompt for every request type. Auto routing is the feature that looks at your query, figures out what kind of task it is, and picks the fastest capable model instead of making you choose. I wanted to know if it actually worked, and more importantly, which model it would send each kind of return to. So I tagged every request as it came in—refund, exchange, damaged item, policy question—and tracked what the system chose.
Here's what I learned, broken down by the query types you'll see in your own post-holiday inbox.
Straightforward refunds: Gemini Flash Lite handled most of them
A straightforward refund is someone saying "I want my money back" with an order number and no drama. No damaged goods, no missing items, just buyer's remorse or a duplicate purchase. Out of two hundred and six requests, ninety-one were this type. Auto routing sent eighty-seven of them to Gemini Flash Lite.
The logic makes sense. These requests need speed more than nuance. Flash Lite confirmed the order number, checked that it was within my thirty-day return window, generated the return label, and drafted a reply with the refund timeline—usually in under three seconds. The four exceptions went to Gemini Pro, and when I looked at those, they all had slightly ambiguous phrasing: "I think I want to return this but I'm not sure if I opened it" or "Can I get a refund even though I threw away the packaging?" Pro handled the edge cases where the policy needed interpreting.
If you're setting this up for your own store, the takeaway is that you don't need a powerful model for the bulk of your refund queue. Flash Lite is fast and cheap, and it'll process the simple ones while you focus on the weird ones.
Exchanges: Claude Sonnet picked up anything involving inventory checks
Forty-three requests were exchanges—someone wanted a different size, a different colour, or a different product entirely. Auto routing sent thirty-nine of them to Claude Sonnet 4.5.
I didn't expect this at first, but when I looked at the queries, they all had one thing in common: they required the model to cross-reference what the customer wanted with what I actually had in stock. "Can I swap the red spatula set for the blue one?" means the model needs to check my Shopify inventory, see if the blue set is available, confirm the price difference (if any), and explain the exchange process. Claude Sonnet is better at that kind of structured lookup than the faster models.
The four that didn't go to Sonnet went to Gemini Pro, and those were all cases where the customer wanted to exchange for something that wasn't a direct swap—"I bought the small baking mat but I actually need the large one, can I just pay the difference?" Pro handled the partial refund math and the new order creation in one pass.
If your Shopify store has a lot of SKU variants, expect Claude Sonnet to do the heavy lifting on exchanges. It's slower than Flash Lite, but it doesn't make inventory mistakes.
Damaged goods and missing items: DeepSeek V3 showed up for anything requiring evidence review
Thirty-two requests involved damaged products or missing items. These are the ones that make you nervous, because you need to look at photos, decide if the damage happened in shipping or after delivery, and figure out whether to issue a refund or ask the customer to file a claim with USPS. Auto routing sent twenty-nine of them to DeepSeek V3.
This surprised me. I'd assumed Claude Sonnet would handle anything that needed judgement, but DeepSeek V3 turned out to be better at parsing image uploads and matching them against my product photos. A customer sent a picture of a cracked bench scraper and wrote "This arrived broken." DeepSeek pulled the original product image from my Shopify listing, compared the crack location to the packaging seams, and flagged it as likely shipping damage. It drafted a reply offering a full refund and a prepaid return label, and it did it in about five seconds.
The three that went to Claude Sonnet were all cases where the customer's description was vague—"Something's wrong with this but I can't explain it" with a blurry photo. Sonnet asked follow-up questions to narrow down the issue before committing to a refund. That's the trade-off: DeepSeek is faster when the evidence is clear, Claude is better when you need to investigate.
If you're dealing with physical products that can arrive damaged, expect DeepSeek to handle the straightforward cases and Claude to pick up the ambiguous ones. Neither model will approve a fraudulent claim, but DeepSeek is more willing to trust a clear photo.
Policy questions: Llama 3.3 70B took the ones that needed your return policy quoted back
Twenty-two requests were pure policy questions—no refund or exchange yet, just someone asking if they could return something. "I bought this three weeks ago, is it too late?" or "Do I have to pay return shipping?" Auto routing sent eighteen of them to Llama 3.3 70B.
Llama's strength here is that it's very good at pulling exact phrases from your return policy and restating them in plain English without editorialising. A customer asked "Can I return something I used once?" and Llama quoted the relevant line from my policy: "Items must be unused and in original packaging," then explained that "used once" would disqualify the item. It didn't try to be generous or make an exception—it just told the customer what the rule was.
The four that went to Gemini Pro were all cases where the question had a hypothetical attached: "If I return this, can I use the refund toward a different purchase, or does it have to go back on my card?" Pro handled the conditional logic better than Llama.
If you get a lot of "Can I…?" questions before people commit to a return, Llama will save you time by answering them consistently. It won't bend your policy, which is either a feature or a limitation depending on how you run your store.
The eighteen requests that didn't fit a pattern: Claude Opus handled the edge cases
Eighteen requests were weird. A customer who wanted to return one item from a bundle but keep the rest. Someone who'd received the right product but claimed it didn't match the description on my site. A person who wanted a refund and to keep the item because "it's not worth the hassle of shipping it back." Auto routing sent sixteen of these to Claude Opus 4.5.
Opus is the model you use when the query has multiple moving parts and no obvious answer. It's slower and more expensive than the others, but it doesn't panic when the request is unusual. The bundle return required Opus to calculate a prorated refund based on the individual item prices, check if the remaining items still qualified for the bundle discount, and explain the math to the customer. It took about twelve seconds, but the answer was correct and the customer accepted it.
The two that didn't go to Opus went to Claude Sonnet, and both were cases where the edge case turned out to be simpler than it looked. "I want to return this but I lost the order confirmation email" sounds complicated, but Sonnet just looked up the order by the customer's email address and moved on.
If you're worried about Auto routing sending everything to the most expensive model, don't be. Opus only showed up when the query genuinely needed it, and even then it was less than nine percent of the total volume.
What Auto routing didn't do well: anything requiring a human decision disguised as a policy question
There were four requests where Auto routing picked a model, the model gave an answer, and I had to override it because the answer was technically correct but commercially stupid. One customer had ordered a $12 silicone mat, received it on time, used it once, and then asked for a refund because "it's not as non-stick as I expected." My return policy says used items aren't eligible for return. Llama quoted that policy. The customer replied with a long explanation of how they'd seen my Instagram post claiming the mats were "completely non-stick" and felt misled.
At that point, the right move was to issue the refund anyway and eat the $12 to avoid a chargeback or a bad review. Auto routing can't make that call. It can tell you what your policy says, but it can't weigh the cost of enforcing the policy against the cost of bending it. I ended up refunding three of these four cases manually.
This isn't a flaw in Auto routing—it's just the boundary of what automation can do. If a return request has a commercial judgement buried inside it, you still need to make that judgement yourself.
The setup that made this work: Shopify connector, Auto routing turned on, and one custom instruction
I'm not particularly technical. The Kryotta Shopify connector took about ten minutes to set up—you paste your Shopify API key, grant access to orders and products, and turn on Auto routing in the workspace settings. The one custom instruction I added was: "When a return request mentions a damaged item, always ask for a photo before approving the refund." That saved me from approving four claims that turned out to be buyer's remorse disguised as damage reports.
Auto routing doesn't require you to write a different prompt for every query type. It looks at the content of the request, figures out what kind of task it is, and picks the model that's fast enough and capable enough to handle it. You can override the choice if you want, but in two hundred and six requests I only did that twice—both times because I wanted to see if a slower model would give a different answer. It didn't.
Questions people ask
Does Auto routing always pick the cheapest model that can do the job?
Mostly, yes. It optimises for speed and cost unless the query has complexity that requires a more capable model. In my test, Gemini Flash Lite handled the highest volume because most refund requests are straightforward.
Can I force Auto routing to always use Claude Sonnet for everything?
You can turn off Auto routing and manually select a model for each query, but that defeats the purpose. If you don't trust the routing logic, you're better off using a single model in a standard chat and skipping the connector.
What happens if Auto routing picks the wrong model and gives a bad answer?
You can regenerate the response with a different model, or just edit the draft before you send it. Auto routing isn't autopilot—it's a first draft generator that saves you from starting every reply from scratch.
Does this work outside Black Friday, or is it only useful for return spikes?
It works year-round. I used it for returns because that's when the volume justified the test, but the same routing logic applies to pre-sale questions, order status requests, and policy inquiries any time of year.
I'm still using Auto routing for every return request that comes in. It's not perfect, but it's faster than writing two hundred replies by hand, and it picks the right model more often than I would if I had to choose manually. If you're running a Shopify store and you're dreading the post-holiday return flood, try the connector and see which model it picks for your first ten requests. You'll know in an hour whether it's worth keeping.
