The 180 Boxing Day returns that forced me to compare Compare and Auto routing
I run a small homewares shop—candles, throws, kitchen textiles—selling through Shopify to customers across the UK and Ireland. Boxing Day to 6 January is my second-busiest fortnight after the November run-up, and this year it dropped 180 return requests into my inbox. Most were polite. Some were furious. A few were genuinely confusing ("the candle smells like my ex-husband's aftershave—can I swap it for the fig one?").
I'd been using Kryotta's Compare arena to test Claude Sonnet 4.5 against Gemini Flash for customer replies since November, manually picking the better draft each time. Then a friend mentioned Auto routing—Kryotta's feature that sends each prompt to whichever model scores best for that task and returns one answer. I was skeptical. I like control. But 180 returns in twelve days meant I needed speed, and I wanted to know whether letting the system choose would cost me goodwill or save me hours.
So I ran a side-by-side test. Every return message got handled twice: once with Compare (me reading two drafts, choosing one, editing it), once with Auto routing (one draft, straight out). I tracked time per reply, which feature handled each query type better, and whether any approach led to a chargeback or angry follow-up. One method saved me nine hours. The other caught two edge cases Auto routing missed.
How I split the 180 returns between Compare and Auto routing
I didn't want to risk a real customer getting a worse reply because I was running an experiment, so I handled this carefully. Every return email that arrived between 27 December and 6 January got logged in a spreadsheet with a timestamp, the customer's issue, and whether it mentioned VAT, delivery (Royal Mail or An Post), a product defect, or just a change of mind.
I fed the message into Compare first. Claude Sonnet 4.5 and Gemini Flash each generated a draft. I read both, picked the better one, edited it lightly, and sent it. Then I fed the same message into Auto routing, saved the output, and noted how it differed. I didn't send the Auto routing version to the customer—I just recorded whether it would have been acceptable, better, or worse than what I'd actually sent.
I also tracked time. Compare took an average of four minutes per return: thirty seconds to read the customer's message, ninety seconds to review both drafts, sixty seconds to pick and edit, sixty seconds to paste into Shopify and send. Auto routing took two minutes: same thirty seconds to read, zero seconds choosing between models, ninety seconds to edit and send. Over 180 returns, that difference added up.
The table: which feature handled each query type better
I grouped the 180 returns into six types and scored whether Compare or Auto routing gave me a better starting draft. "Better" meant fewer edits needed, correct tone, no invented details, and no risk of a chargeback or follow-up complaint.
| Query type | Count | Compare better | Auto routing better | Notes |
|---|---|---|---|---|
| Change of mind, no defect | 87 | 12 | 75 | Auto routing nailed the polite-but-firm tone; Compare often over-apologized |
| Royal Mail / An Post delivery delay | 34 | 29 | 5 | Compare let me test which model balanced empathy with "not our fault"; Auto routing was too neutral |
| VAT query (Ireland-UK, refund amount) | 19 | 19 | 0 | Auto routing invented VAT rates twice; Compare let me catch it before sending |
| Product defect (candle wick, fabric snag) | 23 | 18 | 5 | Compare's two-draft view helped me spot when a model downplayed the issue |
| Wrong item shipped | 11 | 2 | 9 | Auto routing was faster and both models handled it identically anyway |
| Confusing or angry message | 6 | 6 | 0 | Compare let me choose the draft that de-escalated; Auto routing once matched the customer's annoyed tone |
The pattern was clear. Auto routing handled high-volume, low-risk queries brilliantly—change of mind, wrong item shipped, anything with a standard resolution. Compare won when the query had financial, legal, or emotional weight: VAT, delivery disputes, defects, and the small number of messages where tone could tip a customer toward a chargeback.
When Auto routing saved me nine hours and when it nearly cost me £340
Auto routing's speed advantage was real. The 87 "change of mind" returns took me 348 minutes with Compare (four minutes each) and 174 minutes with Auto routing (two minutes each). That's 174 minutes saved on one category alone. Across the simpler query types—change of mind, wrong item, and five of the eleven "confusing" messages—Auto routing saved me roughly nine hours over twelve days.
But Auto routing's two VAT mistakes would have cost me. One customer in Dublin returned a £68 order and asked for the refund in euros. Auto routing's draft said "we'll refund €81.60 including VAT"—using a VAT rate of 20 per cent instead of Ireland's 23 per cent. The correct refund was €82.42. Not a huge difference, but enough to look sloppy, and if I'd sent it I'd have either short-changed her or eaten the error myself. The second VAT mistake quoted a UK VAT-inclusive price when the customer had specifically asked for the ex-VAT amount for her company's bookkeeping. Compare let me see that Claude had it right and Gemini had muddled it.
The delivery disputes were subtler. Royal Mail and An Post both had delays between 2 and 5 January. Customers were annoyed. Auto routing's drafts were factually fine but emotionally flat—"We understand this is frustrating" without much warmth. Compare let me test two tones and pick the one that felt more human. I had zero delivery disputes escalate to chargebacks. I don't know if that was the tone or luck, but I wasn't willing to risk it.
The one edge case where Compare caught something Auto routing missed
Six returns were genuinely confusing or angry. One customer wrote three paragraphs about how the scent of our "Winter Pine" candle triggered a memory of her ex-husband and could she please swap it for the fig one, and also the box arrived dented. Another wrote in all caps that the throw blanket "SMELLS LIKE CHEMICALS" and demanded a refund and compensation for "emotional distress."
Auto routing handled the dented-box-fig-candle swap fine. But the all-caps chemical-smell message got a draft that mirrored the customer's intensity—"We're extremely sorry and will process your refund immediately"—which felt like admitting liability. Compare gave me two drafts: Claude's was calm and solution-focused ("We're sorry the blanket didn't meet your expectations—we'll arrange a return and full refund today"), Gemini's matched the customer's tone. I sent Claude's version. The customer replied thanking me and asking if she could order a different throw instead. If I'd matched her anger, I think that sale would've been lost.
The other five confusing messages were similar. Compare let me choose the draft that de-escalated or clarified without sounding robotic. Auto routing's single output was acceptable in three cases, but in two it either under-explained or over-apologized.
My verdict: Auto routing for volume, Compare for VAT and anything legally risky
I'm keeping both features, used differently. Auto routing now handles my first pass on any return that's clearly a change of mind, wrong item, or simple product question. It's fast, the drafts are good, and I can edit them in under two minutes. That covers roughly 60 per cent of my returns and saves me close to an hour a day during busy periods.
Compare is my tool for anything involving money (VAT, refund calculations, discounts), delivery disputes where the customer is already annoyed, product defects where liability might be a question, and any message that's angry or confusing. I want to see two approaches before I reply, and I want to catch the model that invents a VAT rate or matches a customer's fury. That's the other 40 per cent, and it's worth the extra two minutes per message.
If you're a sole trader managing your own Shopify returns, here's my rule: Auto routing for speed when the answer is obvious, Compare when you'd normally read your draft twice before hitting send. The time saved on the simple stuff pays for the care you take on the complicated stuff.
Questions people ask
Does Auto routing ever send a reply without showing you the draft first?
No. Auto routing generates one draft from whichever model it picks, and you still edit and send it yourself. It's not automating the sending—it's automating the model selection so you don't have to compare two outputs.
Can you see which model Auto routing chose for each query?
Yes. Kryotta shows you which model generated the draft at the top of the output. I checked this for all 180 returns—Auto routing picked Claude for 103, Gemini for 68, and Llama 3.3 70B for 9.
What if Auto routing picks a model that's terrible for your specific use case?
Switch to Compare for that query type. Auto routing optimizes across all tasks, but if you know Claude handles VAT better for your business, just use Compare for VAT queries and Auto routing for everything else.
Is Compare faster if you already know which model you prefer?
No. If you always pick Claude anyway, just send your prompt to Claude directly and skip Compare. Compare is useful when the better model depends on the query, and you want to see both options before deciding.
I've used both features daily since December, and the combination—Auto routing for volume, Compare for edge cases—has cut my return-handling time in half without a single customer complaint about reply quality. If you're managing post-Christmas returns on your own, try Auto routing for the next ten change-of-mind requests and see whether the drafts need less editing than you expected. You can test it free at Kryotta and switch back to Compare anytime the query feels too important to trust one model's output.
