What Auto routing actually does when a message arrives
Auto routing reads your message, checks which models are available, looks at what each model costs and what it's good at, then picks one. You don't choose. It does. I wanted to see if that worked when real customers were asking real questions and I needed answers in under two minutes, not after I'd compared three models myself.
I run a Meesho store selling cotton sarees and unstitched salwar sets. Most orders come through WhatsApp Business. January is wedding season, so I had 150 customer messages in one week: tracking requests, COD confirmations, return questions, bulk-order haggling. I turned on Auto routing in Kryotta and logged which model it picked for each message type, how accurate the reply was, and whether I had to rewrite anything before sending.
The result: Auto routing chose Claude Sonnet 4.5 for 61% of messages, Gemini Flash for 28%, and Llama 3.3 70B for 11%. It picked Claude when the question had nuance (a return request mentioning "colour doesn't match the photo"), Gemini when speed mattered more than perfection (simple tracking updates), and Llama when the message was straightforward and I didn't want to spend ₹0.80 on a one-line reply.
What those 150 messages looked like
Not every customer message needs the same brain. Some are easy: "Where is my order?" Some need judgement: "The blouse piece is missing, I need it before Thursday's function, can you send it separately or should I return everything?"
I grouped the messages into five types and tracked which model Auto routing selected most often for each:
-
Order tracking (42 messages): "Parcel kahan hai?", "Delivered status dikha raha hai but mujhe mila nahi", "Expected date kal tha, update do." Auto routing picked Gemini Flash 38 times, Claude 4 times. Gemini answered faster and the replies were fine—order ID, courier name, expected date. Claude only got picked when the message included frustration or a threat to cancel.
-
COD confirmations (31 messages): Customer places order, I need to confirm COD is available in their pincode and give the total with shipping. Auto routing chose Claude 22 times, Gemini 9 times. Claude's replies sounded more human ("Yes, COD available in 560078, total ₹847 including ₹47 shipping, delivery by Friday") and matched my tone. Gemini's were correct but flat.
-
Return and exchange requests (28 messages): "Size small hai, can I exchange?", "Colour photo se alag hai, refund milega?", "Blouse piece missing." Auto routing went Claude 26 times, Llama twice. Returns need empathy and policy knowledge. Claude nailed both. Llama's two replies were technically accurate but sounded like a FAQ bot, so I edited them before sending.
-
Bulk order questions (19 messages): Boutique owners asking for twenty sarees in mixed colours, or a wedding planner wanting fifteen salwar sets with custom blouse stitching. Auto routing picked Claude every single time. These messages need negotiation, not templates. Claude understood when to offer a discount, when to ask clarifying questions, and when to say no politely.
-
Random questions (30 messages): "Do you have this in green?", "What's the fabric weight?", "Can I pay half now half later?" Auto routing spread these across all three: Claude 9, Gemini 12, Llama 9. Simple factual questions went to Gemini or Llama. Anything that touched payment terms or required me to make an exception went to Claude.
Why Auto routing chose Claude for the hard stuff
Claude Sonnet 4.5 costs more per message (around ₹0.80 for a typical reply versus ₹0.15 for Gemini Flash), but Auto routing still picked it 61% of the time. That felt expensive until I looked at which messages got Claude.
Every return request, every bulk-order negotiation, every message where the customer was upset or asking for something outside standard policy—Auto routing sent those to Claude. And Claude's replies needed almost no editing. It understood context. When someone said "Colour photo se alag hai", Claude didn't just say "return accepted"; it acknowledged the disappointment, explained the return process, and offered to send photos of other colours in stock. That saved me a follow-up message and kept the customer from canceling.
Gemini Flash is fast and cheap, but it doesn't read between the lines. A message that says "Parcel kal aana chahiye tha, function Friday ko hai, ab kya karoon?" is really asking "Can you fix this or should I panic?" Gemini answered the literal question (tracking update). Claude answered the real one (offered express shipping upgrade or a partial refund if it didn't arrive in time).
When Gemini Flash was the right choice
Gemini handled 28% of messages, almost all of them tracking updates and stock checks. These are high-volume, low-stakes questions where speed beats nuance. Customer asks where the parcel is, I need to paste the tracking link and expected date in under sixty seconds, done.
Gemini's replies were shorter than Claude's, which is exactly what I wanted for these messages. "Your order #ME2847 is with Delhivery, tracking number 5738294, expected delivery 24 Jan" is perfect. I don't need three sentences of reassurance when the answer is just data.
The cost difference adds up here. Forty-two tracking messages at ₹0.15 each (Gemini) versus ₹0.80 each (Claude) is ₹6.30 versus ₹33.60. Auto routing saved me ₹27.30 on tracking messages alone by not using Claude when Gemini was good enough.
The eleven times Auto routing picked Llama
Llama 3.3 70B is the cheapest option in Kryotta—around ₹0.05 per reply—but Auto routing only chose it for 11% of messages. These were the simplest questions: "Do you have this in size L?", "What's the return period?", "Is this pure cotton?"
Llama answered correctly every time, but the tone was flat. "Yes, size L is available" versus Claude's "Yes, I have size L in stock, I can send you photos if you'd like to confirm the colour." Both answers are true, but one makes the customer more likely to buy.
I wouldn't use Llama for anything customer-facing that needs warmth, but for internal notes or quick checks it's fine. Auto routing understood that and only picked Llama when the question was so straightforward that tone didn't matter.
What I edited before sending
Auto routing got the model choice right 89% of the time (I sent 133 replies as-written, edited 17). The edits were small: adding a customer's name, tweaking a discount offer, softening a "no" into a "not right now, but here's what I can do."
The seventeen I edited broke down like this:
-
Nine Claude replies: Too formal. Claude sometimes wrote "We regret to inform you" when I would've said "Sorry, that size is out of stock right now." I kept Claude's structure but made the language warmer.
-
Six Gemini replies: Missing context. Gemini answered the literal question but didn't acknowledge the urgency. A message that said "Function Friday ko hai" got a standard tracking update with no mention of expedited shipping.
-
Two Llama replies: Correct but robotic. I added one sentence of empathy before the factual answer.
Auto routing's model choice was right in all seventeen cases—I didn't wish it had picked a different model, I just adjusted the tone. That's a much smaller problem than picking the wrong model and getting a useless answer.
What this cost versus ChatGPT Plus
ChatGPT Plus costs ₹1,650/month. For that I get GPT-4 access, which is good, but I can't switch to a cheaper model when I don't need the full power. Every message costs the same whether it's a complex return negotiation or a one-line stock check.
Kryotta's Auto routing cost me ₹47 for those 150 messages. Breakdown:
- 92 Claude replies at ₹0.80 each: ₹73.60
- 42 Gemini replies at ₹0.15 each: ₹6.30
- 16 Llama replies at ₹0.05 each: ₹0.80
Total: ₹80.70. But Kryotta credits the first ₹500 free each month, so my actual cost was zero. Even if I'd paid full price, ₹80.70 for 150 messages is ₹0.54 per message. ChatGPT Plus would've been ₹1,650 for unlimited messages, but I don't send 3,000 messages a month. I send 150–200. Paying ₹1,650 for that makes no sense.
The math works if you're a solo seller handling customer messages yourself. If you're running a team and everyone needs access, Kryotta's team workspace pricing (₹299/user/month) is still cheaper than giving everyone ChatGPT Plus.
The one thing Auto routing got wrong
Twice, Auto routing picked Gemini for messages that needed Claude. Both were return requests where the customer mentioned a wedding or function. Gemini gave the correct return policy but didn't acknowledge the time pressure. I caught both before sending and switched to Claude manually.
This happened because the customer didn't explicitly say "urgent" or "need it by Friday"—they just mentioned the event in passing. Auto routing read it as a standard return question. Claude would've caught the subtext. Gemini didn't.
I reported both cases to Kryotta's feedback tool (there's a thumbs-down button on every reply). No idea if that improves Auto routing over time, but it felt worth logging.
Questions people ask
Does Auto routing work if my customers write in Hindi or Hinglish?
Yes. All three models (Claude, Gemini, Llama) handled Hinglish without breaking. Messages like "Parcel kab aayega?" or "Size chhota hai, exchange ho sakta hai?" got accurate replies. Claude and Gemini were better at matching the customer's language—if they wrote in Hinglish, the reply came back in Hinglish. Llama sometimes replied in full English even when the question was mixed.
Can I force Auto routing to always pick the cheapest model?
Not directly, but you can turn off the expensive models in your workspace settings. If you disable Claude, Auto routing will only choose between Gemini and Llama. I tested this for one day and saved ₹12, but I had to rewrite six replies that needed Claude's judgement. Not worth it.
What happens if Auto routing picks wrong and I don't notice until after sending?
You fix it in the follow-up. I sent one Gemini reply that was too blunt (customer asked about a missing blouse piece, Gemini said "Contact courier"), customer replied annoyed, I switched to Claude for the next message and smoothed it over. Auto routing doesn't lock you into anything.
Is this faster than just picking a model myself?
Yes, by a lot. Picking a model manually adds fifteen seconds per message—I have to read the question, decide if it's simple or complex, remember which model is good at what, then click. Auto routing happens instantly. Over 150 messages that's 37 minutes saved.
Auto routing isn't perfect, but it's right often enough that I trust it for first drafts. I still read every reply before sending (I'm not handing my customer relationships to a black box), but I'm editing tone, not rewriting from scratch. That's the difference between spending two hours on messages and spending thirty minutes. If you're managing a Meesho store or WhatsApp Business orders during peak season and you're still typing every reply yourself, try Auto routing for a week and log which model it picks. You'll see the same pattern I did: it spends money on the messages that need it and saves everywhere else. Start at Kryotta and turn on Auto routing in workspace settings—it's off by default, which I didn't realize for the first three days and wondered why nothing was happening.
