I started with 140 GST queries in eight weeks and no idea which model could handle tax language without sounding like a notice
I sell handblock-printed cotton dupattas and kurta sets on Shopify. My store is in Pune, my manufacturer is in Jaipur, and my customers are everywhere—Bangalore techies buying gifts, wedding planners in Surat ordering in bulk, a hotel chain in Goa that wanted custom prints for their staff uniforms. I've been doing this for three years. I know my fabrics, I know my margins, and I thought I knew GST well enough to stay compliant.
Then Diwali happened. Between mid-October and the end of December I received 140 GST-related queries and disputes. Some came through email, some through WhatsApp Business, a few through Shopify's contact form. Buyers who'd entered the wrong GSTIN and wanted a revised invoice. Flipkart asking why my place-of-supply didn't match the shipping address on three orders. The GST portal flagging four invoices for reverse-charge errors I didn't even know I'd made. A CA firm in Chennai that bought ₹1.8 lakh worth of fabric for a corporate event and then disputed the HSN code because their accountant said it should be 5208, not 5407.
Every reply took me seven to ten minutes. I'd open the invoice, check the portal, draft an answer, rewrite it to sound less annoyed, check again that I hadn't promised something wrong, and hit send. By mid-November I was spending ninety minutes a day just on tax emails. I needed help, but hiring someone meant explaining my entire setup, and I wasn't sure a VA would get the technical bits right.
I'd been using Kryotta's Compare feature to test different models for product descriptions. I thought: what if I feed these queries to Claude, Gemini, Llama, and GPT-OSS and see which one writes replies that sound human, stay compliant, and don't take three follow-ups to get right? I pulled twenty queries from each category—GSTIN errors, place-of-supply disputes, reverse-charge confusion, amended return requests, and general "why is my invoice like this" questions—and started testing.
Claude Sonnet 4.5 handled reverse-charge and amended return emails better than I could
Reverse-charge queries scared me. A buyer in Haryana ordered ₹45,000 worth of fabric, then emailed saying his accountant told him I should've charged tax under reverse charge because he's registered and I'm the supplier. I knew the threshold is ₹5,000 per day or ₹25 lakh per year for the recipient, but I didn't know how to explain that without sounding like I was reading from a government PDF.
I gave Claude this prompt:
"Draft a reply to a B2B buyer who thinks reverse charge applies to his ₹45,000 textile order. Explain when reverse charge applies, confirm it doesn't apply here, stay polite, and keep it under 120 words."
Claude came back with a reply that cited Section 9(4) of the CGST Act, explained the ₹5,000 daily threshold, confirmed that regular GST applied to his order, and suggested he check with his CA if he had questions. It was firm, polite, and accurate. I sent it with one edit—I added "Hope this clears it up" at the end because the original felt slightly stiff.
Amended return requests were worse. A bulk buyer in Indore wanted me to revise an August invoice because his GSTIN had a typo and he couldn't claim input credit. I needed to explain that I'd file a GSTR-1 amendment in the next return period, but I wasn't sure how to word it so he'd know when to expect the revised invoice on the portal.
Claude's draft explained the amendment process, gave him a timeline (next month's GSTR-1 filing, so mid-October), and told him he could track it on the portal once I uploaded it. It even added a line saying I'd send him a confirmation email when it was done. I used that reply verbatim for six more amendment requests. Average time per email dropped from eight minutes to ninety seconds.
Gemini Flash handled GSTIN format errors and place-of-supply mismatches faster and cheaper
GSTIN errors were the biggest category—68 out of 140 queries. Someone typed 28 instead of 29 in the state code. Someone left out a digit. Someone sent me a TAN instead of a GSTIN and got annoyed when the invoice didn't generate. These emails didn't need deep tax knowledge; they needed patience and a clear explanation of what went wrong.
I tested Gemini Flash because it's fast and cheap. My prompt:
"Buyer entered GSTIN 27AABCU9603R1Z9 but it should be 27AABCU9603R1Z5. Draft a reply explaining the error, asking them to confirm the correct GSTIN, and offering to reissue the invoice once I have it. Keep it friendly and under 100 words."
Gemini came back in two seconds with a reply that pointed out the mismatch, asked for confirmation, and said I'd send the revised invoice within 24 hours. It didn't over-explain. It didn't sound like a tax notice. I used it for 52 GSTIN queries and tweaked maybe eight of them—usually just to add the customer's name or reference their order number.
Place-of-supply disputes were trickier. A buyer in Guwahati ordered fabric, shipped it to his warehouse in Siliguri, and then complained that the invoice showed Assam as the place of supply when it should've been West Bengal. I knew the rule—place of supply is where the goods are delivered—but I'd used his billing address by mistake because Shopify auto-filled it.
Gemini's reply explained the error, confirmed I'd reissue the invoice with West Bengal as the place of supply, and apologised for the confusion. It took three seconds to generate and didn't need any edits. I used the same structure for eleven more place-of-supply queries, changing only the state names and order details.
Llama 3.3 70B was better for "why is my invoice like this" questions when I didn't want a lecture
Some queries weren't disputes—they were just confused customers. "Why is there CGST and SGST instead of IGST?" "Why is the taxable value less than the total I paid?" "What's this cess line for?" These people weren't angry; they just wanted to understand their invoice without feeling stupid.
I tried Claude first. It gave me a technically perfect explanation of intra-state vs inter-state supply, broke down the tax components, and cited the relevant sections. It was 180 words. Too much.
I switched to Llama 3.3 70B with this prompt:
"Buyer in Pune (same state as me) asks why there's CGST and SGST on the invoice instead of one tax. Explain it simply, keep it under 80 words, don't sound like a textbook."
Llama's reply: "Since we're both in Maharashtra, GST splits into two parts—Central GST (CGST) and State GST (SGST), each 9%. If you were in another state, you'd see IGST instead (18% in one line). It's just how the tax gets divided between the Centre and the state. The total you pay stays the same. Let me know if anything else looks off!"
Perfect. I used Llama for 34 general explainer queries. It was conversational, accurate, and short. Twice it got a detail slightly wrong—once it said the cess applied to all textiles when it only applies to some categories—but I caught it before sending and added a line to clarify.
GPT-OSS 120B was the best fallback when I needed a reply in ten seconds and didn't want to compare four outputs
By December I had a workflow: Claude for technical disputes, Gemini for format errors, Llama for explainers. But some emails didn't fit a category. A buyer in Kolkata sent a rambling WhatsApp message asking about three different invoices, one of which had a reverse-charge question, one had a place-of-supply issue, and one just needed a duplicate PDF. I didn't have time to split it into three prompts and compare models.
I used GPT-OSS 120B. One prompt with all three questions pasted in, asked it to draft replies for each, kept the whole thing under 200 words. It came back with three short paragraphs—one explaining reverse charge didn't apply, one confirming I'd reissue the place-of-supply invoice, one saying I'd email the duplicate PDF within an hour. I sent it as-is. The buyer replied "Thanks, sorted" and didn't follow up.
I used GPT-OSS for eighteen mixed or urgent queries when I just needed something decent in under ten seconds. It wasn't as precise as Claude or as cheap as Gemini, but it was reliable and I never had to rewrite a reply from scratch.
I cut my average response time from eight minutes to under two and didn't get a single follow-up complaint
Before I started testing models, 140 queries over eight weeks meant roughly 18 hours of email work. After I built my workflow—Claude for technical, Gemini for format, Llama for explainers, GPT-OSS for mixed—I spent about four hours total. I drafted most replies in 60 to 90 seconds, spent another 30 seconds checking the invoice to make sure the model hadn't invented a detail, and sent them.
I tracked follow-ups. Out of 140 replies, I got nine follow-up questions. Six were people asking for the revised invoice (which I'd already said I'd send, they just wanted to confirm). Two were asking for clarification on a tax term (I sent a one-line reply). One was a genuine error—I'd used Llama for a reverse-charge query and it oversimplified the threshold rule, so the buyer came back confused. I switched to Claude for the follow-up and it cleared it up.
No one complained that the emails sounded like AI. No one said I'd given them wrong information that messed up their taxes. My CA didn't get any panicked calls from buyers saying I'd told them something non-compliant. The replies worked.
I still write some emails myself—anything involving a refund or a genuinely angry customer gets my voice, not a model's. But for the 80% of queries that are just "fix this invoice" or "explain this line," I'm not spending eight minutes per reply anymore.
The models I'd use again and the ones I wouldn't for GST emails
Claude Sonnet 4.5 is the one I'd pay for if I could only pick one. It handles technical tax language, doesn't guess when it's unsure, and writes replies that sound like a patient human who knows the rules. I used it for 31 queries and edited maybe five of them. It's slower and costs more than Gemini, but for anything involving reverse charge, amendments, or compliance risk, I'm not using anything else.
Gemini Flash is the best value for high-volume, low-risk queries. GSTIN errors, place-of-supply fixes, duplicate invoice requests—anything where the answer is straightforward and you just need it written politely. I used it for 63 queries and it cost me almost nothing. If you're handling 200+ tax emails a month and half of them are format errors, Gemini will save you hours.
Llama 3.3 70B is underrated for explainer emails. It doesn't try to sound authoritative; it just explains the thing in plain language and gets out of the way. I used it for 34 queries and only had one follow-up where I had to clarify. If your customers aren't disputing anything and just want to understand their invoice, Llama's the one.
GPT-OSS 120B is the model I reach for when I don't have time to think about which model to use. It's not the best at anything, but it's good enough at everything, and it's fast. I used it for 18 mixed queries and didn't regret it once. If you're drowning in emails and just need replies that won't make things worse, GPT-OSS will get you through the day.
I'm still using Kryotta's Compare feature when I get a query type I haven't seen before—last week someone asked about the GST treatment of gift hampers, and I tested all four models before picking Claude's answer. But for the categories I've already mapped, I know which model to use, and I'm not spending my evenings writing tax emails anymore.
Questions people ask
Which model is most accurate for GST compliance questions?
Claude Sonnet 4.5. It cites sections, explains thresholds correctly, and doesn't oversimplify rules that have exceptions. I used it for every reverse-charge and amendment query and didn't get a single complaint.
Can I use these models if I'm not sure about the tax rule myself?
Not safely. The models draft replies based on what you tell them. If you're unsure whether reverse charge applies or which HSN code is correct, ask your CA first, then use the model to write the explanation. Don't let the model make compliance decisions for you.
Do the replies sound like AI?
Not if you edit the greeting and sign-off. I add the customer's name, swap "Best regards" for "Thanks," and sometimes add a line like "Let me know if you need anything else." The body of the email—the explanation, the steps, the timeline—sounds professional and human. No one's ever asked if I used AI.
How much did this cost in tokens?
I didn't track it precisely, but 140 queries over eight weeks, most of them under 150 words, cost me less than ₹400 in Kryotta credits. Gemini Flash was cheapest, Claude was the most expensive, but even Claude didn't cost more than ₹2 per reply. Compared to eighteen hours of my time, it's not even a question.
If you're spending hours every week on GST emails, invoice disputes, or tax explainers, you don't need to hire someone or write every reply from scratch. You need a workflow, a few good prompts, and models that know when to cite a rule and when to just explain it simply. I built mine in Kryotta because I wanted Claude, Gemini, Llama, and GPT-OSS in one place without switching tabs or managing four subscriptions. It worked. I'm not going back.
