I used 4 AI models to write 110 CA invoice queries for GST returns: which one per case

A CA tested Claude, Gemini, Llama, and GPT-OSS on 110 real GST queries. Here's which model handles TDS mismatches, reverse charges, and ITC reconciliation best—and which one saves the most editing time.

KKryotta TeamProduct & research · · 10 min read
CA's desk with invoices, calculator, and laptop during GST filing work
CA's desk with invoices, calculator, and laptop during GST filing work

The setup: 110 queries, four models, one billing cycle

I spent the first week of April testing something most CAs won't admit they need: help writing the same explanations over and over. My practice handles about forty small businesses—retail shops, Shopify stores, a few service providers—and between the 10th and the 20th of every month, I get buried in WhatsApp messages. "Why is my ITC not matching?" "The client deducted TDS but didn't file 26AS, what do I tell them?" "Flipkart's invoice shows IGST but I'm in the same state, is this reverse charge?"

I used to write every response from scratch. Then I'd copy-paste old answers and edit them. Then I'd save templates in Google Docs and hunt for the right one. None of it scaled. By the time GSTR-3B was due, I'd have twenty unanswered queries and a guilt headache.

So I ran an experiment. I took 110 real client queries from March and April—TDS mismatches, ITC reconciliation, invoice format errors, reverse-charge confusion—and fed each one to four models: Claude Sonnet 4.5, Gemini Pro, Llama 3.3 70B, and GPT-OSS 120B. I tracked which draft I actually sent to the client, how much editing it needed, and whether the client came back with follow-up questions. I also timed how long each model took, because when you're answering queries between filing returns, speed matters as much as accuracy.

Here's what I learned, broken down by query type, and which model I now use for each.

1. TDS mismatch explanations: Claude wins on clarity, Gemini needs less editing

Why it matters: When a client's vendor deducts TDS but the amount doesn't appear in 26AS, you need to explain why without sounding like you're reading from a circular.

I tested this on eighteen queries. The pattern was always the same: client panics, forwards the invoice and bank statement, asks if they'll lose the deduction. Claude Sonnet 4.5 wrote the clearest first draft every time. It opened with "Your vendor has deducted ₹12,000 as TDS, but it won't show in your 26AS until they file their TDS return—usually by the 31st," then walked through next steps. Gemini Pro's drafts were nearly as good but used more jargon ("credit under Section 199" instead of "you'll get credit once they file"). Llama and GPT-OSS both tried to explain why the system works this way, which added two paragraphs the client didn't need.

What I do now: I use Claude for TDS mismatch queries. The drafts need almost no editing. I change the client's name, add the specific amount, send. If the client is a first-timer who might want the longer explanation, I'll use Gemini and keep the extra detail.

2. Reverse-charge errors on marketplace invoices: Gemini is faster, Claude is safer

Why it matters: Flipkart, Amazon.in, and Meesho invoices confuse everyone. The platform charges IGST even when buyer and seller are in the same state, and clients assume they've been billed wrong.

I had twenty-three of these queries. The answer is always the same—marketplaces are required to collect tax under Section 9(5), it's not an error, you can claim ITC—but the tone matters. If you sound annoyed, the client feels stupid. If you're too detailed, they stop reading.

Gemini Pro nailed the tone. Its drafts opened with something like "This looks wrong, but it's actually how the law works for marketplace sales," then gave two sentences of explanation and one line on claiming ITC. Claude's drafts were more formal and added a citation to the GST Act, which is good if the client might forward your answer to their own CA, but overkill for a small seller. Llama's answers were fine but slower to generate—thirty seconds versus eight for Gemini. GPT-OSS missed the nuance twice and suggested the client ask for a revised invoice, which is incorrect.

What I do now: Gemini for marketplace reverse-charge questions. It's fast, friendly, and I've never had a follow-up question. I keep Claude for the rare case where the client is a larger business that might need to show the explanation to a statutory auditor.

3. ITC reconciliation mismatches: Claude for complex cases, GPT-OSS for simple ones

Why it matters: When GSTR-2B doesn't match the purchase register, you're either missing invoices, the vendor hasn't filed, or there's a data-entry error. Clients want to know which.

I tested this on thirty-one queries. The complexity varied wildly. Some were simple—"Vendor ABC's March invoice isn't in 2B"—and the answer was just "They haven't filed yet, check again after the 13th." Others involved three vendors, six invoices, and a client who'd recorded some bills in the wrong month.

For the simple ones, GPT-OSS 120B was surprisingly good. It wrote short, direct answers and didn't over-explain. For anything with more than two moving parts, Claude Sonnet 4.5 was the only model I trusted. It would lay out the steps—"First, confirm ABC filed their GSTR-1. Then check if the invoice number in your books matches the one in 2B. If it does and it's still missing, the amount might be below the threshold"—in a way that didn't assume the client knew what GSTR-1 was. Gemini tried to do the same but sometimes skipped a step. Llama's drafts were thorough but read like a checklist, not a conversation.

What I do now: GPT-OSS for one-line ITC questions. Claude for anything that needs a diagnostic walkthrough. I've started keeping a note in each client's WhatsApp chat that flags which model I used, so if they ask a follow-up I don't switch voices.

4. Invoice format errors (missing HSN, wrong GSTIN): Llama explains, Claude fixes

Why it matters: Clients forward you an invoice their vendor sent and ask if it's "okay for ITC." Usually it's missing the HSN code, or the GSTIN has a typo, or there's no tax breakup.

I had nineteen of these. The client doesn't want a lecture on Rule 46; they want to know if they should ask the vendor for a revised invoice or if they can let it go.

Llama 3.3 70B wrote the best explanations. It would say something like "This invoice is missing the HSN code, which is mandatory if the invoice is over ₹50,000. Technically your ITC claim could be disallowed in an audit, so I'd ask the vendor for a revised copy"—then it would list exactly what the revised invoice should include. Claude's drafts were more cautious and always recommended getting a new invoice, even when the error was minor. Gemini's tone was good but it sometimes said "should" when it meant "must," which is a problem when the client forwards your answer to their vendor. GPT-OSS was hit-or-miss; half the time it didn't mention the ₹50,000 threshold.

What I do now: Llama for invoice format questions. The explanations are clear, the risk assessment is balanced, and clients don't come back confused. If the error is serious—wrong GSTIN, no tax amount—I'll use Claude because I want the "get a new invoice" message to sound non-negotiable.

5. Late fee and interest calculations: Gemini is fast, Claude shows working

Why it matters: Clients miss the GSTR-3B deadline, then ask how much the late fee will be. You need to calculate interest on the tax liability, add the fixed late fee, and explain it without making them feel worse.

I tested this on twelve queries. The math is simple—₹50 per day per return (₹20 if turnover is below ₹5 crore), plus 18% annual interest on the tax amount—but clients always ask you to double-check their sums.

Gemini Pro was the fastest. It would calculate the late fee, show the total, and write one empathetic sentence: "The late fee comes to ₹900, which I know is frustrating, but filing today will stop it from going higher." Claude Sonnet 4.5 did the same but also showed the formula and the daily breakdown, which some clients appreciated and others found patronizing. Llama and GPT-OSS both wrote longer answers that included advice on avoiding late fees next time, which is useful but not what the client asked.

What I do now: Gemini for late fee queries where the client just wants the number. Claude when the client is a repeat late-filer and I want the extra detail to nudge them toward setting a reminder.

6. Explaining amendments after filing: Claude for high stakes, Gemini for low

Why it matters: Client files GSTR-1, then realizes they left out an invoice or entered the wrong amount. They want to know if they can fix it and whether there's a penalty.

I had fifteen of these. The answer depends on whether the return period is still open, whether the mistake affects their tax liability, and how risk-averse the client is. Claude Sonnet 4.5 handled the nuance best. It would say "You can amend this in next month's GSTR-1, and because it's an addition, there's no penalty—but the ITC will flow to your customer one month late, so let them know." Gemini's drafts were similar but shorter, which worked for clients who'd done this before. Llama over-explained the amendment rules. GPT-OSS twice suggested filing a revision when it meant an amendment, which is a different process.

What I do now: Claude for first-time amendment questions or anything involving a tax liability change. Gemini for clients who've amended returns before and just need confirmation they're doing it right.

What didn't work: trying to automate the prompt

I spent two days trying to write one master prompt that would handle all query types. The idea was to paste the client's message, let the model detect the category, and generate the right answer. It failed. The models either guessed wrong—routing a reverse-charge question to the ITC reconciliation template—or wrote generic answers that needed so much editing I might as well have started from scratch.

What works better: a simple routing decision that I make. I read the client's message, decide if it's TDS / reverse charge / ITC / invoice format / late fee / amendment, and paste it into the right model with a two-line prompt: "Client query below. Write a clear, empathetic response in 4–5 sentences. Client's name is [X], business is [Y]." That's it. The model has enough context, I'm not wrestling with a 300-word mega-prompt, and I stay in control.

The cost and time breakdown

I tracked the token cost for all 110 queries. Claude Sonnet 4.5 was the most expensive—₹40 for the full batch—but also needed the least editing, so my billable time per query dropped. Gemini Pro cost ₹8 for the same batch and was nearly as good for most query types. Llama 3.3 70B was ₹12, GPT-OSS was ₹6. If you're answering five queries a day, the difference is invisible. If you're doing fifty a week during filing season, Gemini's speed and cost make it the default, with Claude for the hard cases.

Time-wise, I went from an average of eight minutes per query (reading, drafting, editing, sending) to three minutes. Over 110 queries, that's nine hours saved. I billed six of those hours to other work. The rest I spent not working, which also has value.

Questions people ask

Can I use these models for replies that go into official GST proceedings or audits?
No. These are client-facing explanations, not statutory filings or audit responses. For anything that might be read by a GST officer or used in an appeal, you write it yourself or have a senior CA review it. The models are good at explaining rules; they're not qualified to argue your case.

What if the model gets the law wrong?
It happens, especially with niche scenarios or recent circulars. I spot-check every draft before I send it. If I'm unsure, I look up the rule or ask a colleague. The models save time on structure and tone, not on legal research. Don't outsource your judgment.

Do clients notice the answers are AI-assisted?
Not once. The replies sound like me because I edit them to match how I'd say it. The models give me a first draft that's 80% there; I adjust the last 20% for the client's familiarity with GST and how much detail they usually want.

Which model should I start with if I'm new to this?
Gemini Pro. It's fast, cheap, and handles the majority of CA query types well enough that you'll see the time savings immediately. Once you're comfortable, add Claude Sonnet 4.5 for the complex cases—TDS mismatches, ITC reconciliation with multiple vendors, amendments. You can run both in one workspace at Kryotta and route queries based on how much explanation the client needs.

Filing season doesn't have to mean ignoring half your messages until the 21st. Pick the right model for the question, edit the draft to sound like you, and get back to the work that actually needs a CA.

K
Written by
Kryotta Team
Product & research

Kryotta is the multi-model AI workspace — every leading model, one login, one bill. Try it free →

Related reading