The setup: 140 queries, three models, one very nervous April
I handle invoicing and customer data for a Shopify store selling industrial supplies—mostly fasteners, adhesives, and small tools—to B2B customers across eight EU countries. We do about €65,000 a month, which puts us in the awkward middle: large enough that the Finanzamt notices when something's wrong, small enough that I'm the one fixing it.
Between mid-March and late April, we received 140 queries from tax offices. Most came from Germany's Finanzamt, but we also heard from the Dutch Belastingdienst, Austria's Finanzamt, and France's DGFiP. The emails fell into five categories: missing or incorrect VAT IDs on invoices, reverse-charge confusion (we'd charged VAT when we shouldn't have, or vice versa), invoices with formatting errors (wrong date formats, missing line items), cross-border B2B documentation gaps, and a few "please explain this transaction" requests that required digging into three-month-old Slack threads to remember what actually happened.
I used to write every response by hand. I'd open the original invoice PDF, cross-check our Stripe metadata, re-read the EU VAT Directive section on reverse charge for the dozenth time, then draft a reply in what I hoped was appropriately formal German or French. Each response took thirty to fifty minutes. I'd send it to our tax advisor in Hamburg for a quick check (€95 per review), wait a day, revise, send. By the time I cleared the queue, three more queries had arrived.
So I tested Claude Sonnet 4.5, Gemini Pro, and DeepSeek V3 on all 140 queries. I fed each model the Finanzamt email (translated to English when it wasn't already), the relevant invoice details, and a short prompt: "Draft a formal reply to this tax-office query. Explain what happened, cite the correct VAT treatment, and attach the corrected invoice if needed." I tracked which draft I sent, how much editing it required, whether the tax office accepted the response or came back with more questions, and whether the model invented any details. I also noted which model handled German bureaucratic tone without sounding like a nervous tourist, and which ones understood reverse-charge mechanics well enough that I didn't have to rewrite the VAT logic from scratch.
Here's what I learned, broken down by the type of query that landed in my inbox.
Missing VAT ID queries: Claude nailed the tone, Gemini rushed the apology
Forty-three queries were about missing or incorrect VAT IDs. A customer's VAT number wasn't on the invoice, or we'd written "DE123456789" when the validated number was "DE 123 456 789" (spaces matter, apparently), or we'd issued an invoice before the customer's VAT registration came through and now the Finanzamt wanted to know why we charged German VAT on a B2B sale.
Claude Sonnet 4.5 wrote replies that sounded like they came from someone who'd done this before. Formal but not stiff. It opened with a direct acknowledgment—"Thank you for your inquiry regarding invoice 2024-03-0847"—explained what happened ("The customer's VAT ID was validated on 14 March, but the invoice was generated on 12 March using our standard B2C template"), cited the correction ("We have issued a revised invoice with the validated VAT ID DE 234 567 890 and applied the reverse-charge mechanism under Article 194 of the VAT Directive"), and closed with "Please find the corrected invoice attached. Should you require further documentation, we are happy to provide it."
I sent thirty-one of Claude's drafts with only minor edits—usually adding the customer's company name or adjusting the attachment reference. The Finanzamt accepted twenty-nine without follow-up. Two came back asking for proof of VAT validation, which was fair; I sent the VIES screenshot and that closed it.
Gemini Pro wrote replies that worked but felt over-apologetic. It opened with "We sincerely apologize for the inconvenience caused by this oversight" even when the issue was the customer's VAT registration timing, not our mistake. The explanations were accurate—Gemini got the reverse-charge rules right—but the tone made us sound like we'd committed fraud rather than hit a two-day registration gap. I used eight of Gemini's drafts, all of them rewritten to dial back the apology and add more neutral phrasing.
DeepSeek V3 invented invoice numbers. Twice it cited "invoice 2024-03-0991" in the reply when the actual invoice was 2024-03-0889. Once it wrote "the customer's VAT ID DE 987 654 321 has been validated" when the real number was entirely different. I caught these before sending, but it meant I couldn't trust any draft without line-by-line checking. I didn't send a single DeepSeek response for this category.
Reverse-charge confusion: Claude explained it clearly, DeepSeek hallucinated the law
Thirty-eight queries were about reverse-charge errors. We'd charged German VAT on a B2B sale to a French customer (we shouldn't have; the customer should self-assess French VAT). Or we'd applied reverse charge to a sale within Germany (wrong; that's a standard German VAT sale). Or we'd issued a credit note for a reverse-charge invoice and the Finanzamt wanted to know why the VAT correction didn't appear in our return.
Claude Sonnet 4.5 explained reverse charge in plain language without dumbing it down. For a query about a €3,400 sale to a Lyon-based distributor where we'd mistakenly charged 19% German VAT, Claude wrote: "This was a cross-border B2B supply of goods. Under Article 194 of the EU VAT Directive and Section 13b of the German VAT Act (UStG), the place of supply is France, and the reverse-charge mechanism applies. The customer is liable for French VAT; we should not have charged German VAT. We have issued a corrected invoice with no VAT charged and a note: 'Reverse charge – customer to self-assess VAT.' The original VAT of €646 has been credited in our April return."
I sent twenty-two of Claude's reverse-charge replies with minimal edits. The tax offices accepted twenty without follow-up. Two asked for the credit note, which I sent, and that was the end of it.
Gemini Pro got the rules right but buried the key point. It spent two paragraphs explaining what reverse charge is before saying what we did wrong. For the same Lyon sale, Gemini opened with a definition of cross-border B2B supplies, then explained Article 194, then finally admitted the error in paragraph three. I used six of Gemini's drafts, all of them restructured to put the correction up front.
DeepSeek V3 cited a section of German tax law that doesn't exist. It wrote "Under Section 13b(5) of UStG, the reverse-charge mechanism applies to cross-border supplies" when Section 13b has no subsection 5 in that context. It also claimed we'd "submitted a corrected VIES return" for a transaction that didn't require VIES reporting. I didn't send any DeepSeek drafts for reverse-charge queries. The risk of a tax office catching a fake citation was too high.
Invoice formatting errors: Gemini caught the details, Claude missed one date format
Twenty-nine queries were about formatting: wrong date formats (we'd used MM/DD/YYYY on a few invoices when the German standard is DD.MM.YYYY), missing line-item descriptions (we'd written "Fasteners" instead of "Stainless steel hex bolts M8x50mm, 500 units"), or invoices that didn't show the customer's full registered address.
Gemini Pro was excellent here. It spotted the formatting issue in the prompt, explained what the standard required, and drafted a reply that cited the specific line: "Invoice 2024-03-1024 used the date format 03/17/2024. German invoicing standards (Section 14 UStG) require the format 17.03.2024. We have issued a corrected invoice with the proper date format and updated line-item descriptions to include product specifications." I sent nineteen of Gemini's drafts with almost no changes.
Claude Sonnet 4.5 was nearly as good but missed one detail: it didn't always flag that the date format was wrong unless I explicitly mentioned it in the prompt. For a query about a missing customer address, Claude wrote a clear reply and attached a corrected invoice, but when the same invoice also had a date-format error, Claude didn't mention it. I caught it during review and added a line. I sent seven of Claude's drafts for formatting queries, all of them lightly edited.
DeepSeek V3 rewrote invoice numbers again. For a query about invoice 2024-04-0234, DeepSeek's draft referred to "the corrected invoice 2024-04-0235" when we hadn't issued a new invoice number—we'd replaced the PDF under the same number. I didn't use any DeepSeek drafts here.
Cross-border B2B documentation: Claude handled the nuance, Gemini wanted more proof than required
Eighteen queries asked for additional documentation: proof that the goods left Germany (CMR shipping documents, carrier tracking), proof that the customer was VAT-registered at the time of sale (VIES printouts), or proof that the sale qualified as an intra-community supply rather than a domestic sale.
Claude Sonnet 4.5 wrote replies that provided exactly what was asked for and no more. For a query about a €5,200 sale to a Polish customer, Claude drafted: "The goods were shipped on 22 March 2024 via DHL to the customer's registered address in Kraków. Attached are the CMR document, the carrier's proof of delivery, and the VIES validation printout showing the customer's Polish VAT ID was active on the date of sale. This qualifies as an intra-community supply under Article 138 of the VAT Directive." I sent eleven of Claude's drafts with zero edits.
Gemini Pro over-explained and suggested attaching documents the tax office hadn't requested. For the same Polish sale, Gemini's draft included a paragraph about how we validate VAT IDs and offered to send "our internal shipping logs and Stripe payment confirmation" when the Finanzamt had only asked for the CMR and VIES printout. I used three of Gemini's drafts, trimmed down.
DeepSeek V3 didn't understand what a CMR document was. It wrote "Attached is the shipping receipt" when the query specifically asked for a CMR (Convention relative au contrat de transport international de marchandises par route—the standard international consignment note). I didn't send any DeepSeek drafts for documentation queries.
"Please explain this transaction" requests: Claude asked clarifying questions, DeepSeek made up a story
Twelve queries were open-ended: "Please explain the nature of this transaction" or "Why was this invoice issued twice?" These required me to dig into Slack, check our Shopify order notes, and sometimes email the customer to confirm what actually happened.
Claude Sonnet 4.5 was the only model that asked me for more information before drafting. When I fed it a query about an invoice issued twice, Claude replied in the workspace: "I need more context. Was this a duplicate invoice sent in error, or was a second invoice issued to correct the first? If it's a correction, what changed?" That saved me from sending a vague reply. Once I provided the details—we'd issued a second invoice because the first one had the wrong VAT rate—Claude drafted a clear explanation. I sent eight of Claude's drafts for this category.
Gemini Pro guessed. For the duplicate-invoice query, Gemini assumed it was a system error and drafted an apology without knowing what had actually happened. I didn't send any of Gemini's drafts for open-ended queries.
DeepSeek V3 invented a narrative. For a query about a €1,800 invoice to a Belgian customer, DeepSeek wrote: "This invoice was issued to correct a previous billing error in which the customer was undercharged by €340." We hadn't undercharged anyone. The invoice was for a separate order. I didn't send any DeepSeek drafts here either.
What I'd do differently next time
I'd use Claude Sonnet 4.5 for anything that requires formal tone, reverse-charge explanations, or replies to open-ended queries. It's the only model that sounded like a competent human and didn't invent details. I'd use Gemini Pro for formatting and documentation queries where the task is mostly "list what's attached and confirm it meets the requirement." I wouldn't use DeepSeek V3 for any tax-office correspondence. The hallucinated invoice numbers and fake legal citations are disqualifying.
I'd also build a shared prompt library in Kryotta with standard clauses for reverse charge, VIES validation, and CMR explanations, so I'm not rewriting the same context every time. And I'd keep a log of which queries each model handled well, because the next time the Finanzamt emails, I'll know which model to open first.
The 140 queries took me nineteen hours total, down from an estimated seventy hours if I'd written every reply from scratch. Our tax advisor reviewed twelve of the trickier responses (the ones involving credit notes and multi-country shipments) at €95 each, which cost €1,140 but saved us from filing incorrect corrections. The Finanzamt accepted 134 responses without follow-up. Six came back with clarifying questions, all of them reasonable, and we closed those within a week.
Questions people ask
Can I use AI to reply directly to tax-office emails without reviewing the draft?
No. I caught hallucinated invoice numbers, fake legal citations, and wrong VAT logic in about 15% of drafts. Always review before sending, especially for anything involving money or compliance deadlines.
Which model is best if I only handle German Finanzamt queries?
Claude Sonnet 4.5. It handled formal German bureaucratic tone better than the others and didn't invent details. Gemini Pro is a decent backup for straightforward formatting queries.
What if the tax office writes in German and I don't speak German?
I translated every email to English using DeepL before feeding it to the models, then translated the final draft back to German. The models wrote better replies in English than in German, and DeepL's translations were accurate enough that the Finanzamt didn't question them.
Do I need to cite the EU VAT Directive in every reply?
Not always, but it helps. For reverse-charge and cross-border queries, citing Article 138 or Article 194 showed the tax office we knew what we were talking about. For formatting errors, citing Section 14 UStG was enough.
If you're facing a pile of tax-office queries and you'd rather not spend three days writing the same explanation twelve different ways, try drafting a few replies in Kryotta and see which model gets your tone right. Start with Claude for anything formal, Gemini for anything detail-heavy, and always review the VAT logic before you hit send.



