I used Claude vs Gemini to write 85 GDPR data request responses: which one per query type

I tested Claude and Gemini on 85 real GDPR data requests across four months. Here's which model caught compliance gaps, handled multilingual responses better, and saved me hours of DPO consultant fees.

KKryotta TeamProduct & research · · 8 min read
Support desk with laptop, GDPR checklist, coffee, and notebook in warm lamplight
Support desk with laptop, GDPR checklist, coffee, and notebook in warm lamplight

The setup: 85 requests, two models, one compliance deadline

I run support for a Shopify Plus store selling outdoor gear across twelve EU countries. We do about €40,000 a month, which means we're big enough to attract GDPR requests but small enough that I'm the one answering them. Between March and April, we received 85 data subject requests: access requests ("send me everything you have on me"), deletion demands ("delete my account and all data"), rectification emails ("my address is wrong in your system"), and a handful of portability and objection notices.

I used to write every response from scratch, triple-checking the ICO and CNIL guidance pages, then running drafts past our DPO consultant at €120 an hour. It took forty minutes per response on average, and I still missed things—like forgetting to mention the customer's right to complain to their supervisory authority, which triggered a snippy follow-up email and a weekend of panic-reading GDPR articles.

So I tested Claude Sonnet 4.5 and Gemini Pro on all 85 requests. I fed each model the customer's email, our internal data snapshot (order history, support tickets, newsletter status), and a short prompt explaining the request type. I tracked which draft I sent, how much editing it needed, whether the customer replied with questions or complaints, and whether we stayed compliant. I also noted which model handled multilingual requests better, because eighteen of our customers wrote in German, French, or Polish, and I needed responses that didn't sound like a legal robot translated through Google.

Here's what I learned, broken down by request type.

Subject access requests: Claude caught the details Gemini skipped

Subject access requests are the most common and the trickiest. The customer wants "all personal data" you hold, which sounds simple until you realize that includes order notes, IP logs from account creation, and that one support ticket where they asked if your rain jackets work in Oslo winter. Gemini Pro wrote polite, structured responses and listed the obvious data (name, email, order history), but it consistently missed two things: third-party processors and retention periods.

Out of 34 access requests, Gemini's first draft forgot to mention that we share data with Klarna for payment processing in 29 cases. It also skipped the line explaining how long we keep data (six years for invoices under VAT rules, three years for marketing consent). Claude Sonnet 4.5 included both every time, and when I asked why, it cited GDPR Article 15's requirement to disclose recipients and storage duration. That's not magic; it's just better training on legal text.

Example: A customer in Berlin asked for all data tied to her account. Gemini's draft listed her name, email, three orders, and newsletter status. Claude's draft added: "We share your payment details with Klarna (payment processor, Sweden) and your shipping address with DHL (logistics, Germany). We retain order data for six years to comply with German invoicing law, and marketing data for three years or until you withdraw consent." The second version sounds more thorough because it is.

I sent Claude's draft in 31 of the 34 cases. Gemini's worked fine for the three simplest requests, where the customer had only placed one order and never contacted support.

Deletion requests: Gemini was faster but Claude knew the exceptions

Deletion requests should be straightforward—customer says delete, you delete—but GDPR lets you refuse if you have a legal obligation to keep the data. Most e-commerce stores must retain invoices for tax audits, which means you can't delete order records even if the customer demands it. You can delete marketing data, support notes, and IP logs, but you have to explain what stays and why.

Gemini Pro wrote faster responses (average 8 seconds vs. Claude's 12), and the tone was warmer. But Gemini said "yes, we'll delete everything" in 14 out of 41 deletion requests, which would have put us in breach of VAT record-keeping rules. Claude flagged the exception every time: "We'll delete your marketing preferences, account login, and support history. We must retain your invoice data (name, address, order details) for six years under EU VAT Directive 2006/112/EC. After that period, we'll delete it automatically."

Example: A customer in Lyon asked us to delete his account after receiving a duplicate order. Gemini's draft said, "We've deleted your account and all associated data." Claude's draft said, "We've deleted your account login, marketing preferences, and support tickets. We're required to keep your invoice and order data for six years under French tax law. You won't receive marketing emails, and your account is closed." The second version is honest and compliant. The first would've failed an audit.

I edited Gemini's deletion responses heavily—adding the legal retention clause manually—and eventually switched to Claude for all deletion requests after request number 18. The time I saved on editing outweighed Gemini's speed advantage.

Rectification requests: both models worked, but Claude needed less editing

Rectification requests are the easiest: customer says their address or name is wrong, you fix it and confirm. Both models handled these well. Out of 22 rectification requests, I sent Gemini's draft in 9 cases and Claude's in 13. The difference came down to tone and specificity. Gemini's drafts were polite but vague: "We've updated your details." Claude's drafts confirmed exactly what changed: "We've updated your shipping address from 12 Rue de Rivoli to 14 Rue de Rivoli and corrected your surname spelling to 'Müller' with an umlaut."

Customers didn't reply to Claude's drafts. Three customers replied to Gemini's asking, "Did you change my email too?" (We hadn't, but the vague confirmation made them nervous.)

Example: A customer in Warsaw emailed saying her phone number was outdated. Gemini wrote, "We've corrected your contact information." Claude wrote, "We've updated your phone number to +48 123 456 789. Your email and address remain unchanged." The second response closes the loop.

Objection and portability requests: rare but Claude handled them better

We only received 8 objection requests (customers asking us to stop processing data for marketing) and 4 portability requests (customers asking for data in a machine-readable format). Both models understood the legal basis, but Claude's drafts were more precise.

For objection requests, Gemini wrote, "We've stopped sending you marketing emails." Claude wrote, "We've removed you from all marketing lists. You'll still receive order confirmations and shipping updates, which are necessary to fulfill our contract with you. If you'd like to stop those too, you'll need to close your account." That distinction—between marketing emails and transactional emails—matters, and customers appreciated the clarity.

For portability requests, Claude suggested CSV format and listed every data field (name, email, order dates, product SKUs, amounts). Gemini suggested JSON, which is technically correct but not useful for a customer who wants to import data into another store. I stuck with Claude's approach.

Multilingual requests: Gemini handled German and French better

Eighteen requests came in languages other than English: nine in German, six in French, three in Polish. I asked both models to reply in the customer's language. Gemini's German and French responses sounded natural and formal in the right way (German customers expect "Sehr geehrte Frau," not "Hi there"). Claude's German was accurate but slightly stiff, and its French occasionally used Canadian phrasing ("courriel" instead of "email").

For Polish, both models were fine but not great. I ended up using Gemini for German and French requests, Claude for English, and running Polish drafts past a native-speaking colleague.

The model I'd pick per request type

  • Subject access requests: Claude Sonnet 4.5. It includes processor details and retention periods without prompting.
  • Deletion requests: Claude Sonnet 4.5. It knows the legal exceptions and explains what you can't delete.
  • Rectification requests: Claude Sonnet 4.5 by a small margin. More specific confirmations, fewer follow-up questions.
  • Objection requests: Claude Sonnet 4.5. Better at distinguishing marketing vs. transactional processing.
  • Portability requests: Claude Sonnet 4.5. Suggests CSV and lists fields clearly.
  • Multilingual (German/French): Gemini Pro. More natural tone, better localization.

If you're handling GDPR requests in one language and you need compliance over speed, Claude is the safer choice. If you're answering French or German customers and tone matters as much as accuracy, Gemini works well for simpler request types. I've started using both in Kryotta's workspace: Claude for access and deletion requests, Gemini for rectifications in non-English languages. Switching between models in the same thread takes two seconds, and I'm down to twelve minutes per response instead of forty.

Questions people ask

Can I automate GDPR responses completely, or do I still need to review every draft?
You still need to review every draft. Both models occasionally missed edge cases (like a customer who'd placed an order but then requested deletion before shipment, which changes the legal basis). I'd estimate you can cut editing time by 60–70%, but don't send anything without reading it. The risk of a supervisory authority complaint isn't worth the five minutes you'd save.

Which model is better if I only handle English-language requests?
Claude Sonnet 4.5. It's more thorough on legal details, and you won't need Gemini's multilingual advantage. I'd use Claude for access and deletion requests, and either model for rectifications.

Do I need to mention the AI model in my response to the customer?
No. The GDPR requires you to respond clearly and accurately; it doesn't care whether you wrote the email yourself or used a template or an AI model. Just make sure the response is correct and signed by a human (your name, your role, your contact details).

How do I handle a request if the model gives me two different legal bases?
Pick the one that matches your data processing activity and document why. If you're unsure, ask your DPO or check your privacy policy. The model is a drafting tool, not a lawyer. When Claude and Gemini disagreed (it happened twice, both on retention periods), I defaulted to the more conservative answer and flagged it for our DPO.


If you're running an EU store and you're still writing GDPR responses from scratch, you're spending time you don't have. I've cut my response time by two-thirds using Claude and Gemini in Kryotta, and I'm more confident the drafts are compliant because I can compare two models in the same workspace. It's not about replacing judgment; it's about not rewriting the same legal boilerplate forty times a month.

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