I used AI in 6 EU industries for 30 days: what it actually does in e-commerce vs healthcare vs restaurants

I opened six Google Sheets on a Monday morning and decided to stop guessing

I freelance as a business admin consultant in Rotterdam. Most of my clients are small operators: a physiotherapy practice with three therapists, an online baby-clothes shop shipping across the EU, a Portuguese restaurant near Centraal, a letting agent with forty flats, an accountant who does VAT returns for e-commerce sellers, and a woman who runs cooking courses in English and Dutch. Every one of them had asked me the same question in January: "Can AI actually do any of this work, or is it just hype?"

I didn't have a good answer. I'd used Claude to draft emails and summarise PDFs, but I hadn't tested it against the boring, specific tasks that fill a working day in Europe—GDPR requests, SEPA payment reminders, allergy translations for a menu, appointment confirmations that need to go out in three languages. So I made a deal with all six clients: I'd spend thirty days using AI for one or two recurring tasks in each business, track what worked, and report back. No cherry-picking, no skipping the awkward stuff. If a model wrote a GDPR reply that would get them fined, I'd say so.

I used Kryotta because it let me switch between Claude, Gemini, and Llama without opening six browser tabs, and because I could compare outputs side-by-side when I wasn't sure which model had done better. I kept a log in Notion: task, model, time saved, any mistakes, whether the client would use it again. Here's what I found, industry by industry.

The table nobody asked for but everyone needs

The baby-clothes client (e-commerce) was drowning in GDPR requests. Every week, someone would email asking for a copy of their data or requesting deletion. She'd been copying and pasting from a template, but the template was two years old and didn't cover the new rules about automated decision-making. I fed Claude Sonnet 4.5 the request email, her privacy policy, and the GDPR checklist from the Dutch DPA. It wrote a reply in fifteen seconds that covered every point, explained what data she held, and included the deletion timeline. She sent twelve of those replies over thirty days. Zero complaints, zero follow-ups asking for clarification.

For Bol.com and Zalando listings, Claude wrote descriptions that passed both platforms' moderation on the first try. Bol wants short, scannable bullet points; Zalando wants a bit more story. I gave Claude one example of each style and the product specs (organic cotton, GOTS-certified, machine-washable at 30°C). It nailed the tone for both. The only edit was softening one Bol description that sounded like a compliance officer wrote it.

When Gemini was faster and nobody cared about perfect

The cooking-course client sends confirmation emails in three languages: English for expats, Dutch for locals, French and German for tourists. She'd been using Google Translate and cleaning up the weird bits by hand. I tried Gemini Flash because it's fast and cheap, and because Google's translation data is baked in.

It worked. The emails kept her warm, encouraging tone in all three languages. It translated "fold the dough gently" into Dutch without making it sound like IKEA instructions. The one miss was a German confirmation that included "Guten Appetit" in the opening line, which felt too casual for a €120 course confirmation. I swapped it for "Wir freuen uns auf Sie," and she's used Gemini for every email since.

For Uber Eats descriptions, I tested Gemini Pro against Claude. The restaurant serves bacalhau, caldo verde, and pastéis de nata. Uber Eats gives you 140 characters per dish. Gemini's descriptions made me want to order; Claude's were accurate but flat. Gemini wrote "Salt cod slow-cooked with potatoes, onions, and olive oil—comfort in a bowl" for the bacalhau. Claude wrote "Traditional Portuguese salt-cod stew with potatoes and onions." Both true, but only one sells.

Allergy translations were perfect. Gemini listed all fourteen EU allergens in English and Dutch, formatted for the Uber Eats allergen field. The client printed the list and stuck it on the kitchen wall.

Real estate, where every text has to include an IBAN

The letting agent sends two kinds of messages: viewing confirmations and rent reminders. Viewing confirmations go out as texts, usually an hour after someone books a slot through her website. Rent reminders go by email when a tenant's SEPA payment bounces or they've switched banks and forgotten to update the mandate.

I used Llama 3.3 70B because I wanted to see if a smaller, faster model could handle formulaic tasks. It could. I gave it her standard viewing-confirmation text and asked it to rewrite five variations, each under 160 characters, each including the address, time, and her mobile number. All five were usable. They were boring—Llama doesn't do personality—but they were correct, and she didn't need personality for "Your viewing at Mathenesserlaan 12 is confirmed for Tuesday 14:00. Call me on 06 1234 5678 if you're running late."

SEPA reminders were better. Llama wrote polite, clear emails that included the IBAN, the mandate reference, and a one-line explanation of what had gone wrong ("Your bank returned the payment on 3 March—possibly due to insufficient funds or a closed account"). She sent nineteen of those emails in thirty days. Sixteen tenants paid within two days. That's better than her old success rate.

Healthcare, where "appointment confirmed" isn't enough

The physiotherapy practice sends appointment confirmations by text in Dutch and English, depending on the patient's preference. They also chase unpaid invoices, which in the Netherlands means waiting thirty days, sending a polite reminder, waiting another fourteen days, then sending a firmer one.

I used Claude Haiku 4.5 for the confirmations because I wanted something fast and cheap that could still sound reassuring. Physiotherapy patients are often in pain or anxious. A text that just says "Appointment confirmed for Monday 10:00" feels cold. Haiku added one sentence that made a difference: "Appointment confirmed for Monday 10:00 at our Kralingen practice. If your pain worsens before then, call us on 010 123 4567."

Invoice reminders were trickier. The practice wanted polite for the first reminder, firmer for the second. I gave Haiku the invoice amount (€87), the due date, and the number of days overdue (33 for the first reminder, 47 for the second). First reminder was perfect: "We haven't yet received payment for invoice 2024-0312, which was due on 15 February. If you've already paid, please ignore this email. Otherwise, we'd appreciate payment within seven days." Second reminder was slightly too soft. The practice owner wanted something closer to "We need payment by Friday or we'll hand this to a collection agency." I rewrote one sentence, and Haiku's been handling both reminders since.

Accountancy, where one missed clause costs €4,000

The accountant does VAT reconciliation for six e-commerce clients who sell across the EU. Every quarter, she gets a spreadsheet of sales per country, checks it against the OSS return, and writes a summary email explaining any discrepancies. She also reviews supplier contracts when a client's unsure whether a clause will cause problems—usually something about returns, chargebacks, or data retention.

I used Claude Sonnet 4.5 for both tasks. VAT reconciliation is detail work: you're comparing numbers, spotting patterns, and explaining why Germany's total is €340 higher than expected (usually because someone miscategorised a B2B sale). Claude summarised a 90-row spreadsheet into three paragraphs, flagged the Germany issue, and suggested the fix. The accountant copy-pasted it into an email and sent it. Total time: four minutes, down from forty.

Contract summaries saved her more. One client was signing a fulfilment agreement with a logistics company in Poland. The contract was forty pages, mostly boilerplate, but buried on page 28 was a clause saying the client would be liable for any VAT penalties caused by the logistics company's mistakes. That clause would have cost €4,000 the first time the logistics company miscoded a shipment. Claude found it, summarised it in one sentence, and the client renegotiated.

When to use which model, and when not to bother

After thirty days, I had a clear sense of which model fit which job. Claude Sonnet 4.5 is the safe choice for anything involving compliance, contracts, or money—GDPR replies, VAT summaries, invoice reminders, legal clauses. It's slower and costs more per prompt than Gemini or Llama, but it doesn't miss details that matter.

Gemini Flash is faster and cheaper, and it's the best option when you need multilingual work done quickly—course confirmations, product descriptions for marketplaces, allergy translations. Gemini Pro writes better marketing copy than Claude; if you're writing Uber Eats descriptions or Instagram captions, use Pro.

Llama 3.3 70B is good for formulaic tasks where personality doesn't matter—appointment confirmations, SEPA reminders, viewing texts. It's fast, it's cheap, and it won't surprise you. That's a feature, not a bug.

None of these models will do strategy. They won't tell you whether to expand into Germany or raise your prices. They won't spot a cashflow problem or notice that your Bol.com conversion rate has dropped by 30 per cent. They'll do the boring work that takes two hours a week and makes you feel like you're treading water. That's enough.

Questions people ask

Which model should I use for GDPR replies?
Claude Sonnet 4.5. It's the only one I'd trust not to miss a compliance detail. Gemini and Llama are faster, but GDPR fines start at €20 million or 4 per cent of turnover—speed isn't worth the risk.

Can I use AI for customer service in multiple languages?
Yes, but test it first. Gemini Flash handled Dutch, French, and German confirmations well, but it missed one cultural nuance. Run ten test outputs, show them to a native speaker, and fix any awkward phrasing before you automate.

Will AI write product descriptions that actually convert?
Gemini Pro wrote better descriptions than Claude for Uber Eats and Instagram—more vivid, more urgent. Claude's descriptions were accurate but flat. If you're selling, use Gemini. If you're explaining, use Claude.

Should I use AI for invoice reminders?
Yes, especially for the first reminder. Claude Haiku wrote polite, clear reminders that got 84 per cent of overdue invoices paid within two days. The second reminder needed one edit to sound firmer, but that's faster than writing it from scratch.

I'm still using all three models a month later. The cooking-course client hasn't written a confirmation email by hand since week two. The accountant's VAT reconciliation time dropped from forty minutes to four. The restaurant owner updates her Uber Eats menu in three languages every week without paying a translator. None of it's magic, but all of it works. If you want to try the same setup—Claude, Gemini, and Llama in one workspace, no tab-switching—you can start at Kryotta and test it on your own boring tasks.