Start with the three terms that change what you pay every month
You don't need a computer science degree to pick an AI model for your Shopify store. You need to know which jargon affects your invoice. Tokens are the first word that matters. Every AI model charges per token—roughly four characters of text, or three-quarters of a word. When you paste a customer's complaint about a delayed shipment into Claude and ask for a reply in German, French, and Polish, you're paying for the input (the complaint, your instructions, any order history you included) and the output (three responses).
If you're using AI to answer twenty support tickets a day across four languages, tokens add up. Claude Sonnet 4.5 costs more per token than Gemini Flash, but it writes better apologies and catches tone problems that Flash misses. The checklist question: are you paying for a premium model to draft routine "your order has shipped" emails, or only for the complicated refund negotiations? If you're using Sonnet for everything, you're spending €15 a month when €4 would work. Use Flash or Llama 3.3 70B for confirmations and FAQs. Save Sonnet for the customer who's threatening a chargeback because their package went to a ParcelShop in the wrong town.
Context window is the second term. It's how much text the model can remember in one conversation—your instructions, the customer's message history, and the reply it's building. Gemini Pro has a massive context window. Claude Sonnet's is smaller but still plenty for most support threads. This matters when a customer emails three times about the same order. If you copy-paste all three messages into the AI so it understands the full story, you need a model that won't forget the first email halfway through writing the reply.
The decision: does your support flow involve long threads, or do most tickets get resolved in one exchange? If you sell on Bol or Zalando and customers message once to ask about sizing, a small context window is fine. If you run a Shopify store with complicated custom orders and customers email five times with changes, pay for Gemini Pro or Claude Sonnet. Don't use a model that forgets the customer already said they need the invoice reissued with a VAT number.
Hallucinations are the third term, and the only one that can get you in legal trouble. A hallucination is when the model invents information. It might tell a customer their order shipped yesterday when it hasn't, or cite a return policy you don't have, or promise a refund you never approved. This happens more with smaller, faster models. GPT-OSS 20B hallucinates more than Claude Sonnet. DeepSeek R1 sometimes makes up tracking numbers.
The checklist test: are you using AI to state facts about orders, inventory, or policies, or just to rephrase what you already told it? If you're generating a reply that says "according to your order history, you bought this on March 3rd," make sure the AI is pulling that date from real data, not guessing. Kryotta lets you compare outputs across models. When I draft a response to a GDPR data request, I run it through Sonnet and Gemini Pro, then check that both cite the same information. If one invents a detail, I know not to send it.
The GDPR terms that decide whether you can use a model at all
Grounding means the AI is working from documents or data you gave it, not from memory or the open web. If you upload your return policy, your size chart, and last month's order CSV, a grounded model will answer questions using only that material. An ungrounded model might blend in something it learned during training, which could be wrong or irrelevant.
For GDPR, grounding matters because you need to know what data the model saw and whether it's storing anything. If a customer in Germany asks to see all the data you hold about them, and you've been using an AI that caches conversations or trains on your inputs, you have a problem. Most closed models (Claude, Gemini, GPT-OSS) don't train on your API requests, but you should check the terms. Open-weight models like Llama or Mistral give you more control—you can run them in a way that guarantees nothing leaves your server—but that's overkill unless you're processing thousands of tickets a month.
The decision: are you pasting customer names, addresses, or order details into the AI? If yes, use a model with a clear data policy. Kryotta's privacy page confirms that none of the models it routes to will train on your inputs. If you're running your own instance of Llama, you're responsible for that guarantee yourself. For most Shopify or marketplace sellers, the simpler path is to use a hosted model that contractually won't store personal data, and to strip out unnecessary details before you send the prompt. You don't need to include the customer's full address to draft a reply about a sizing question.
Open-weight versus closed models is the last piece of jargon that changes your options. A closed model (Claude, Gemini) is a black box. You send text in, you get text out, and you don't know how it works inside. An open-weight model (Llama, Mistral, DeepSeek) publishes its architecture and weights, so technically you could download it, inspect it, and run it on your own hardware.
For e-commerce, this distinction matters in two cases. First, if you're handling sensitive data and your legal team won't let you use a third-party API, you need an open-weight model you can host yourself. That's rare for small sellers, but if you're doing €500k a year on Allegro and your contracts require data to stay in Poland, it's a real constraint. Second, if you want to fine-tune the model on your own product catalogue or support history, open-weight is easier. You can train Llama to recognize your SKU codes or to match your brand voice without sending your data to Anthropic.
The checklist question: do you have a compliance reason to self-host, or do you want to customize the model's behaviour beyond what prompts can do? If no, use a closed model. It's faster, cheaper, and someone else handles the updates. If yes, start with Llama 3.3 70B or Mistral Large through Kryotta, test whether the output quality is good enough, then decide if you need to run your own instance.
Which jargon doesn't matter (and what to ignore)
Parameters and training data cutoff dates come up in every AI comparison chart, and neither one affects your Shopify workflow. Parameters are a measure of model size—more parameters usually means better quality, but it's not a simple trade-off. Llama 3.3 70B has fewer parameters than some older models but writes better product descriptions because the training was smarter. You can ignore this number. Test the output instead.
Training cutoff dates matter if you're asking the AI about current events, but you're not. You're asking it to rewrite a return policy or translate a shipping update. The model doesn't need to know what happened last week. If you're using AI to draft blog posts about "EU sustainability regulations in 2025," then yes, check the cutoff and add recent details yourself. For customer support, it's irrelevant.
Inference speed shows up in benchmarks—how many tokens per second the model generates. This matters if you're building a chatbot that needs to reply instantly, but for back-office tasks like writing Bol listings or drafting Klarna payment reminder emails, a two-second delay is fine. Don't pay extra for speed you won't notice.
The one speed issue that does matter: if you're using AI during a live chat with a customer, test whether the model is fast enough that you can read and edit the reply before the customer gets impatient. Gemini Flash is faster than DeepSeek R1. If your live chat tool shows a "typing" indicator for ten seconds while R1 thinks, the customer assumes you're ignoring them. Use Flash for live support, R1 for drafting emails you'll send later.
How to test whether the jargon makes a real difference in your workflow
Pick one task you do every week—say, writing product descriptions for new arrivals, or answering "where is my order" emails in three languages. Run it through two models with different trade-offs. Use Claude Sonnet 4.5 (high cost, low hallucination rate, excellent multilingual quality) and Gemini Flash (low cost, faster, occasionally awkward phrasing).
Time yourself. Count the tokens (Kryotta shows this in the output panel). Note which reply you'd send without editing. If Sonnet's output is perfect and Flash needs three corrections, and you're doing this twenty times a day, the time you save is worth the extra €8 a month. If both replies are fine and you'd edit them anyway, use Flash.
The mistake: testing once, with one prompt, and deciding forever. Your workflow changes. In December I used Sonnet for everything because I was handling complicated Christmas delivery complaints and needed perfect German. In January I switched half my prompts to Flash because the questions got simpler and I wanted the invoice under €10. Check your usage every month. If you're spending €25 on a model you're using for tasks a €5 model could handle, move those tasks.
Questions people ask
Do I need to understand tokens to use AI for my Shopify store, or can I just ignore the pricing page?
You can ignore it for the first month, but once you're using AI daily you should check how many tokens your typical tasks consume. A 200-word product description in English costs about 300 tokens to generate. If you're doing ten a day, that's 90,000 tokens a month—€3 on Gemini Flash, €12 on Claude Sonnet. Knowing this helps you pick the right model and avoid a surprise bill.
If a model hallucinates a tracking number or a return deadline, am I liable under EU consumer law?
Yes. The AI is your tool, and you're responsible for what you send to customers. If you promise a refund date the model invented, and the customer relies on it, that's on you. Always check facts before you hit send. Use grounding (upload your real policies) and compare outputs across two models if the stakes are high.
Can I use open-weight models like Llama without worrying about GDPR, or is that a myth?
Open-weight doesn't automatically mean GDPR-safe. If you're using Llama through a hosted API (like Kryotta), the privacy terms are the same as any closed model—check that the provider won't train on your data. If you self-host Llama, you control everything, but you're also responsible for securing it. For most sellers, a hosted model with clear terms is simpler and just as compliant.
Does a bigger context window mean I can paste my entire product catalogue into one prompt, or will that break something?
Technically you can, but it's expensive and often ineffective. A huge context window costs more tokens, and the model might not focus on the relevant part of your catalogue. Better approach: paste only the five products related to the customer's question, or use a search step first to narrow it down. Save the giant context for cases where you genuinely need the AI to cross-reference fifty SKUs at once.
If you're selling across Shopify, Bol, or Allegro and you're tired of guessing which AI terms matter, Kryotta gives you access to Claude, Gemini, Llama, and Mistral in one workspace. Test two models side by side, see the token count for every reply, and switch models mid-task without opening a new tab. No lock-in, and you only pay for what you use.
