Most of what a shopper types into a store chat does not need a large language model. "Hi", "show me oak tables", "anything under 200 euros in size L?" are questions your Magento catalog can already answer. Sending every one of them to the most expensive model is how AI chat budgets get out of hand. Claspwell uses three-tier AI routing to avoid that. Here is how it works and why it matters when you compare vendors.
The three tiers
- Scripted. Predictable turns are handled by fixed logic and templates, with no model call at all.
- Light intent. When the assistant needs to understand what the shopper wants, a small, low-cost model turns the message into a structured intent: which category, which attributes, what kind of question.
- Full model. A larger model is used only when the first two tiers can't produce a good answer. On Growth and above, this frontier-model escalation is capped, so a busy month can't quietly multiply your costs.
The model routes, the catalog answers
An important detail: in Claspwell the model does not invent products. Once the intent is clear, the product lookup runs in code against your synced Magento catalog, category first and then attributes, typically in no more than two catalog queries per turn. The product cards the shopper sees are built from that catalog data, with real SKUs and prices, not from text the model wrote. That is also why good attribute data matters so much. Our attribute checklist covers what to fix first.
Why this matters for your bill
On every Claspwell plan, only sessions that reach the AI path count toward your included conversations. A session handled entirely by scripted and catalog logic doesn't use up your allowance. In practice that means:
- Your allowance (750, 3,000 or 10,000 conversations a month, depending on the plan) is spent on conversations where AI actually did the work.
- Overage is billed per extra conversation at your tier rate, so costs stay predictable instead of depending on token counts.
- If usage reaches 200% of the allowance, the AI path is throttled and the assistant falls back to deterministic catalog and knowledge answers until you upgrade. Shoppers still get product answers; the store doesn't go silent.
Full numbers per plan are on the pricing page.
Why it matters for answer quality
Routing is not only a cost trick. A generic GPT wrapper that answers everything from the model can describe products you don't sell or quote prices that changed last week. Keeping the model in the job of understanding the question, and leaving the answer to your catalog and your curated store knowledge, keeps replies tied to what is actually on your storefront.
Questions to ask any Magento chat vendor
- Does every message go to an LLM, or only the ones that need it?
- Are product results pulled from the Magento catalog, or generated by the model?
- What exactly counts as a billable conversation?
- What happens when we go over the allowance: hard stop, surprise invoice, or a fallback?
- Can we cap use of the most expensive model?
Where Claspwell fits
Claspwell is a catalog-aware AI sales assistant for Magento 2 and Hyvä, with three-tier routing on every plan. We install the Magento module with you, typically in 3 days to 1 week once scope and access are ready. See how the Magento 2 AI chatbot works, or request an integration.