When merchants look at Magento AI chatbot pricing, the first question is rarely the euro amount. It is: what actually counts as a conversation? If every widget open, every stock check, and every “yes/no” reply burned quota, plans would be useless. Claspwell only counts sessions that reach the AI path — scripted catalog turns stay free of the conversation counter.
Three paths, one counter
A catalog-aware Magento assistant should not call a language model for every message. Claspwell uses three-tier routing:
1. Scripted / catalog-only. Category picks, attribute clarifiers, product cards from Postgres SQL and keyword search, and stock/price replies that never need a model. These turns do not count as conversations.
2. Light intent routing. a low-cost LLM classifies the shopper’s goal when the scripted path is not enough. That is the AI path: it counts.
3. Full reply generation. The model drafts a natural-language answer (still grounded in catalog data and a curated knowledge base for policy/B2B questions). That also counts.
There are no embeddings and no vector database. Product discovery stays SQL/keyword search against the synced Magento catalog; the model is for routing and wording, not for “remembering” the whole catalog.
What a merchant sees on each plan
Conversation allowances are per store (each store or isolated store view is its own tenant):
Essential — 750 conversations/month (€99 net), overage €0.10.
Growth — 3,000 conversations/month (€249 net, recommended), overage €0.06, hourly sync.
Scale — 10,000 conversations/month (€499 net), overage €0.05, up to three stores, 15-minute sync.
Agency starts from €990 with multi-store and contract terms. Prices are net EUR; VAT follows the usual EU rules. Billing is by bank-transfer invoice — there is no self-serve card checkout in v1. Full detail lives on the pricing page.
Grace, overage, and the 200% throttle
At 100% of the monthly allowance the merchant is notified. There is one grace window (month or quarter, depending on the billing period) up to 120%. After that, overage is billed at the plan rate. At 200% of allowance, the AI path throttles: the assistant keeps serving deterministic catalog and curated knowledge answers until the plan is upgraded. Shoppers still get product help; the expensive model path pauses.
Why this design matters on Magento
Magento catalogs are attribute-heavy. Most useful turns are “show me red size 42 in this category” — work that category-first discovery and SQL search already handle. Counting those as AI conversations would punish good catalog hygiene. Counting only the AI path rewards merchants who keep attributes clean and lets agencies demo the widget without burning quota on every filter click.
Install is still guided: Magento module + catalog sync + attribute mapping, typically 3 days to 1 week once scope and access are ready. Hyvä gets the Alpine storefront widget; Luma is in scope for manual onboarding.
Practical checklist before you buy
Ask any Magento AI vendor — including Claspwell — these four questions:
1. Do catalog-only / scripted turns count toward the quota?
2. Is the counter per store view or shared across the whole Magento install?
3. What happens at 100%, 120%, and 200% of allowance?
4. Is product search done against a live/synced catalog (SQL/keyword), or is the catalog dumped into a generic GPT context?
If the answers are vague, you will learn the hard way mid-season.
Claspwell is a catalog-aware Magento 2 / Hyvä AI sales assistant for the storefront — not an admin copilot. It syncs categories, products and shopper-facing attributes into Postgres, finds products with SQL and keyword search, and uses a low-cost LLM only for intent routing and replies when a scripted path is not enough. There are no embeddings and no vector database. Install time is typically 3 days to 1 week once scope and access are ready. A curated knowledge base covers B2B and policy questions the catalog cannot answer.
If you want transparent conversation counting on a Magento 2 or Hyvä store, start at claspwell.com or request an integration.