La Salve
AI customer service chatbot for La Salve: it resolves customer queries 24/7 in any language, and turns the rest into tickets that are already classified and routed automatically.
The Challenge
La Salve, the craft brewery from Bilbao owned by the Mahou group, managed its online store's info@ inbox by hand, with more than 20 recurring email types: shipment tracking and delays (the most frequent), coupons and promotions, availability of out-of-stock products, transport damage, address changes, duplicate orders, payments and account access. Every email had to be read, classified and forwarded by hand to whoever it concerned, and the customer waited until someone was available to reply. The inbox had become a bottleneck that consumed the team's hours and delayed every response.
The Solution
We built a custom WordPress plugin with an AI assistant connected to the store (WooCommerce) and to a full ticketing system. The bot understands each query in the customer's own language, verifies their identity when needed, checks real order data and resolves on the spot: shipment tracking, address changes, invoice resending, coupon validation or adding the customer to the waiting list for an out-of-stock product. When it cannot resolve, it escalates to a ticket that is born already classified and assigned to the right transport agency or team. Underneath, a layered architecture over a REST API, seven dedicated tables, Claude Sonnet 5 by Anthropic as the reasoning engine and prompt caching to lower the cost of every conversation.
Technologies.
The tools that made this project possible
How a request flows.
- 01 The customer writes in the store widget, in their own language; the assistant detects it and always replies in that same language
- 02 Claude interprets the query and decides which tool it needs: order, shipment, invoice, stock, cart or discounts
- 03 Identity verification is handled by code, not by the model: without a name, email and order number that match WooCommerce, no data is shown
- 04 The tools query real data: the order and invoice in WooCommerce, shipment tracking via the transport agency's API
- 05 When it can, it resolves on the spot: shipment tracking, address change, invoice resending or adding the customer to the waiting list for an out-of-stock product
- 06 When it can't or shouldn't, it escalates: it creates a ticket that is born already classified into one of the 8 categories and routed to the right team or transport agency
What is AI and what is deterministic.
AI components
The model (Claude, by Anthropic) understands the query in the customer's language, reasons about which tool to invoke, drafts the reply and classifies the conversation by category. It also translates the knowledge base content on the fly when the customer writes in a language other than Spanish.
Deterministic components
Everything sensitive runs through conventional code: verifying that the order belongs to whoever is asking (server-side email comparison), which address fields can be changed and in which order states, the list of discounts that can be announced, ticket routing by category and carrier, and automatic closures. The model can't bypass these rules: they're enforced on the server even if the conversation asked otherwise.
What the system does not do.
- It never reveals anything about an order without verifying identity: the email comparison against WooCommerce is done by the server, not the model
- It doesn't change the address once the order is already with the carrier: the block is by order status, at code level, and the request escalates to a ticket
- It can't touch amounts or order lines, or process refunds: that capability doesn't exist in the system; those cases always end up in a ticket for the team
- It only mentions discounts from a list approved by marketing; private or test coupons are never exposed
- It is forbidden from inventing prices, promotions or availability, discussing internal company data or giving legal advice; faced with an ambiguous query, it escalates instead of improvising
Results.
Metrics in detail.
| Metric | Result |
|---|---|
| Recurring email types mapped during discovery | 20+ |
| Ticket categories with automatic routing | 8 |
| Store tools the assistant can invoke | 8 |
| Dedicated database tables | 7 |
| Assistant response time | 1-9 s |
| Availability | 24/7 |
Methodology: Design and configuration figures for the system in production (July 2026), verified against the plugin's code. Operational metrics (automatic resolution rate, classification accuracy, monthly volumes) are being instrumented; we'll publish them with their period and measurement method.
Process.
Discovery
Mapping the info@ inbox: more than 20 email types, volumes and who each one was forwarded to
Flows
Defining the incident flows and the 8 ticket categories with their routing together with the client
AI + WooCommerce
Claude Sonnet 5 assistant connected to WooCommerce, with identity verification and real order lookups
Tickets and portals
Ticketing system with statuses, attachments and one portal per transport agency, each seeing only its own shipments
Production and improvement
Go-live and continuous improvement of the bot from real conversations
The Platform.
Let's Talk Results.
Tell us your challenge. We'll respond within 24h.
Start Conversation