Case study · Travel & Hospitality
A blocked WhatsApp inbox, turned into a 24/7 AI sales channel
Condense built ARIA, a production conversational agent on WhatsApp Business API with Claude as its reasoning core, taking Erga Tourism from missed messages to prefilled checkout carts in seconds.
1000+
customer conversations handled across two WhatsApp numbers
~20 sec
median time from customer message to agent reply, live bookings included
300
prefilled checkout carts generated by the agent, ~1 in 2 converting to a paid order
4 / 5
outbound replies written by the AI, freeing staff for cases that need judgment
The challenge
Five out of ten potential customers were already lost
Erga Tourism is a ~20-person incoming DMC selling excursions, car and boat rentals, airport transfers and ticketing across Sicily, Sardinia, Puglia and Umbria. In high season it fields more than 100 contacts a day, split evenly between pre-sales and post-sales.
Phone was the primary channel; WhatsApp existed but was fragmented across different numbers per brand and connected to nothing. An earlier attempt at a manually staffed chat had already been shut down as operationally unsustainable, and product knowledge was scattered across four separate back-office systems, so even a simple availability check required a human to go and look.
By the team's own estimate, roughly five out of ten potential customers were already being lost for lack of a timely answer.
The solution
ARIA: a production agent, not a scripted chatbot
Condense put Claude at the reasoning core of a WhatsApp Business API agent, wired to Erga's own systems with deterministic tools, so it can actually complete a booking — not just answer questions about one.
1
Working memory
Claude is the single writer of conversation state, holding the thread across slot-filling, corrections and mid-conversation changes, even across sessions.
2
Fifteen tools
Live calls to Erga's booking API for catalog, availability and pricing, across an eight-step order flow.
3
Prefilled cart
A prefilled guest-checkout cart that turns a conversation straight into a ready-to-pay order.
4
Bilingual knowledge
A hybrid knowledge layer with cross-lingual embeddings answers pre- and post-sales questions in the customer's own language, auto-detected from the first message.
Erga's staff can oversee and manage conversations from a multi-tenant operator console with a WhatsApp-Web-like inbox and seamless bot↔human takeover, backed by a full audit log and analytics.
Continuous evaluation aligned with business metrics
Every iteration of the agent is validated against business-critical outcomes rather than subjective impressions. End-to-end multi-turn purchase simulations and blind head-to-head evaluations on actual production conversations safeguard conversion rates, booking accuracy, and brand consistency at every release.
The roadmap
TALK TO CONDENSE →
What's next
Building on the same Claude-powered runtime, Condense and Erga are now automating outbound quote generation for sales campaigns: an agent that qualifies leads in real time over WhatsApp, pulls live pricing from Erga's API, and drafts a branded quote for human approval before it ever reaches the customer.
Building something similar?
Condense designs and ships production Claude agents wired to real business systems, not demos.