Results
The challenge
A Mumbai-based D2C beauty brand with ~15,000 monthly orders was hitting WhatsApp support ceiling. The human team of 6 could handle ~400 conversations per day at high quality; the brand was getting 1,000+ on festival weeks. The result was a predictable pattern: slower responses → angrier customers → more escalations → even slower responses.
The standard D2C answer — hire more — didn't work. Finding and training quality WhatsApp support agents in Mumbai who could handle the brand's bilingual (English + Hindi + sometimes Marathi) tone was hard. Attrition was high. Training cost was ~₹30k per agent before they hit full productivity.
Leadership wanted to keep human agents for the hard conversations (product complaints, refund negotiations, escalations) and automate the FAQ-level work.
What we built
Arthat built a WhatsApp support agent for the brand with the following capabilities:
Ticket classification: Every incoming message is classified in under 2 seconds — order query, product question, return/refund, delivery issue, complaint, upsell signal, general chat. Classification happens in the language the customer wrote in (English, Hindi, Marathi, occasional Hinglish).
Auto-response for high-confidence FAQs: Order status ("where's my order?"), policy questions ("what's your return policy?"), product info ("is this shade vegan?"), delivery timelines, discount code clarifications — all handled conversationally in the customer's language, with real order data pulled from Shopify.
Guided return flow: When a customer initiates a return, the agent walks them through — asks for reason, checks return window, arranges pickup via Shiprocket, creates the return in the admin, confirms refund timing. If the return is outside policy, the agent asks clarifying questions before escalating to a human.
Escalation with context: For anything complex (complaints, custom requests, refund disputes, product quality issues), the agent writes a summary of the conversation, extracts key details (order ID, SKUs involved, customer sentiment, what's been promised), and hands off to the next available human agent via Freshdesk.
Human-in-the-loop on money: Any action that moves more than ₹500 (refund, discount, replacement) requires human approval. The agent drafts the action and routes it; the human taps "approve" or adjusts.
Multi-language native handling: Hindi conversations stay in Hindi. Marathi stays in Marathi. The agent doesn't condescend by translating to English and back — it responds natively.
Tone calibration: The brand has a specific, playful, slightly irreverent tone. The agent was fine-tuned with 300+ approved examples of real brand responses. It sounds like the brand, not like a generic bot.
Arthat AI
Want results like these at your company?
We built this for them. We can build something similar — or something completely different — for you. Start with a 30-minute discovery call and we'll tell you honestly whether AI is the right tool for your situation.
