A multilingual AI relationship executive calling every customer after purchase
An outbound AI agent that calls recent customers of a ten-branch jewellery chain, collects genuine feedback on their purchase, then introduces the gold savings schemes most relevant to them — in English, Telugu or Hindi.
Client: Mukunda Jewellers · Hyderabad, Telangana · Status: live
The challenge
A jewellery chain with ten branches had no consistent way to follow up after a purchase. Feedback was collected ad hoc at the counter, and the savings schemes that drive repeat business were explained only when a customer happened to ask.
What we built
We built an outbound voice agent that runs a two-part call: first a short, genuine feedback conversation about the recent purchase, then a tailored introduction to whichever savings scheme fits that customer. The agent knows the full branch network, the exchange and purity policies, and all three schemes including the rules that must never be misstated, and it can query a knowledge corpus mid-call when a customer asks something specific.
Engineering highlights
- Detects the customer’s language from their first reply and switches fully to English, Telugu or Hindi — and asks permission before switching rather than assuming.
- Hard limit of one to two sentences per turn, so the call never turns into a monologue about schemes.
- Scheme rules that carry financial consequence — redemption windows, what the funds may be spent on, instalment terms — are pinned in the prompt so they cannot be improvised.
- A retrieval tool answers detailed questions mid-call, with the instruction to synthesise one natural sentence rather than read retrieved text aloud.
- Built twice, on Pipecat and on Ultravox, so the deployment is not locked to a single voice stack.
- Silence handling: one warm check-in, then a polite close, rather than leaving a customer on an open line.
Technology
Services used
Want something like this?
Tell us your problem and we'll tell you exactly how AI solves it and what it costs.