A bank's inbound calls, answered by a voice assistant in Croatian

Retail banks and NBFCs in India run phone desks that drown in repetitive requests: card activation, blocked-card replacement, wallet setup, and 'where is my statement' queries. Most of this still flows through a rigid touch-tone IVR that forces callers down menu trees designed for the bank's convenience, not the customer's. The operational question is whether a natural-language voice layer can absorb the predictable, low-judgement traffic without degrading the experience for anyone who genuinely needs a human.

The interesting part is not deflection for its own sake but what happens to the routing spine. Legacy IVRs punish callers who pick the wrong branch, and abandonment often hides the real cost. A voice assistant that understands intent could, in theory, shorten the path to the right queue. For an Indian BFSI desk, the design tension is language coverage and the sensitivity of anything touching card and account actions.

PolyAI reports that its Croatian-language inbound voice assistant, named Mia, was deployed at Zagrebačka banka (a UniCredit subsidiary) to handle card activation and replacement, wallet setup, and online-banking FAQs, with natural-language routing intended to replace the legacy IVR.

According to PolyAI, the assistant automates 27% of all calls and routes callers 83% faster than the previous IVR. PolyAI also states that call abandonment fell by 10% and that NPS rose by 14 points within the first six months. No publication date is stated on the source, so these figures carry no date and should be re-verified before use.

An Indian BFSI or NBFC desk should first ask what the 27% is measured against: is it 27% of total inbound, or 27% of eligible intents after filtering out calls the bot never attempts? Ask how card-action requests are authenticated, since activation and replacement touch sensitive instruments, and confirm whether the assistant handles the full task end-to-end or hands off after collecting details. Clarify the language scope: Croatian is a single-language deployment, whereas an Indian desk may need Hindi plus regional coverage.

Also inspect the honest metric. NPS moving 14 points matters more than deflection because a bot that deflects while frustrating callers simply defers cost to complaints and repeat calls. Ask how NPS was sampled, whether it was collected only from automated calls or all callers, and over what call volume. Confirm applicable rules, consent, retention and escalation design with qualified counsel or compliance owners for voice recordings, and check whether routing decisions are auditable when a customer disputes an action taken over the phone.

This is a large European retail bank inside a multinational group, operating in one primary language with call volumes and IT integration that most Indian SMBs and smaller NBFCs will not match. A 27% automation rate assumes a high concentration of standardised, repeatable intents; a smaller lender with messier account states and more human-mediated relationships may see a very different mix, so the headline percentage should not be treated as portable.

The single-language assumption is the sharpest break. Indian callers switch between languages and dialects mid-call, and card-action flows raise India-specific authentication and data-handling questions to confirm with qualified counsel or compliance owners rather than infer from a European setup. The reported routing and abandonment gains depend on replacing a specific legacy IVR; a desk without that baseline has nothing equivalent to improve against, so the same intervention could produce a smaller or differently-shaped effect.