BRD: 85% of banking questions auto-resolved, 24/7

BRD — one of Romania's largest retail banks — was fielding a high and growing volume of repetitive customer questions across phone and digital channels. Product eligibility, fees, card operations, branch procedures, document requirements: the same questions, asked thousands of times, were saturating the call centre, pushing up wait times and pulling senior agents away from the cases that genuinely needed a human.

The bank wanted to deflect that repetitive load without degrading the customer experience — and without the robotic, dead-end feel of a traditional IVR or rules-based bot. The bar was high:

• Accurate, current answers about real banking products and procedures — not hallucinated ones
• Natural Romanian, including the informal phrasing customers actually use
• Sub-2-second responses, 24/7, at call-centre-scale concurrency
• A clean, context-preserving handoff to a human agent the moment a question exceeds the bot's competence
• Auditability and safety controls appropriate for a regulated financial institution
• Integration with the bank's existing systems rather than a siloed widget

We built a production assistant on OpenAI's GPT-4, grounded in BRD's own knowledge so it answers from the bank's reality rather than the model's general training. Rather than fine-tuning on static data, we used a retrieval-grounded architecture: the assistant retrieves the relevant, current product and procedure information at question time and answers strictly from it — which keeps answers accurate as products and fees change, and makes every answer traceable to a source.

Trained on real conversations. We analysed 50,000+ real BRD customer conversations to understand intent distribution, the vocabulary customers actually use, and the long tail of edge cases — then shaped retrieval, prompting and handoff thresholds around what customers really ask, not what a product manual assumes they ask.

Romanian-first NLP. Romanian retail-banking language is full of diacritic-free typing, anglicisms and informal phrasing. We tuned intent handling and retrieval for that reality, so an informally typed request is understood as readily as a formal one.

Safety and guardrails for a regulated bank. The assistant is scoped to answer only what it can ground in approved content; out-of-scope, sensitive or account-specific requests are routed to a human rather than guessed at. Every exchange is logged for audit, and senior engineers — not the model — own the prompts, the retrieval sources and the evaluation suite that gates changes before they reach production.

Latency budget. To hit a sub-2-second feel at scale, we put a Redis caching layer in front of common intents and retrieval results and streamed responses, so the customer sees an answer forming immediately instead of waiting on a full generation.

Human handoff, done right. When confidence drops or the topic is account-specific, the conversation — with its full context — is handed to a human agent, so the customer never repeats themselves and the agent starts informed.

Stack: OpenAI GPT-4, Python, FastAPI, Redis, a retrieval/NLP pipeline, and integrations with the bank's existing channels and APIs.

The assistant now resolves 85% of incoming questions automatically, answers in under 2 seconds, and is available 24/7 — reaching 96% customer satisfaction on its interactions.

The operational impact compounds from there: the call centre is no longer saturated by repetitive, low-complexity questions, so human agents spend their time on the cases that actually need judgement. Customers get instant, accurate answers at any hour instead of queueing. And because the assistant answers from grounded, auditable content with a clear escalation path, the bank gets deflection without taking on the risk of an unsupervised bot improvising about financial products.

Just as important for a regulated institution: senior engineers own the quality, safety and evaluation loop end to end. AI is a capability inside a governed product — not a black box bolted onto the contact centre.