Tonik Bank

Tonik Bank, a digital-only neobank serving 1.5 million customers in the Philippines, faced a support crisis driven by its own growth. Customer interactions surged 2.5X as the bank rapidly expanded its product lineup, overwhelming an existing NLP-based chatbot that could not keep pace. The core limitation was structural: every product update, policy change, or new feature required manual retraining cycles, creating a persistent lag between what the bank offered and what the bot could accurately explain. In the neobank sector — where speed-to-market and customer trust are existential — this gap eroded service quality and forced the bank to choose between costly headcount expansion or degraded support experiences.

Tonik partnered with Gupshup to replace the static NLP system with a Generative AI architecture built around Gupshup's ACE LLM, a fine-tuned large language model designed for financial services conversational use cases. Rather than replacing the existing NLP layer entirely, the deployment used a hybrid multi-model approach — combining the ACE LLM with traditional NLP models to preserve intent-classification strengths while adding generative reasoning for open-ended queries. The critical architectural decision was grounding AI responses directly in live data: the system continuously reads from Tonik's website and internal policy documents, enabling self-updating knowledge without manual retraining. The solution integrated into Tonik's existing in-app chat channel, ensuring customers experienced no friction in the transition while the AI layer operated transparently beneath the surface.

The Gen AI deployment delivered measurable gains across resolution rate, accuracy, and operational efficiency:

• 75% of all customer queries now resolve autonomously end-to-end
• 95% AI accuracy rate, establishing a reliability baseline sufficient for financial services compliance
• 4.3X increase in customer care agent productivity — agents handle more complex escalations rather than repetitive FAQs
• 9 in 10 queries are contained within the in-app chat channel, reducing call center volume
• $20 million in projected cost savings over three years
• Headcount growth held below 20% despite 2.5X growth in customer interaction volume

The productivity multiplier is particularly significant: agents were redeployed to higher-value interactions rather than eliminated, aligning cost control with service quality.

• Hybrid LLM + NLP architectures offer a lower-risk upgrade path for teams with existing intent-classification infrastructure — don't discard what works.
• Grounding generative AI in live documents (not static training data) solves the retraining bottleneck that makes traditional chatbots expensive to maintain in fast-moving product environments.
• For neobanks and digital-first financial services, autonomous resolution rate matters more than raw deflection — 75% end-to-end resolution is a different standard than 75% containment.
• Productivity multipliers (4.3X) often unlock more long-term value than headcount reduction alone; frame AI ROI around agent capacity reallocation, not just cost cuts.