Discovery Bank

Discovery Bank operates at the intersection of banking and behavioral science, running a shared-value model that rewards clients for measurable improvements in financial wellness — from meeting savings goals to purchasing healthy groceries. Translating that depth of individualized insight into real-time client interactions at scale was the central challenge. Their behavioral modeling engine produced rich next-best-action recommendations, but communicating them effectively through digital and human-agent channels required generative AI capabilities the bank had not yet deployed. Compounding the difficulty, initial AI response times of 5–6 seconds created noticeable friction — undermining the smooth, premium experience Discovery needed to differentiate itself in a competitive retail banking market.

Discovery built 'Discovery AI' by layering Azure OpenAI in Foundry Models on top of Azure Databricks, which serves as the bank's data lake and behavioral modeling platform. Rather than deploying a single general-purpose model, the team fine-tuned five separate Azure OpenAI 4o-mini and 4.1-mini models for distinct functions — consolidating what had previously required three processing steps into one, directly addressing the latency problem. The system powers two parallel surfaces: an agent assist tool that equips contact center staff with personalized prompts and next-best-action recommendations during live calls, and a client-facing interface accessible 24/7 via WhatsApp and the Discovery Bank app. The deployment supports multimodal inputs — text, images, and voice — enabling interactions such as photographing a grocery receipt to verify healthy food rewards in real time.

Client engagement with next-best-action recommendations doubled following the rollout, with 70% of all client interactions now flowing through the personalized recommendations engine. Response latency dropped from 5–6 seconds to 1.5–2 seconds on average — a reduction of more than 50% — making AI-powered conversations feel smooth. Contact center staff now handle approximately 3,000 AI-assisted questions daily, and satisfaction scores improved specifically in interactions where agents offered next-best actions versus pure service inquiries. Adoption accelerated sharply: traffic through Discovery AI nearly doubled within a single month.

• 2x client engagement with next-best actions
• 50%+ reduction in AI response latency
• 3,000 daily agent questions processed
• 70% of interactions routed through the recommendations engine

• Proprietary behavioral data is the real moat: combining a mature modeling stack with generative AI produces outcomes neither system achieves independently, and is hard for competitors to replicate without equivalent data depth.
• Fine-tuning multiple smaller, task-specific models outperforms a single large model on latency and accuracy — and lets teams systematically fix domain-specific failure modes (terminology gaps, inconsistent phrasing) iteration by iteration.
• Deploy agent-assist and client self-serve channels simultaneously; each reinforces adoption of the other and accelerates overall engagement growth.
• Multimodal inputs (text, image, voice) lower the barrier to engagement — natural interactions meaningfully increase client trust and usage frequency.