Redefining personal banking with Discovery Bank and Azure OpenAI

Behavioral modeling + generative AI = transformative financial services

Over three decades ago, Discovery was founded on a clear purpose: To make people healthier and to enhance and protect their lives. Guided by this principle, Discovery has grown into a global organization spanning banking, investments, and a comprehensive range of insurance solutions across health, life, and motor—designed to meet clients’ evolving needs. Supporting clients across these areas enabled Discovery to pioneer its shared-value model, which offers them a range of pathways and incentives to change their behavior to decrease their personal risk. For instance, Discovery Bank clients can earn meaningful rewards for purchasing healthy food items and reaching financial savings goals, ultimately shaping their actions over time to improve their long-term financial and personal wellbeing. 

This ability to personalize the experience for its customers and encourage specific behaviors at the individual level is what sets Discovery apart. It also requires a powerful AI stack operating behind the scenes. Using Azure Open AI in Foundry Models, Discovery makes it happen. “We view the future of financial services as rewarding financial wellness,” says Stuart Emslie, Head of Actuarial and Data Science at Discovery Bank. “We’ve developed extensive behavioral modeling to create a client’s profile that helps determine the actions we should be recommending to our clients.” Aiming to evolve the experience further, the Discovery team created Discovery AI, which is a generative AI solution that clients can interact with through their preferred channels. “That’s really where we see our competitive advantage—being able to combine our behavioral stack with generative AI,” says Emslie. 

Azure Databricks at the core of hyper-personalization

So much of Discovery’s engagement with its clients relies on hyper-personalization. On the backend, a robust AI engine powers unique client profiles based on financial indicators, as well as behaviors like spending patterns and engagement. But being able to scale that level of machine learning and behavioral modeling was a challenge. “We didn’t want to worry about platform management, and we needed the latest technology to drive the personalization our clients expect,” says Emslie. “Azure Databricks was by far the best platform we evaluated because it gives us the agility and flexibility we need without the management overhead, allowing us to build data products quickly.” Serving as Discovery’s data lake, Azure Databricks drives the full lifecycle of the client experience—from targeting via digital advertising to optimizing financial health. 

Azure OpenAI in Foundry Models powers ‘a private banker in every pocket’ 

The next step for the Discovery team was to use Azure OpenAI in Foundry Models to layer generative AI on top of its behavioral modeling. “We shifted our focus to better communicating our modeled next-best actions based on our personalized client data and behavioral profiles,” says Emslie. This gave way to an agent assist tool as well as personalized interactions through the Discovery Bank app. Now, when a client connects with a human agent, Discovery AI pulls from a knowledge base and client profiles to generate personalized responses and prompt next-best actions. This allows human agents to not only solve banking issues quickly, but also offer ways to keep clients on their journey to financial health. So far, clients have given higher ratings to the interactions where agents have offered next-best actions, as opposed to interactions where they only dealt with a service inquiry. 

At the same time, Discovery Bank clients can access generative AI through WhatsApp 24/7. “Discovery AI gives every single one of our clients a private banker in their pocket,” says Emslie. “They can ask questions, receive personalized recommendations, and even perform actions like setting budget reminders.” Whether asking about their forecasted spending, requesting a tax document, or viewing their top spending categories, Discovery AI provides clients with tailored answers through WhatsApp or in the Discovery Bank app.

Reducing latency with fine-tuned Azure OpenAI 4o-mini models

To achieve this level of precise personalization, the team fine-tuned five separate versions of Azure OpenAI 4o-mini models and Azure OpenAI 4.1-mini models for different functions. Instead of using a single model, fine-tuning multiple smaller models allowed Discovery to maintain performance while improving accuracy and reducing latency. “We consolidated three steps into one,” says Emslie. “Response times that were previously five or six seconds came down to one and a half to two seconds on average.” This approach made the system more efficient and the 50% reduction in latency made conversations with Discovery AI feel seamless.

Each iteration of fine-tuning also targeted specific model failures that could impact the customer experience, from inconsistent phrasing to misunderstanding company-specific financial terms. “It’s about getting that last 5% of accuracy right,” says Dean Bunce, Head of Data Science at Discovery Bank. “We focused on fixing very specific pain points, ensuring the model generates the right SQL query format or interprets custom templates correctly. Once those were resolved, the entire system became more reliable.”

Beyond text interactions, Discovery AI’s multimodal capabilities extend to documents as well. Clients can upload photos, voice notes, or documents and Discovery AI uses Azure Open AI in Foundry Models to interpret the information and respond intelligently. For example, customers might snap a photo of a grocery receipt to confirm whether it qualifies for healthy food rewards. “Through text, images, or voice, multimodal experiences transform how clients engage with us—creating interactions that feel natural, intelligent, and human,” says Abdullah Esmael, Senior ML Engineer at Discovery Bank. “Right now, our Azure OpenAI models are flexible enough to power these multimodal experiences and as we move toward more critical use cases such as KYC, Azure AI Document Intelligence and Azure OpenAI will form a key part of our generative AI stack.”  

Azure-based architecture supports exponential growth for Discovery AI

Both the agent assist tool and the Discovery AI application run on Azure Container Apps and Azure App Service. Initially, the team hosted internal AI tools on Azure App Service, later migrating to Azure Container Apps for greater flexibility and scalability. Lightweight containers now host APIs, engine logic, and orchestration layers that connect Discovery Bank’s contact center, Azure Databricks, and Azure OpenAI endpoints. 

At the front door of this architecture sits Azure API Management, acting as the secure gateway for every AI request. It standardizes and monitors traffic across Discovery’s internal and external applications while ensuring consistency, governance, and security. “API Management is the entry point to our entire AI world,” says Adit Mehta, Head of MLOps at Discovery Bank. “It allows us to handle security, throttling, and monitoring from a central point so we can move fast without losing control.”

For data storage, Discovery Bank uses Azure Cosmos DB to manage unstructured chat data and context from ongoing client interactions. Azure Cosmos DB offers a flexible schema and high-speed read/write capabilities, making it ideal for storing conversation history and contextual data across thousands of simultaneous sessions. In addition, Event Hubs and Azure Data Factory handle data streaming and ETL workflows between on-premises systems and the cloud. Real-time transactions, event data, and updates flow into Azure environments, ensuring every recommendation and model decision reflects the most current information. Finally, Discovery maintains security and compliance through Azure Key Vault, which safeguards all service credentials, certificates, and secrets across the AI stack. Meanwhile, Azure DevOps powers version control and continuous integration/continuous deployment (CI/CD) to automate and manage model releases. 

As the Discovery Bank team built out their sophisticated architecture, their partnership with Microsoft allowed them to move fast and troubleshoot. “The support we’ve received from Microsoft has been phenomenal,” says Mehta. “Anything we needed help with, they were more than happy to assist. They’ve been on speed dial for us.”

What started as an application to drive AI-driven personalization has evolved into a cornerstone of Discovery Bank’s client engagement. Discovery Bank’s servicing agents now process roughly 3,000 questions daily, with 70% of clients engaging with personalized recommendations. “In the past month alone, we’ve seen traffic through Discovery AI nearly double,” says Emslie. “It’s going to see exponential growth as more clients use it.” By pairing behavioral modeling with generative AI, Discovery Bank can expand its mission to help people lead healthier, more financially empowered lives.

Learn more about Discovery on LinkedIn and YouTube.