How to scale governed AI personalization in wealth management

The team also helped the development of a structured evaluation framework, testing tools and model options through phased proofs of concept and controlled experimentation. The evaluation moved from workshops and controlled test environments to hands-on experimentation and defined review points, helping the teams assess what could scale before committing to a production approach. Clear handover steps were also incorporated so internal teams could continue development and iteration without having to restart the process for each use case. The team also helped prepare curated training data and refine models to reflect real-world policy and domain requirements, including the distinction between general information and personalized financial advice. Where technical or compliance constraints emerged, including limitations on reusing training data across tools, the approach was adapted to maintain compliance while keeping development progress measurable.
 

Governance and monitoring capabilities were embedded throughout the development lifecycle to provide visibility into model behavior while protecting sensitive data. In parallel, EY teams helped establish a more self-service governance framework that enabled domain teams to fine-tune and deploy AI capabilities within defined governance, risk and cost controls. Taken together, these elements created a reusable operating pattern that connected model selection, testing, governance, deployment and ongoing refinement, rather than treating each AI use case as a separate technology experiment. This approach reflects the ey.ai vision of helping organizations move beyond isolated AI pilots by combining technology, trusted data, governance and domain experience to create sustainable business value.

The engagement established a scalable foundation for responsible AI adoption across wealth management and created a repeatable operating model for future AI deployments. Early pilot results demonstrated strong leading indicators across model quality, latency and governance. While the work remained at the pilot stage, the results provided evidence that specialized models could improve both technical performance and the practical operation of AI controls in high-sensitivity workflows. This gave the client a clearer basis for deciding where and how to expand AI adoption responsibly.
 

The governed SLM evaluation and refinement process established by the project team enabled the client to define clear performance measures for targeted AI use cases. Once evaluation criteria and success metrics were established, a focused set of models was tested and refined using client-specific data within the client's controlled technology environment. The result was a set of proprietary models designed to meet the organization's requirements for performance, responsiveness, cost efficiency and data privacy, while remaining aligned to governance and compliance expectations.
 

In financial advice detection use cases, specialized SLMs demonstrated stronger performance than both legacy rules-based approaches and general-purpose model baselines on policy-aligned test scenarios, improving the accuracy of classifications tied to regulatory requirements. The improved classifications were especially relevant for nuanced policy definitions where generic models tended to over- or under-classify. Operationally, reducing false positives can help avoid unnecessary friction in review processes, while reducing false negatives can strengthen the identification of potential compliance exposure. Performance also improved materially, with early testing showing up to 10x lower latency than prior large language model (LLM)-centric approaches. That improvement supported more responsive client-facing interactions and more efficient enforcement of AI guardrails within the request process.