TD Bank

TD Bank, one of Canada's largest financial institutions, faced a fundamental scalability problem with its AI infrastructure. Traditional predictive models were purpose-built for single tasks, meaning engineers had to construct an entirely new model from scratch each time a new use case emerged across the bank's many business units. This one-model-per-task architecture made AI deployment slow, expensive, and difficult to scale. As demand for personalized client experiences grew across retail banking — where understanding individual financial needs is central to competitive differentiation — the inability to generalize AI capabilities became a significant operational bottleneck that constrained innovation bank-wide.

TD's response was to build AI Prism, a foundational predictive model developed by its in-house AI research lab, Layer 6 (acquired in 2018). Rather than optimizing for a single task, AI Prism was architected as a general-purpose predictive engine capable of being fine-tuned and redeployed across multiple business units — from personal banking to wealth management. The model uses large language models and generative AI techniques to process far greater data volumes than legacy systems, ingesting over 100 times more data per prediction cycle. TD centralized its data on a secure cloud platform to enable rapid model development and deployment at scale. Pilots ran across contact centers, branches, TD Wealth, TD Insurance, and TD Securities through 2025, validating the foundational model strategy before broader rollout.

AI Prism demonstrated measurable performance gains over legacy predictive systems during its 2025 pilots:

• 20–30% improvement in identifying client needs compared to traditional predictive models
• 100x more data ingested per model run versus legacy architectures
• 12% reduction in contact center call transfers, attributed to Gen AI virtual assistants operating alongside AI Prism insights
• $1 billion in total AI value targeted for 2026, with AI Prism as a core contributor

Beyond the numbers, the foundational model approach fundamentally changed how TD's engineering teams work — shifting from bespoke model construction to fine-tuning a shared capability, compressing development timelines across the organization.

• Build once, deploy many: Foundational models eliminate the redundant engineering cost of single-task AI, making each new use case faster and cheaper to launch.
• Data centralization is a prerequisite: Consolidating data on a unified cloud platform was what made AI Prism's scale possible — without it, the model's data advantage couldn't be realized.
• Pilot before scaling: TD's disciplined 2025 pilot program across multiple business units validated the strategy before committing to bank-wide deployment.
• Governance must scale with capability: TD's dedicated Trustworthy AI team embeds fairness, privacy, and accountability checks into every model — a non-negotiable as AI touches more client interactions.
• Internal AI research labs compound over time: The Layer 6 acquisition in 2018 seeded the capabilities that produced AI Prism — a reminder that foundational AI investment has a long lead time.