This article examines how the success of AI investments in financial services increasingly depends on strong governance, trust, and risk management frameworks. Featuring insights from Joe Breeden, the piece explores how generative AI challenges traditiona
Artificial intelligence promises transformative returns for financial institutions, but without robust risk management, those gains may prove fragile.
In the second of a three-part series for Connect Magazine, Joe Breeden, CEO at Deep Future Analytics, argues that protecting the return on investment in AI depends not on innovation alone, but on building trust, governance, and control frameworks that can withstand real-world complexity.
The scale of adoption is already clear. Financial services ranks among the most active users of AI, reflecting its potential to drive efficiency and enhance decision-making.
As Breeden explains, however, rapid deployment has outpaced the industry’s ability to manage risk effectively.
“AI risk management is not a compliance burden - it enables positive ROI,” he says, emphasizing that without proper controls, projected gains can quickly be eroded by failures, regulatory breaches, or reputational damage.
Traditional model risk management frameworks provided a strong foundation in the past. They were built around deterministic systems, predictable behavior, and well-defined validation
processes.
According to Breeden, these principles worked because models operated within known boundaries, supported by clear metrics and accountability structures. But generative AI fundamentally changes that equation.
“LLM outputs vary even with identical inputs,” Breeden notes, highlighting the breakdown of determinism that underpinned earlier frameworks.
Unlike traditional models, large language models operate across effectively limitless input spaces, making comprehensive validation impossible. As outlined in the presentation, the number of potential prompts is so vast that pre-deployment testing cannot anticipate all failure modes.
This creates a new category of risk. Failures are not confined to technical errors but extend to ethical, regulatory, and operational domains.
Breeden points to examples where AI systems have crossed from helpful assistant to unintended accomplice, subtly enabling behavior that contravenes policy.
“The difference is subtle - not something keyword searches or vendor guardrails would catch,” he explains.
The implications for return on investment are significant. AI initiatives are often justified on efficiency gains or revenue growth, but these projections rarely account for downside risk.
Breeden’s framework reframes the issue by quantifying expected losses from model failure, regulatory non-compliance, or cybersecurity breaches, and then measuring how controls reduce those losses. The result is a clearer picture of
risk adjusted ROI.
