Priya Jaiswal stands as a formidable figure in the intersection of banking and advanced technology, bringing years of seasoned perspective to the rapidly evolving world of financial market analysis and portfolio management. As global markets grapple with the transition from artificial intelligence as a speculative tool to a core operational necessity, Jaiswal has become a key voice in defining what “return on AI” actually looks like for institutional giants. In an era where investors are increasingly skeptical of “token maxing” and runaway compute costs, her insights offer a roadmap for navigating the delicate balance between high-level innovation and the gritty reality of corporate P&L statements. This conversation explores how the banking sector is moving beyond the initial euphoria of generative AI toward a disciplined, federated model that prioritizes measurable value over mere technological novelty.
The following discussion explores the strategic shifts occurring within large-scale financial institutions as they transition from the experimental phase of AI adoption to full-scale production. Key themes include the implementation of a “four E” framework for internal adoption, the shift from frontier-only models to a balanced ecosystem of proprietary and open-source tools, and the rigorous filtering of use cases based on high-frequency impact. We also delve into the metrics of success, such as the dramatic reduction in merchant quoting times and the optimization of software development cycles, while addressing the critical challenge of “work slop” in probabilistic systems.
Many organizations struggle to move past the initial hype phase of technology adoption. How do you structure an internal framework that moves AI from a mere curiosity to a functional part of the corporate workflow?
The transition from a novelty to a necessity requires a very specific, disciplined operating model that I like to categorize into four distinct pillars: evangelizing, educating, enabling, and executing. In a corporate setting, the first hurdle is helping thousands upon thousands of builders and software engineers understand how the tools they use in their personal lives can be responsibly applied to sensitive financial data. Evangelizing isn’t just about showing off a demo; it’s about illustrating the value proposition so that an analyst or developer feels the “excitement” of what is possible while remaining anchored in corporate policy. Education follows closely behind, where we focus on the “how”—how to build, how to integrate, and how to use these probabilistic technologies without compromising our rigorous risk standards. Enabling is perhaps the most resource-intensive phase, as it involves providing the right platforms, upskilling developers, and ensuring that we aren’t just giving them more screens to look at, but rather collapsing interfaces to simplify their daily workflows. Finally, execution is where the rubber meets the road, shifting the focus from “what can we do?” to “what must we do?” to deliver value to customers and shareholders alike.
With the rise of massive frontier models from providers like OpenAI and Google, there is a growing debate about whether banks should rely on these external giants or build their own. How do you navigate the choice between using frontier models and developing proprietary ones?
We have found that the most resilient strategy is maintaining a balanced environment where we are not tethered to a single provider. While we certainly leverage the immense power of frontier models from OpenAI, Anthropic, and Google, we also heavily utilize open-source and proprietary models that we build in-house. The primary driver for building proprietary models is the protection of our highly confidential data, which is the secret sauce that makes these models significantly more effective for our specific needs. We also employ specialized tools like Mythos and Fable to hunt for vulnerabilities, ensuring that our internal ecosystem is as secure as it is innovative. By avoiding a “one-size-fits-all” approach, we can match the right model to the right task, ensuring we don’t use a massive, expensive frontier model for a job that a smaller, more specialized proprietary model could handle more efficiently.
There has been a noticeable shift in the industry away from “token maxing,” where the goal was simply to increase AI usage as much as possible. Why is this shift happening now, and how are you managing the associated costs?
The era of rewarding people simply for how much they use AI is effectively over, and we have never been proponents of token maxing for its own sake. This shift is driven by three primary factors: first, model providers have moved from flat enterprise licenses to per-use charging, forcing a natural re-evaluation of expenses. Second, the initial phase of wild experimentation is ending, and stakeholders are rightfully demanding to know what the actual return on investment looks like. Third, and perhaps most importantly, the people who were experimenting are now seeing where the real value lies and are getting serious about integrating AI into core business processes. We encourage exploration, but it must be in the service of our stakeholders—customers, shareholders, and employees—rather than just being technology for technology’s sake. We want to understand exactly what we are getting out of every interaction, ensuring that the “excitement” of the new capability is backed by a cold, hard look at the value it generates.
Can you share some concrete examples of where this technology is delivering measurable impact today, particularly in areas that were previously bogged down by manual processes?
One of our proudest achievements involves our merchant payment services and the Elavon technology we provide. Previously, when our sales force engaged with merchants, it could take several weeks to digest their existing spend data and provide a comparable quote and proposal. By integrating generative AI, vision models, and predictive analytics, we have transformed that multi-week ordeal into a near real-time experience where information is digested and a proposal is generated almost instantly. Another significant win is in the realm of software development, where we aren’t just generating code faster, but we are testing more of it to ensure higher quality. We track developer productivity through measures like cycle time and production incidents, allowing us to see a real dollar impact in how we onboard clients and maintain our digital infrastructure. These aren’t just theoretical improvements; they represent a fundamental shift in how quickly we can respond to the needs of the market.
When you look across a massive enterprise with diverse needs, how do you decide which AI use cases deserve funding and which ones should be discarded?
We use a very specific filter through a centralized governance body that segregates ideas into three buckets: growth drivers, cost savings, and intangible productivity improvements. We prioritize the first two buckets because they offer a clearer path to measurable ROI, such as improving the customer onboarding experience or reducing manual labor in high-volume areas. A key piece of guidance we give our teams is to look for high-frequency work; if a process is incredibly complex but only happens once a month, it probably isn’t a good candidate for AI compared to a simpler task that happens thousands of times a day. We also place a massive premium on reusability, ensuring that if one part of the organization develops a brilliant capability, it is federated across the entire enterprise rather than being reinvented in eight different silos. This disciplined approach ensures that we aren’t just chasing the latest trend, but are instead building a library of tools that can be leveraged at scale.
The term “AI work slop” has emerged to describe the time wasted fixing errors or hallucinations generated by these models. How do you prevent AI from becoming a net-negative for productivity due to these inaccuracies?
Minimizing “work slop” is all about the rigor you apply to the engineered solution before it ever reaches a team member. We don’t just “turn on” a model; we define exactly what subset of data it can act upon and what the specific intended outcome should be for that use case. Because these are probabilistic technologies, we have to build in significant guardrails and proactive testing to ensure we aren’t creating a situation where fixing the output takes longer than doing the work manually. This is especially true in risk processes and control testing, where accuracy is non-negotiable and the cost of an error is high. We are constantly matching the right model to the right dataset to ensure the output is useful from the start, rather than a “slop” of information that requires human intervention to be made functional.
What is the “disciplined approach” for moving a project from a successful pilot or experiment into enterprise-wide production?
The jump from experimentation to production is often where projects fail, so we timebox our experiments to ensure we aren’t stuck in a perpetual state of “innovating” without delivering. From the very moment we select a use case, we are already designing what the production version will look like, ensuring we have the right skillset and technological infrastructure to scale it. We have a defined, rigorous process that acts as a gatekeeper, and we are very disciplined about killing experiments that do not work as intended or fail to meet our baseline metrics. This rigor upfront means that when something does move toward production, it has already been battle-tested and its value has been clearly articulated. We’ve found that this level of discipline actually allows us to see more things scale successfully because the teams are focused on the “how” of production from day one.
There is a lot of talk about an “AI bubble” and whether current valuations are sustainable. How do you view the current market euphoria compared to the actual utility of the technology?
It is vital that we separate value from valuation, as the two often get conflated during periods of technological breakthrough. There is a tremendous, undeniable value in the capability of AI itself—value for individuals, corporations, and society at large—but the euphoria of the market often runs ahead of the reality of adoption. History shows that with every major breakthrough, we get ahead of ourselves in the short term, thinking it will change the world overnight, while underestimating the difficulty of adopting it at scale in a corporate setting. Adopting this technology requires significant effort, from upskilling employees to restructuring data, and that level of difficulty is often underestimated by those focused purely on stock prices. Whether or not the valuations are in line with reality is for the analysts to decide, but from an operational standpoint, the utility is real, and the focus must remain on the long-term integration rather than the short-term noise.
What is your forecast for the role of AI in the financial sector over the next five years?
I believe we are entering a phase of “invisible AI,” where the technology will move from being a standalone topic of discussion to being the foundational plumbing of every financial transaction. In the next five years, the “four E” framework will have matured into a standard operating procedure across the industry, leading to a much more federated landscape where every department—from HR to high-frequency trading—will have its own specialized, highly efficient models. We will see a dramatic collapse in the time it takes to perform complex risk assessments and customer service interactions, moving from days or weeks to instantaneous, real-time responses. The real winners will be the organizations that didn’t just chase the hype, but instead spent this time building the disciplined governance and reusability structures that allow AI to function at scale without constant human intervention. Ultimately, the success of AI in banking won’t be measured by the complexity of the models, but by how seamlessly it disappears into the background of a better, faster, and more secure customer experience.
