In the latest roundtable hosted by the CFA Institute, industry specialists examined how agentic AI is reshaping investment workflows through new skill standards and advanced data models, while issuing a cautionary note that investors must remain vigilant about inherent behavioral biases—such as loss aversion—embedded within AI systems.

The conversation, moderated by Rhodri Preece, Senior Head of Research at the CFA Institute, brought together Brian Pisaneschi, Head of Applied Investment Practices and Tools, and James Tate, a Data Science Research Fellow. Together, they dissected the construction pathways of agentic AI within investment operations, explored novel solutions for addressing gaps in financial data, and uncovered often-overlooked bias blind spots associated with model selection and governance.

Agentic AI Transforming Workflows: A Shift From Code-Driven to Skill-Driven Approaches

Historically, financial professionals required advanced coding expertise to leverage AI. Today, the barriers to building agentic workflows are rapidly diminishing, as evolving industry standards are elevating AI from a mere support tool to a central operational hub.

Brian Pisaneschi observed that just a year ago, constructing an agent required proficiency in various Python libraries. Now, the industry has transitioned to using Markdown files—semi-structured text—to articulate specific workflows, known as "Skills."

Even more transformative is the proliferation of MCP (Model Context Protocol) servers. Pisaneschi emphasized that MCP servers have become an industry benchmark, explaining that their primary function is to enable large language models to interact with external data APIs in an almost native capacity.

This evolution is fundamentally altering daily habits for finance professionals. Previously, an analyst would log into a Bloomberg terminal to extract data and then import it into Excel. Now, the logic is inverted. Pisaneschi noted that the primary method of task execution now begins with interacting directly with an AI platform. Once all necessary data sources are connected, the AI can autonomously determine what is relevant, extract the appropriate information, and launch the corresponding dialogue and workflow.

The elegance of this model lies in its simplicity and iterative nature. Analysts can continuously correct errors and append revised information to skill files. Pisaneschi highlighted that even without improvements to the underlying model, one can build remarkably powerful workflows purely through accumulated knowledge and iterative refinement within these task-specific skill files.

Overcoming Historical Data Bottlenecks: How Synthetic Data Is Revolutionizing Stress Testing

Financial market datasets are frequently incomplete, and AI is offering novel approaches to fill these voids. Beyond traditional methods like linear interpolation and k-nearest neighbors imputation, James Tate pointed to the commercially significant potential of generative AI in producing "synthetic data."

This approach holds substantial practical value in portfolio management. Tate noted its applicability in scenario simulation, portfolio stress testing, and backtesting trading strategies. Traditional stress tests typically depend on historical empirical distributions, whereas generative AI operates without preset linear or distributional assumptions. This AI-driven methodology does not require predefining a functional form or assuming a given distribution; instead, it adopts a more empirical approach, theoretically capable of modeling any potential relationship between variables. Analysts can input macroeconomic factors such as inflation, yields, or GDP, allowing the AI to extrapolate possible portfolio trajectories through data-driven reasoning, offering significant flexibility for adapting to shifting market dynamics.

Open-Source Versus Proprietary Models: Navigating Cost Optimization and Winner-Takes-All Dynamics

With a plethora of AI models available, financial institutions face complex decisions regarding selection, balancing performance against data privacy and cost considerations.

Concerns over privacy often drive institutions toward open-source solutions. Pisaneschi acknowledged that open-source large language models may lag behind mainstream proprietary offerings, such as those from OpenAI, by roughly three months. In the proprietary domain, a winner-takes-all dynamic prevails, as leading entities achieve significant parallelization and optimize their tooling frameworks specifically for agentic tasks. Tate added that proprietary models boast overwhelming advantages in scalability, allowing users to rapidly expand operations by sending 50,000 batch requests via API and receiving results within 24 hours.

However, the core commercial value of open-source software lies in cost optimization. Pisaneschi pointed out that processing massive datasets with frequent calls to OpenAI's API could become prohibitively expensive. Consequently, institutions can train and run small language models (SLMs) locally on high-quality datasets. If you can obtain a sufficiently capable model, running it locally on a small language model can substantially reduce costs, making operations almost free.

Addressing a Critical Misconception: AI Does Not Eliminate Bias—It May Even Exhibit Loss Aversion

In governance and compliance, a common fallacy persists that delegating decisions to machines removes human emotional biases. This roundtable presented a starkly contrasting conclusion.

Tate issued a strong warning regarding bias within large language models. He noted that extensive research indicates these models largely mirror investor biases because they are trained on human-generated data and naturally reflect those human tendencies. He highlighted a particularly striking finding: studies reveal that large language models exhibit loss aversion bias. This means if tasked with managing a portfolio containing both winning and losing positions, the model is more likely to hold onto the losing portions, rather than reallocating those funds toward more prudent investments. Furthermore, due to training data composition, these models demonstrate a significant Western-centric bias and a preference for US technology stocks.

Building Trust and Mitigating Bias in AI-Driven Investment Environments

Preece raised the pertinent issue of trust and governance in an AI-dominated world, asking how firms can responsibly utilize tools and datasets while protecting client confidentiality and fulfilling fiduciary duties. Both panelists agreed that the path to trust is rooted in control and iterative validation.

Pisaneschi drew parallels to quantitative investing, where basic tools like stop-loss orders help mitigate emotional biases. He argued that by embedding operations within rulebooks and investment strategies, one can significantly reduce emotional interference. Yet, large language models are complex entities with decision-making authority that is susceptible to bias. The remedy lies in techniques like red teaming, borrowed from cybersecurity, where systems are probed for vulnerabilities. By red teaming models for specific biases and creating skill files to mitigate them, practitioners can instill greater predictability and control over AI outputs.

Tate reinforced the importance of benchmarking different models. He suggested using a custom-developed model as a baseline, testing it across numerous workflows, and then running the same processes through free open-source alternatives to compare outcomes. He shared his experience of nearly replicating the performance of a proprietary OpenAI model on a specific task using a much smaller open-source model. However, he noted that scalability remains a challenge, as running even small language models locally requires substantial infrastructure investment, making proprietary API solutions more practical for many firms.

The discussion concluded with a consensus that the ultimate goal of all governance and ethical considerations in AI is earning trust. As investment professionals, overseeing AI outputs with domain expertise—evaluating, correcting, and iterating—mirrors how one would manage and mentor a human employee over time. This process of continuous assessment and adjustment is how trust is scaled across an entire organization in an AI-centric world.

For those interested in diving deeper, the CFA Institute Research and Policy Center offers a wealth of insights, forward-looking analysis, and research findings on these topics.