4 Key Takeaways from our Workshop at HCII2024

Four design principles from Kensho's HCII2024 workshop on building human-centered AI applications for financial services.

By Houssam Kherraz, Yuanfei Zhao, and Shirley Anderson

During our recent workshop at the Human-Computer Interaction International Conference (HCII2024), we delved into the intersection of the evolving landscape of AI models, human-centric design, and the growing field of financial AI. Here are four key takeaways from our workshop, offering insights into how these areas shape designing AI applications.

1. Navigating the spectrum of AI models

We explored the spectrum of AI models from classical models like multiple linear regression and decision trees to modern, complex systems such as neural networks and Large Language Models (LLMs).

Classical models: These models are favored in finance for their statistical backing and explainability. They provide clear insights into which factors contribute most to decisions, making them invaluable for a range of tasks.

Generative AI models: These models, including LLMs, are powerful but less interpretable. While they offer unprecedented capabilities, their complexity makes them challenging to explain.

A key takeaway to consider is the trade-off between power and interpretability. As models become more sophisticated, they lose the clear factor attribution that makes classical models so interpretable. This lack of interpretability poses significant challenges, especially when dealing with the phenomenon of AI hallucinations.

2. Emphasizing human-centric AI design

Human-centered AI is the approach that advocates putting people first and valuing meaningful human control when designing for AI products and services. It emphasizes incorporating users’ considerations and constraints into the early stages of algorithm design. It also advocates evaluating not only the algorithm accuracy, but also the acceptability, user experience and the social impact of the AI system. Human-centered thinking is crucial in the development of most technology products, but it holds even greater significance for AI products.

Avoiding the “hammer and nail” trap: When designing AI models, it is vital not to become fixated on the technology as a solution in search of a problem. Instead, the focus should be on addressing user needs to prevent developing solutions that fail to tackle real issues.

Specific vs. generalized solutions: While AI models are often designed for broad applications, users have specific needs. For example, document summarization for meeting notes will differ significantly from a summary for a company’s 8-K report. Understanding the specific needs of users can greatly enhance the model’s effectiveness and usefulness.

AI as part of the experience: AI cannot address all the tasks users need to complete. It is one component within a larger process. As product designers, our goal is to map out the entire user journey, identify pain points in their workflow, and create innovative solutions that effectively address their needs.

3. Driving financial AI innovation

NVIDIA’s 2024 report reveals that 91% of financial companies are either using or evaluating AI. Of these, 55% are exploring generative AI (GenAI) and LLMs. This technological leap has already enhanced operational efficiency for 43% of these firms, and 97% are planning to ramp up their AI investments soon. These statistics underscore a pivotal shift, where AI is becoming indispensable across various financial functions — from fraud detection to customer service.

Start with the customer problem: To create successful financial AI products, focus on understanding the specific challenges your customers face. Avoid chasing the latest tech trends and focus on solving real problems.
Understand genAI: Dive deep into GenAI from both a technical and market perspective. Knowing its capabilities, risks, and realities will help you leverage its potential effectively.
Jobs to be done (JTBD) framework: Use the JTBD framework to create a minimum desirable product. This approach grounds your thinking and ensures you’re building a product that truly meets customer needs. Constructing a JTBD statement involves defining the context, identifying barriers, clarifying the assistance needed, and stating the desired outcome. Example of a JTBD Statement: “I’m considering buying a home (context), but I’m unsure how much I can afford (barrier). Help me calculate my budget and potential mortgage payments (goal), so I can make a realistic and financially sound decision (outcome).”

4. Achieving product-market fit

Product-market fit goes beyond having a unique product. It’s about ensuring your product effectively addresses customer needs in a compelling way. This sweet spot lies at the intersection of AI trends, technological feasibility, and market context. The case studies below illustrate how banks are achieving product-market fit and successfully using AI to address critical issues in finance.

1. Barclays: Fraud Detection: Barclays employs AI to monitor merchant payment transactions in real-time, predicting and preventing potential fraud.

2. Bank of America: AI in Research Analysis: Their platform, Glass, helps sales and trading employees uncover hidden market patterns, consolidating data across asset classes and regions with in-house models and machine learning.

Conclusion

Building impactful financial AI products requires a deep understanding of the customer needs and a thoughtful approach to integrating AI into existing workflows. By evaluating AI models, applying human-centered design principles, and prioritizing solving real user problems, it is possible to develop solutions that have a significant and meaningful impact. Remember, AI is part of the experience and not the entire experience.