Context

Artificial intelligence (AI) systems are deployed in a myriad of contexts, each with unique requirements and evaluation methodologies. This diversity often results in disparate evaluation metrics, leading to ‘apples-to-oranges’ comparisons that hinder effective assessment and improvement of AI systems. The need for a consistent and coherent evaluation framework grows as AI systems become more integral to critical sectors like finance, healthcare, and security. Transparent, rigorous evaluation paradigms can support stakeholders in understanding AI system performance relative to their specific contexts, fostering trust and enabling safer, more reliable deployments.

The complexity of AI systems further necessitates an approach that factors in human impacts and operational realities. Engaging subject matter experts (SMEs) in the evaluation process ensures that these systems are assessed against realistic scenarios reflecting real-world applications, challenges, and opportunities.

The Research

The research presented by Choong et al. aims to bridge the gap in AI evaluation by advocating for methodological transparency, operational grounding, and the application of human-centered design (HCD) principles. The authors propose a structured process that translates high-level use cases into detailed evaluation scenarios using inputs from subject matter experts. This approach is exemplified in the financial services sector, showcasing key AI applications such as cyber defence enablement and credit memo generation.

The methodological framework is built around an ‘AI Use Case Worksheet,’ which captures six essential elements: use case, sector, users (direct and indirect), intended outcomes, expected impacts (positive and negative), and KPIs and metrics. This structured approach ensures that evaluations are grounded in sector-specific realities and that they reflect the needs and contexts of diverse stakeholders involved.

Key Finding

Central to the study is a repeatable, three-stage expansion pipeline that integrates large language model (LLM) prompting with iterative human reviews. This process generated 107 scenarios from the initial AI use cases identified by SMEs in the financial services sector. The pipeline is instrumental in maintaining operational grounding by integrating human feedback at every stage — from defining scenario titles and descriptions to articulating core elements such as users, benefits, risks, and evaluation objectives.

The iterative human review process serves as a checkpoint, ensuring that each scenario remains reflective of realistic usage contexts. This approach not only bridges the gap between abstract AI functionalities and practical applications but also embeds human-centric values into the evaluation process by considering the potential impacts and interactions with human users.

Practical Implications

For business operators and founders, adopting such a structured and transparent evaluation process offers myriad benefits. In the automation, CRM, and broader digital infrastructure domains, understanding AI systems through consistent scenarios enables better decision-making and risk management.

The AI Use Case Worksheet provides a template for organisations aiming to develop a comprehensive evaluation framework. By articulating clear use cases and metrics, businesses can unequivocally assess AI performance, improving reliability and accountability in their digital operations. In the financial services, this could mean more efficient SAR filing systems and enhanced internal call centre support, reducing operational costs and boosting service efficiency.

Implementation Considerations

For operators considering this evaluation approach, it is crucial to focus on stakeholder engagement and iterative feedback mechanisms. SMES play a crucial role in ensuring that scenarios are relevant and accurately represent real-world challenges and opportunities. Integrating such stakeholder insights not only enriches scenario relevance but also aligns AI systems more closely with strategic goals and user needs.

However, organisations should also be mindful of the resource demands involved in consistently engaging experts and maintaining a rigorous review process. Careful planning and phased implementation can ensure these activities are sustainable and productive.

References

Choong, Y.-Y., Greene, K., Qian, A., Marasli, M., Yang, Z., Chen, S., Dabbish, L., Rao, A., & Shen, H. (2023). Towards Apples to Apples for AI Evaluations: From Real-World Use Cases to Evaluation Scenarios. arXiv preprint arXiv:2605.07986v1. Retrieved from http://arxiv.org/abs/2605.07986v1

Note: This paper is a preprint and has not yet undergone formal peer review.