Connect to Financial and Operational Data
For AI solutions to deliver value to FP&A users, they must be integrated with the organization's financial and operational data from systems such as the ERP, HR platform, and other business applications. Without this integration, the tool is merely a generic chatbot for document search or leisure, incapable of providing the financial insights FP&A teams expect from AI.
Data Security
However, when AI is integrated with financial and operational data, data security must be a primary consideration. Just as any user should only access data they are authorized to see, the same principle must apply to an AI solution; this is the most fundamental requirement.
Consequently, when different users pose the same question, the AI must extract distinct datasets based on their individual permissions and perform its analysis exclusively on that authorized data.
Response Accuracy
Although a tendency to hallucinate is inherent to current AI models, and complete elimination may not be feasible, the ongoing development of these models focuses on significantly reducing its frequency. However, as the architect of an FP&A AI solution, it is critical to mitigate this risk at the system level. This begins with ensuring the accuracy of data sourced from underlying systems before it is passed to the model for reasoning and analysis. By implementing robust data validation and processing mechanisms, we can guide the model's behavior and minimize hallucinations.
Response Performance
While the market has set an expectation that AI responses may be slow, particularly for complex tasks, it remains critical to minimize response times to avoid user frustration. Users should not be expected to wait for extended periods while the system processes a query.
To achieve this performance, it is essential to optimize the AI's computational efficiency by having it reason only with necessary information. For instance, using the AI to perform large-scale data aggregations—such as summing transaction data via NL2SQL—is an inefficient use of its capabilities. Such tasks are better handled by dedicated data processing systems.
Response Relevancy
Users expect AI systems to provide answers that are directly relevant to their queries. To meet this expectation, it is essential to carefully design the system's instructions, which guide how the AI interprets and responds to questions. This often requires fine-tuning these instructions for specific AI models to maximize response accuracy.
Involving business stakeholders in developing these instructions is crucial. For complex scenarios, a multi-agent architecture can be employed, where different stakeholders contribute specialized instructions for various topics. The key to success is structuring the guidance in a way that enables the AI to accurately understand and address the most common user questions.
Easy to Maintain and to Train
AI solutions are not static products; they are dynamic systems that require ongoing maintenance and updates to ensure peak performance and relevance. Therefore, they must be designed with administrator-friendly maintenance in mind. This includes the capability for straightforward model retraining and fine-tuning, which is essential for preserving the solution's long-term value and accuracy.
Others
Of course, the features listed above represent what we consider the Minimum Viable Product (MVP). However, specific use cases and industry requirements may introduce additional considerations.
As with any system, it is important to manage investment carefully. We recommend starting with a small-scale implementation, experimenting to gather insights, and then iterating based on the results. This approach is particularly relevant for AI, which is still a novel area for many organizations. While adoption can involve a learning curve and initial hesitation, the prevailing trend is toward leveraging AI for significant productivity gains.
Given that this shift is becoming inevitable, there is a clear advantage in beginning to experiment today.
