FRM: An Intelligent Assistant for Financial Resource Management

Andrew Gelman, Susan Altman, Matt Pallakoff, Ketan Doshi, Catherine Manago, Thomas C. Rindfleisch, Bruce G. BuchananAAAI 1988

Introduction

Financial Resource Management (FRM) represents a significant departure from traditional spreadsheet-based budgeting. While spreadsheets are powerful tools for numerical calculation, they are often "brittle" when faced with the complex, symbolic, and judgmental aspects of organizational planning. Developed in the late 1980s, FRM was designed as an intelligent assistant that treats budgeting as a knowledge-intensive constraint satisfaction problem. It moves beyond simple grids of numbers to incorporate institutional policies, project roles, and multi-perspective reporting requirements.

System Architecture

The FRM system is built upon an object-oriented architecture, utilizing the CLASS/HYPERCLASS system to separate data management from user presentation. This modularity allows the system to maintain a single, consistent source of truth while providing various "perspectives" or views of that data.

The modular architecture of the FRM system, showing the interaction between the user interface (FORMAN), data management (CONFRM), and the reasoning engines.

The architecture consists of several integrated components:

  • FORMAN (User Interface): A visual "form-filling" environment that provides a WYSIWYG experience.
  • CONFRM (Data Management): The central repository for budget data, decoupled from how that data is displayed.
  • Constraint Manager: A symbolic reasoning engine that enforces both numerical and logical rules.
  • Hierarchical Planner: An automated module that suggests corrective actions when constraints are violated.
  • Explanation Facility: A tool that provides transparency by explaining the derivation of specific budget values.

Budgeting as Constraint Satisfaction

At the heart of FRM is the insight that a budget is a network of interconnected constraints. These include "hard" constraints, such as institutional policies or fixed salary caps, and "soft" constraints, which reflect managerial preferences. Unlike standard spreadsheet formulas, FRM's constraints can be symbolic. For instance, the system can recognize job titles and apply logic based on personnel roles. A notable feature is the "Super-Item," which allows managers to group multiple related items (like several part-time clerical staff) into a single high-level entity for planning purposes while retaining the underlying granular data for detailed reporting.

User Interaction and Transparency

The system prioritizes user understanding through its Explanation Facility. When a user questions a value—whether it was entered manually or generated by a constraint-driven corrective action—the system can trace the logic back to its source.

The FORMAN interface showing a detailed budget form and an explanation window clarifying the calculation of a specific personnel cost.

This transparency is crucial for "judgmental" planning, where a manager needs to know why a certain adjustment was suggested by the system's planner. If a budget exceeds its limit, the hierarchical planner uses heuristics and meta-rules to propose a sequence of modifications to bring the budget back into compliance.

Hierarchical Perspectives

Large organizations frequently need to report the same financial data to different external bodies, such as the NIH or the NSF. FRM addresses this through hierarchical perspectives—tree-like structures that organize budget items according to specific reporting standards. By separating the underlying data from these views, FRM ensures that updates in one perspective are automatically and consistently reflected across all others, eliminating the manual errors common in traditional reformatting tasks.

Evaluation and Performance

The FRM prototype was evaluated using internal budgetary data modeled after Stanford University’s financial environment. It was compared against a legacy procedural FORTRAN program. The results demonstrated that FRM’s declarative, knowledge-based approach offered significantly greater flexibility than previous systems. While the prototype, running on Xerox 1186 hardware, occasionally experienced response times of up to 15 seconds for complex operations, this was considered an acceptable trade-off for the advanced reasoning capabilities it provided.

Conclusion and Future Directions

FRM successfully demonstrated the feasibility of an intelligent assistant for financial management. However, the authors identified several areas for growth, including the need for better "negotiation expertise" to handle conflicting constraints and the ability for the planner to generate multiple alternative scenarios rather than a single solution. Future iterations also aimed to implement distributed database support to better serve large, decentralized organizations.

Summary updated

BibTeX
@inproceedings{gelman1988frm,
  author = {Andrew Gelman and Susan Altman and Matt Pallakoff and Ketan Doshi and Catherine Manago and Thomas C. Rindfleisch and Bruce G. Buchanan},
  title = {{FRM: An Intelligent Assistant for Financial Resource Management}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}
Figure