On Supporting the Use of Procedures in Office Work

Richard Fikes, D. Austin Henderson Jr.AAAI 1980

In their seminal 1980 paper, Richard Fikes and D. Austin Henderson, Jr. challenged the prevailing paradigm of office automation. At the time, computer scientists viewed office procedures as analogous to computer programs: rigid, executable sequences of steps. Fikes and Henderson argued that this model acted as a "straight-jacket," failing to account for the inherent unpredictability and "open-ended" nature of real-world organizational tasks. Instead, they proposed a shift toward "symbiotic processing," where artificial intelligence techniques support human workers in navigating the dynamic process of planning and problem-solving.

The Problem: The Illusion of Routine

The fundamental issue with traditional office automation is the assumption of a closed world. In a closed system, all variables are known, and procedures can be executed without deviation. However, the authors argue that the office is an "open-ended" environment where assumptions frequently fail, resources vanish, and goals shift.

To illustrate this, the authors cite an ethnographic case study of a Xerox customer order entry clerk. A standard procedure might require a fixed street address for a copier delivery. However, the clerk encountered a situation where the customer was located on a moving barge. A rigid, automated system would have rejected the input or crashed. The human clerk, however, reasoned about the underlying goal—successful delivery—and created a workaround by using a phone number to track the barge's location. The challenge, therefore, is building systems flexible enough to support this human-centric "replanning" rather than merely automating brittle scripts.

Proposed AI Paradigms for Office Systems

Fikes and Henderson suggest that office systems should be built using core AI paradigms to facilitate a partnership between human and machine:

  • Planning and Execution Monitoring: Rather than "executing" a procedure, the system should treat it as a "plan." Drawing on models like STRIPS, the system monitors preconditions and expected results. This allows the system to help the user detect when a plan has gone off the rails and requires modification.
  • Hierarchical Planning: Borrowing from Earl Sacerdoti’s work, the authors suggest representing tasks at varying levels of detail. This enables the system to track high-level goals (e.g., "Complete the Sale") even if it does not understand the specific, ad-hoc sub-tasks a human is performing to achieve them.
  • Specialization-Based Knowledge Representation: Using formalisms like KL-ONE, the authors propose a "taxonomic lattice." This allows the system’s "world view" to evolve over time; for instance, a new concept like a "barge address" can be defined as a specialization of a standard address, allowing the system to grow with the organization.
  • Semi-Formal Descriptions: A key innovation is the use of "semi-formal" structures. These mix data the system understands (dates, prices) with uninterpreted text or diagrams intended only for humans. This acts as an "escape valve" for information that is too complex for the system to process but essential for the worker.

Key Insights: Procedures as First Approximations

The paper offers several transformative insights into the nature of work:

  1. Procedures as Guides, Not Programs: A written office procedure is not a program; it is a "first approximation." The worker’s real job is to determine if the procedure’s assumptions hold in the current context and to fill in missing details.
  2. Symbiotic Processing: The authors reject the "man vs. machine" dichotomy. They advocate for a partnership where the computer manages complexity—tracking interfaces, reminders, and consistency—while the human handles semantic understanding and open-ended reasoning.
  3. The Multi-Agent Nature of Work: Office work is a distributed, multi-processor environment. Tasks are spread across people and systems who make "commitments" to one another. In this view, replanning is essentially the renegotiation of these commitments when circumstances change.
  4. The Reality of Exceptions: The distinction between "routine" work and "exceptions" is often artificial. Most routine cases are simply those where the human partner did not need to intervene to fix the plan. Systems should be designed to handle all cases with the same underlying flexibility.

Conclusion and Future Directions

Fikes and Henderson provide a conceptual framework for the next generation of office systems, emphasizing three requirements: flexibility, goal-awareness, and evolution. They acknowledge that fully automatic planning is impossible because a system will never truly "understand" the open-ended world.

The most promising path forward, they suggest, lies in multi-processor commitment-based planning, where the system acts as an intelligent assistant tracking the promises made between different agents. However, they highlight that representing "partial knowledge"—the incomplete and often conflicting descriptions of the world—remains a significant, unsolved challenge for the field of knowledge representation.

Summary updated

BibTeX
@inproceedings{fikes1980supporting,
  author = {Richard Fikes and D. Austin Henderson Jr.},
  title = {{On Supporting the Use of Procedures in Office Work}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1980},
}
Figure