SAP Speaks PDDL

Jörg Hoffmann, Ingo Weber, Frank Michael KraftAAAI 2010 · DOI

Abstract

The primary obstacle to deploying AI planning in real-world industrial settings is the "modeling gap"—the prohibitive cost and complexity of creating formal domain descriptions. This research reveals that SAP, a leader in enterprise software, has spent decades maintaining a massive, PDDL-like model of business logic known as Status and Action Management (SAM). By leveraging this existing architecture, the authors automate business process generation and provide the research community with a high-quality, real-world benchmark consisting of over 400 business object types and half a million planning instances.

Introduction: The Modeling Bottleneck

In Business Process Management (BPM), experts manually arrange IT transactions into workflows to achieve specific business goals, such as processing a customer quote through to archiving. Given the vast complexity of enterprise software, this manual process is slow, expensive, and error-prone. While AI planning—the automated generation of action sequences to reach a goal—is a natural fit for this problem, it requires formal models (typically in PDDL, the Planning Domain Definition Language). Historically, building such models for systems as large as SAP was considered an insurmountable task.

Leveraging the SAM Model

The authors bridge this gap by discovering that SAP's engineers had independently developed a variant of PDDL over 40 years of software engineering. The Status and Action Management (SAM) model describes "Business Objects" (BOs) using status variables and "actions" (IT transactions) that modify those statuses.

Example of a SAM-based business process for a Customer Quote (CQ), illustrating non-deterministic outcomes like completeness and consistency checks.

Handling Real-World Complexity

Unlike classic deterministic planning, business actions often have uncertain outcomes. For instance, a "Consistency Check" might result in a state of "Consistent" or "Inconsistent." The authors address this through several technical adaptations:

  • SAM-AO* Algorithm: A search algorithm based on AO* that generates "contingency plans" to account for different possible transaction outcomes.
  • Heuristic Adaptation: The team adapted the Fast Forward (FF) heuristic by "determinizing" the model—calculating distance estimates by assuming the most favorable outcome for non-deterministic actions.
  • Abstraction: SAM succeeds by abstracting thousands of low-level data fields into a manageable set of "status variables" relevant to business logic.

The SAP-PDDL Benchmark

To evaluate their approach, the authors introduced SAP-PDDL, an anonymized translation of the SAM model. This benchmark is unprecedented in scale for the planning community, covering 404 Business Object types and 2,418 transactions. The evaluation suite includes 548,987 planning instances, representing one of the largest real-world datasets in the field.

Results and Performance

The experimental results demonstrate that automated planning is highly viable for enterprise-scale logic. For 275 out of the 371 BO types tested, the planner achieved perfect coverage, solving every instance.

Performance metrics across the SAP-PDDL benchmark: (a, b) show coverage and failure rates, while (c, d) illustrate search effort in terms of state evaluations.

Key performance highlights include:

  • Responsiveness: Most solved instances were completed in under 2.6 seconds, requiring fewer than 1,000 state evaluations. This speed is critical for integration into interactive user interfaces.
  • Complexity: While many tasks were straightforward, "hard" Business Objects required over 54,000 evaluations, indicating that the domain remains a fertile ground for further research.
  • Scaling: The planner encountered difficulties when tasks required the simultaneous coordination of multiple different Business Object types, as the search space expanded exponentially.

Discussion and Future Work

The research highlights a pragmatic shift in planning definitions. In the real world, "strong plans" (guaranteed success) are rare because mandatory checks can fail. Instead, the authors advocate for plans where at least one path reaches the goal and failures are provably unavoidable.

Future work will focus on extending the SAM model to capture dependencies between different Business Objects (e.g., how a "Quote" triggers an "Invoice"). Additionally, long-term pilot studies with SAP customers are needed to validate the business value and ensure that the generated high-level plans can be seamlessly executed by IT systems in production environments.

Summary updated

BibTeX
@inproceedings{hoffmann2010sap,
  author = {Jörg Hoffmann and Ingo Weber and Frank Michael Kraft},
  title = {{SAP Speaks PDDL}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7737},
}
Figure