Back to AI Research

AI Research

Process-aware graphs help agents choose valid enterprise API calls

Key Takeaways

  • The study combines endpoint retrieval with workflow order, payload schemas and allowed code lists.
  • ### An endpoint description is not a complete process
  • The paper [Grounding LLM Agents in Business Processes](https://arxiv.org/abs/2610.37242) examines multi-step enterprise API use.
  • An agent must select an endpoint, follow process order, use valid payload keys and choose values from predefined code lists.
  • Retrieving each endpoint as an isolated document can miss those relationships.

An endpoint description is not a complete process

The paper Grounding LLM Agents in Business Processes examines multi-step enterprise API use. An agent must select an endpoint, follow process order, use valid payload keys and choose values from predefined code lists. Retrieving each endpoint as an isolated document can miss those relationships.
The proposed approach combines OData metadata and BPMN reference processes in a typed property graph. A hybrid retriever uses dense embeddings to find candidate endpoints, then follows process edges to retrieve relevant subsequent steps.

The process layer explains part of the gain

The service-retrieval evaluation covers 42 enterprise services. For 74 queries with low lexical overlap between the request and downstream entities, the authors report a 5.7-point improvement in Recall@3. The method matches dense retrieval on other queries and shows no improvement on OpenAPI services without a process layer.
That boundary matters. The study is evidence for exposing useful process structure, rather than proof that adding any graph improves every service-discovery task.

Valid calls and completed tasks are different metrics

In an 81-query multi-step agent evaluation, the method completes 38.7% of tasks, compared with 10.3% for a chunk-based baseline. A separate elicitation probe finds that providing valid code lists removes invented codes but still leaves invalid payload keys in many write runs. Full typed context reduces invalid calls to three of 101 runs, with no invented codes.
The call-validity result should not be mistaken for a 97% end-to-end task-completion rate. Correctly formed calls can still be part of an unsuccessful workflow.
For teams building enterprise agents, the study suggests inspecting what retrieval omits: process order, field names and permitted values may be as important as a natural-language endpoint description. Reproducing the result requires a meaningful process layer and evaluation of completed work, rather than counting only whether a tool request was syntactically accepted.

Comments