Abstract
As AI assistants transition from simple episodic tasks to complex, multi-step enterprise workflows, the underlying orchestration often becomes a "black box" to the user. This paper presents a planner-based approach for dynamically composing "aggregated assistants" from independent skills. By leveraging Explainable AI Planning (XAIP) techniques, the system provides transparency, allowing users to query the "What," "How," and "Why" of the assistant's automated reasoning.
Introduction: The Transparency Gap in Aggregated Assistants
Modern voice and enterprise assistants are moving beyond single-step commands toward complex Robotic Process Automation (RPA). These systems are often built as "aggregated assistants," where a central orchestrator combines various atomic "skills"—such as querying a database or processing a form—to achieve a high-level goal.
However, when an AI planner automatically strings these skills together, it creates a transparency problem. If an assistant autonomously fetches data and makes decisions in the background, the final output can feel unsettling or confusing to the user. This lack of context may also conflict with data privacy regulations, such as the GDPR's "right to explanation."
Methods: Planner-Based Orchestration
The authors propose a transition from stateless event-broadcasting to a stateful Planner-Based Orchestrator.
- Skill Specification and PDDL Compilation: Individual skills are defined by their required inputs and outputs. These specifications are compiled into a Fully Observable Non-Deterministic (FOND) planning problem using the Planning Domain Definition Language (PDDL).
- Goal Reasoning: As shown in the system flow, a Natural Language Understanding (NLU) engine maps user intents to symbolic goals. A planner then generates a sequence of skills to satisfy these goals while monitoring execution to handle failures or loops.
- Explainable AI Planning (XAIP): To ensure transparency, the system supports three query types:
- What? (Summarization): Uses planning "landmarks" to provide a concise summary of the critical path taken, filtering out irrelevant background processes.
- How? (Drill-down): Identifies the specific skill and data inputs used to establish a particular fact.
- Why? (Justification): Employs causal chain regression to trace the reasoning from a final goal back to a specific action (e.g., explaining why an email address was needed for a loan status check).
Key Insights: Handling Complexity and "Creepiness"
A significant advantage of this stateful approach is the ability to handle "graceful digression." Because goals are managed on a stack, a user can interrupt a workflow (e.g., a loan application) to ask a side question (e.g., "What is my credit score?"). The assistant fulfills the secondary goal and then seamlessly returns to the primary task.
Furthermore, the authors highlight that transparency is a fundamental UX requirement. Without explanations, autonomous background actions—like fetching a credit score based on an email address—can feel "creepy." By using landmarks to filter the "noise" of online planning, the assistant provides a clear, human-readable narrative of its actions.
Results and Future Work
The system was qualitatively validated using a banking scenario involving loan and credit card applications. The prototype successfully demonstrated dynamic skill composition and the generation of natural language explanations. For instance, when asked why it collected an email ID, the system could trace the requirement through a database query service back to the ultimate goal of processing a loan.
Future research aims to reduce the overhead of manual PDDL modeling by learning skill specifications from execution traces. Additionally, the authors plan to incorporate "metric" goals to optimize for business KPIs and chain orchestration explanations with model-level explainability (e.g., explaining the specific features that led to a credit score decision).
Summary updated