Dagstuhl Seminar on AI-Driven Process Execution and Adaptation

I spent an intense week with energetic colleagues from 20+ companies and academic institutions (IBM, SAP, Apromore, Oxford, Ottawa, ..) discussing current trends in the field of AI-driven process execution, in the beautiful setting of Dagstuhl Castle in Germany.

We discussed all sorts of emerging concepts and concerns related to Agentic Process Management (APM), including "framed execution", "conversational actionability", "explainability", "adaptability".

There were four workgroups exploring different themes throughout the week: (1) Autonomy; (2) Adaptability; (3) Explainability; and (4) Conversational Actionability;

I was part of the group on Conversational Actionability. We brainstormed on required capabilities of Agentic Business Process Management Systems (ABPMS), with a focus on their support for conversation and actioning. We'll write more extensively about it in a white paper to be published in a couple of months. Meantime, I wanted to share an initial architecture we converged to after several rounds of discussions.

We elicited four key subsystems in an AI-driven business process execution system: a data layer, a process intelligence layer, an action layer, and a conversational fabric.

The data layer brings together structured and unstructured data about a business operation, including event logs of business processes, other business data (financial data, time series), business process documentation (including text documents and process models), as well as data from physical systems relayed via IoT sensors. This layer allows the ABPMS to sense the current state and evolution of the business process and its environment. It provides a querying interface allowing other layers to leverage business process data and metadata from a variety of sources.

On top of the data layer sits a process intelligence layer, which provides capabilities for process discovery, performance and compliance analysis (process mining), predictive capabilities (based on simulation and machine learning), as well as prescriptive capabilities (recommending interventions and quantifying their effects). In other words, the process intelligence layer provides capability to describe and explain the current state of the process and to predict future states of the process under different scenarios.

Next to the process intelligence layer, the action layer provides capabilities for triggering actions affecting one or more business processes, or interactions with external actors (suppliers, customers, etc.). Example of actions include:

1. Creating or altering the state of a case in a business process orchestrated by a classic (model-based) Business Process Management System (BPMS);

2. Triggering a software bot, e.g. an RPA bot driven by a script;

3. Sending notifications via a communication platform;

4. Updating records in a CRM, ERP or other system of records;

5. or triggering actions that affect the physical environment via IoT actuators.

The fourth component is the conversational fabric. This layer makes available the capabilities of the data layer, process intelligence layer, and action layer to different types of agents. The capabilities of the lower layers are exposed via tools, leveraging the Model Context Protocol (MCP).

An agent may leverage some of these tools to, for example, detect degradations in the performance of a process that may lead to a violation of a Service Level Agreement (SLA). Having detected this risk, the agent may leverage the process intelligence tools to determine interventions that may be triggered to prevent this SLA violation, and it may then leverage the tools coming from the action layer to trigger actions or notifications.

The agents in the conversational fabric receive instructions and goals from end users, whether directly via a chat interface, or indirectly via agents operating in other systems, such as an agent running in a CRM platform.

Some of the agents in the conversational fabric rely on general-purpose LLMs, others are based on fine-tuned or domain-specific models, and some possess explicit planning or reasoning capabilities. Each agent operates autonomously but may collaborate with other agents by passing tasks, sharing context, or requesting specific operations.

Thanks to the participants for highly engaging discussions, Marco Montali, Marco Comuzzi, Daniel Amyot, Irene Teinemaa, and many other seminar participants, and a special shout-out to my seminar co-organizers Lior Limonad Fabiana Fournier Timotheus Kampik Giuseppe De Giacomo.

Acknowledgment: The vision of MCP tools on top of process intelligence and execution owes a lot to ideas exchanged with my colleagues at Apromore, including Marcello La Rosa and Jesus De Carlos.

This is just a brief account of a torrent of discussions we had. We'll post more in the coming weeks in the seminar's page.