Decision Orchestration with Agentic AI

In the Trisotech Examples folder is a simple decision service, Calculate BMI Category. A BKM computes the Body Mass Index from Height and Weight, and then a decision table matches that to a category.

We can use MCP to provide a simple conversational interface to this service. Of course, Agentic AI can do more powerful things, but this illustrates the integration. To create the agent calledTrisotech Helperin Azure Copilot we use Microsoft Copilot Studio.

In this case we want the agent to execute the decision service, which is exposed in the Trisotech Service Library as an MCP endpoint.

In Copilot Studio, we selectAdd toolwith type MCP and point it to this URL.

Now we can execute the service through a natural language prompt, such as “Calculate my body mass index category assuming that I’m 6 feet tall and weigh 175 pounds.” The agent must convert this prompt into the service inputs. In this case, the units are specified in the BKM Description panel:

The agent knows this means use Height = 72 (inches) and Weight = 175. That’s pretty amazing. It then executes the service, returning “Normal”. The response generated using ChatGPT is “Your Body Mass Index (BMI) is in the ‘Normal weight’ category”.

Agent as a Performer

In the second use case, a BPMN User task performer, normally a person, can be replaced by an agent.

In this case, the agent is created in similar fashion to the previous example, but instead of invoking the service via MCP, the agent is defined as a custom performer type that can be selected to replace a human performer for certain tasks. Here is a simple example.

In this process a human task validates that a submitted prescription contains the required elements, again as specified in the task Description panel.

Now we can configure the Performer of the User task to be a particular AI agent type, in this case an agent called “Medical Assistant”.

The agent must be created using the Azure OpenAI tools with the goal of validating that the data inputPrescriptioncontains the elements listed in the task Details, true or false. When you test this in the Decision Modeler with complete information, the agent returns true. Omit some, such as the prescribing doctor, and the agent returns false.

Safeguarding Agentic AI

While Agentic AI is new and exciting, it must heed the lessons of the past in order to succeed. Trisotech’s long experience in Decision Orchestration offers a compelling advantage.

Just as BPM evolved to solve the “white space” problem confounding handoffs between departments and systems, multi-agent workflows face similar issues of gaps in responsibility and conflicting priorities. Trisotech’s embrace of BPM+ standards, designed to shrink the white space through explicit coordination of tasks and responsibilities with enforceable policies and rules, ensures that Agentic AI doesn’t fall into the same traps.

Vague and incomplete business requirements, the traditional bane of software engineering, pose the same issues for Agentic AI. But Trisotech’s suite of graphical business-oriented modeling tools – for processes, decisions, and dynamic cases – ensures that the intent of business stakeholders is clearly and completely reflected in the requirements for Agentic AI, just as they are for conventional Decision Orchestration.

Multi-agent systems require mechanisms for conflict resolution, prioritization, and utility alignment. Without those, agents may act at cross-purposes, compete for resources, or exploit loopholes in their reward structures, just as humans do. Trisotech lets you encode policies and compliance constraints directly into decision, case, and process models, ensuring agents operate within the same rules and incentives that guide human teams.

Agentic AI is intelligent but not 100% correct. When agents are overconfident, underinformed, or misaligned, their decisions can cascade into systemic failure. We must ensure mechanisms for escalation and human override. Trisotech features such as Attended Tasks allow human oversight of critical agentic recommendations, and agents can be designed to recognize when to defer to humans or external decision services for safe and appropriate behavior.

Shared terminology and understanding across the white spaces is critical for success on a large scale. With Trisotech Digital Enterprise Graph (DEG) and Knowledge Entity Models (KEM), semantic models and ontologies are embedded into the modeling layer. This provides agents with a shared, machine-readable understanding of concepts, relationships, and context, ensuring alignment and reducing “semantic drift” across distributed agent teams.

While adding new agents may increase functional capacity, it can exponentially increase orchestration complexity. Decentralized control requires robust protocols for delegation, synchronization, consensus, and trust. Trisotech acts as the Decision-Centric Orchestration layer for agents. It defines who does what, when, and under what conditions, in models that can be tested, versioned, and governed.

These are still early days for Agentic AI. The inherent power is obvious, but trust still must be earned. By integrating it with proven lessons of the past, Trisotech is showing the way forward.