From Decision Models to AI Agents

When making critical business decisions, it is essential to keep powerful — yet often unpredictable — generative AI on a controlled and reliable path. This can be achieved by guiding large language models (LLMs) to leverage proven decision intelligence tools such as business rules, machine learning, and optimization. OpenRules addresses this need by enabling customers to automatically transform decision models into independent, intelligent MCP servers, seamlessly accessible by leading LLMs such as ChatGPT and Claude.

By combining business rules, machine learning, optimization, and AI agents, OpenRules empowers customers to build smarter, more adaptive decision-making systems—all within a unified decision intelligence platform.

All the necessary tools, resources, and prompts are derived directly from decision models and published as local or remote MCP servers. No coding or complex manual configuration is required. A subject matter expert can then ask any LLM to identify the correct decision for a given problem, invoking the decision models already deployed as MCP tools.

Throughout this dialogue, the LLM does the following:

• Determines which tool to use and when, guided by the available prompts
• Automatically generates and dispatches JSON-formatted requests to the appropriate MCP server for execution
• When input data is incomplete, the LLM prompts the user accordingly and proceeds to execute all related decisions and sub-decisions
• Provides transparent explanations of which tools were invoked and why, rendering technical logic in plain English
• Transforms all decision outputs into natural language responses, capturing both the result and the underlying reasoning.

Thus, using LLMs with decision models deployed as MCP Servers, OpenRules supports a tool-augmented, decision-making LLM workflow that implements the autonomous reasoning loop:

Understand → Decide → Act (tool) → Interpret → Explain

Here’s a step-by-step explanation of what’s happening:

1. Deciding Which Tool to Use

The LLM reads the user’s input and figures out what kind of task it is:

• Is it a question? → answer directly
• Needs external data? → use a search tool
• Needs computation or file work? → use a code tool

It uses its internal reasoning to select the most appropriate tool.

2. Generating Tool Requests (JSON)

Once a decision tool is chosen, the LLM:

• Builds a structured request (usually JSON)
• Sends it to the chosen tool

3. Handling Missing Information

If the input is incomplete, the LLM:

• Pauses and asks clarifying questions, OR
• Makes reasonable assumptions

Then it continues the workflow, making all necessary sub-decisions along the way.

4. Explaining Tool Usage (Transparency)

A well-designed system will:

• Explain which tool was used
• Explain why it was needed

This turns hidden system behavior into something the user can understand.

5. Converting Everything into Natural Language

Finally, the LLM:

• Takes raw outputs (data, results, API responses)
• Combines them with its reasoning
• Produces a clear, human-readable answer.

Real Example

Let’s consider how to convert a pure rule-based decision model, “Patient Therapy“, into an LLM-based dialogue.

<work in progress>