Hybrid AI Workflow Demo: AI Agents + Deterministic Automation

In this demo, I will show how Universal Automation Center connects file transfers and orchestrates the applications that use the data. We will follow sales data through AI processing, SAP analytics, and planning. We will also see AI investigate a failure and open an incident ticket automatically in ServiceNow. Then use Robi AI to help fix it with my confirmation. Finally, a Kong AI task, we create an execution report and post it to Slack. Let me explain the demo workflow. Uploading sales data to SharePoint triggers a workflow. UAC downloads the file to an SAP application server on AWS, then starts two parallel streams. The first stream marks sensitive fields and waits for manual approval before starting the model training pipeline through the Google AI agent pipeline task. UAC triggers and monitors that pipeline. The second stream loads the sales data into SAP through an iFlow. Other programs then calculate commissions. After that, the workflow branches again. One pass runs Snowflake's analytics, while the other harmonizes data in SAP DataSphere and loads sales data quantities into SAP IBP for planning. For analytics, SAP Business Data Cloud Connect for Snowflake may... Makes published SAP data products available to Snowflake through zero copy sharing. Think of a data product as a prepared dataset with business context. Snowflake can use that shared data without a separate copy for this connection. UAC checks that sales data product is ready before starting the analytics task. Once analytics and planning pass finishes, the report task brings their results together for the team. Let me upload the sales file to SharePoint. The file is now on the SAP Linux server and the downstream task have started. Let's look at the AI masking step. This task uses Azure OpenAI to mask the data before it reaches the Gemini pipeline. I choose an authentication method such as service principle and configure the tenant and credentials. Then I select the model available for this connection. For this demo, I'm using a smaller model to keep costs low. The system prompt defines a masking rules. The conversational prompts supplies the input data. Here is the output with the selected field masks. Before the Google AI pipeline starts, I've edited a manual approval step assigned to Nils. I can handle this in Universal Portal on my mobile device. The new Universal Portal supports many languages and runs on any device. You can even use it with a mobile with one hand because it has a responsive design. Let me approve the task. The approval is complete and the Google AI agent pipeline has started. Meanwhile, one task has failed. Any error in Universal Automation Center automatically triggers a Universal Event, which is captured by our configured webhook. This webhook launched the Kong AI task. The task retrieves the failed instance logs through the Stonebench MCP server, uses Dynatrace tools for additional context, and asks the model to analyze the issue. It then creates automatically a ServiceNow incident through the ServiceNow MCP server. So three MCP servers are involved in this scenario. As you can see here, the task can connect to multiple MCP servers, subject to the access configured in Kong Konnect. Its model request pass through the central Kong AI gateway. It can also limit which tools are available to the task. I can enter prompts directly or use a versioned script from the UAC script library with variables for the current run. Here, we described instructions how to investigate a failure and create the incident. Here you see the incident, including the AI analysis. The SAP task uses the wrong upper variant list d instead of list p. A mistake like this can happen when variant names differ between development and production. This happened to me already several times in projects. Let me fix the issue with Robi. I will ask Robi to verify the issue, then update the variant to list p and restart the instance. Robi asked for confirmation before proceeding. I will confirm the change and check the result. The SAP task has now completed successfully. I will close the ServiceNow incident and verify its status. The incident is closed, and the workflow has continued through DataSphere and SAP IBP. The final Kong AI task retrieves the workflow task log through the Stonebench MCP server and compiles a structured execution report. The report is then posted automatically to our workload automation Slack channel, giving the team a summary of the run. That's a complete flow from a SharePoint upload through AI processing, SAP, analytics and planning with human approval, AI assistant recovery, and automatic reporting coordinated by UAC.