AI Workflow Automation

See how Stonebranch brings large language models from Azure OpenAI, AWS Bedrock, and Google Vertex AI directly into Universal Automation Center workflows.

In this demo, I will show you how you can use AI in the Universal Automation Center workflow. To make this possible, we created AI tasks for the main cloud platforms, Azure OpenAI, AWS Bedrock, and Google Vertex AI. These tasks allow you to use large language models provided by each platform. All access is handled through the cloud platform, which ensures security and also supports GDPR and EU AI Act compliance. With these integrations, you can use AI directly inside your workflows. This means your automation can understand, generate, and transform content. For example, you can analyze logs, create summaries, classify data, convert unstructured data into a structured format, or make decisions based on context. Here you see the AI task for Azure OpenAI, AWS Bedrock, and Google Vertex AI. All three work in a very similar way. The main difference is how authentication is handled for each cloud platform. Let me show you the Azure OpenAI task as an example. First, you choose the authentication method, for example, a service principal, and enter the required details, like the tenant ID and credentials. Then you select the AI model. You can choose any model that is available in your Azure environment. In my case, I have access to models like GPT-4 mini, GPT-5 nano, and o4 mini. For this demo, I will use GPT-4 mini to keep the costs low. Next, you define two things: the system prompt and the conversational prompt. The system prompts tell the AI how to behave. This is where you define rules, tone, and constraints. The conversational prompt is actually the input, what you want the AI to do. This works just like in ChatGPT. Let me show you a simple example. I'm a big windsurfing fan, so I was wondering what a speed of fifty knots means in kilometers per hour. So I ask, please convert fifty knots to kilometers per hour. You can also enrich the prompt with files or USC variables, and you can reuse the output in other tasks. In this case, I want the result in JSON format, so I define this in the system prompt. Let's run the task. Here you see the result, the converted value in JSON format. You can also see the token usage here to estimate costs. This is one of the many options you can set on the task level. This simple example shows that configuring the task works just like in shared GPT. You define the instructions and provide an input. Now let's look at a real business workflow. In this example, we have a company that sells kitchen machines. The purpose of the following workflow is to automate the calculation of seller commissions. In this workflow, seller commissions are calculated based on sales data. Then we enrich the result with sentiment data from customer feedback. After that, the results are sent to SAP, and emails are created and sent automatically. Sellers with high sales and a positive sentiment score are categorized as premium sellers and receive an additional bonus. We use two input files: a sales Excel for the commission calculation and a sentiment Excel file to calculate a score between minus one and plus one. Let's look at the workflow. The first AI task calculates commissions, assigns a seller category such as low, medium, or top. The Excel file is used as input. For this task, we use GPT-4 mini as this model is well-suited for calculations. The output is generated as JSON file containing sales, commission, and category. The business rules are defined in a markdown system prompt, so no coding is needed. In this prompt, you define the Excel input structure, the business rules, the commission rates, and the output format. The second AI task calculates the sentiment score for each seller. Again, the rules are defined in the system prompt. The AI should perform the following steps. Pass and classify the feedback, group and aggregate by seller, and determine final sentiment output in JSON. The result is a JSON file with the sentiment values. For this task, we use GPT-5 nano. Before running the sentiment analysis, we also have an AI task that anonymizes the data to remove any personal information. The anonymization task should perform the following steps: define input and role, anonymize customer data, and preserve structure and output format. Next, we have a merged AI task. This task combines the commission data and the sentiment data into one JSON object per seller. This task reads both the sentiment and the sales commission file as input. The AI can perform then the following steps: define input and merge key, merge and resolve data, and format the output JSON. Finally, we have an AI task that generates emails. It identifies premium sellers, those with sales above five thousand euros, and a positive sentiment, and applies the bonus rules. In the system prompts, I define the following instructions. First, I define role and input, then I apply the qualification rules and processing rules, and I generate the email output, so my email template. Let's run the workflow. First, sales data is uploaded to SharePoint, which starts the commission calculation. Then sentiment data is uploaded to an AWS s3 bucket, which triggers the sentiment processing. The data is anonymized and analyzed, and then both results are merged. After that, approval is done via Slack to trigger the SAP workflow. Finally, the AI credit emails are generated and sent. Here you see the final email. There are several top sellers, but only three sellers are classified as premium because their sales are above five thousand euros and their sentiment score is positive. They also get an additional bonus. Let me summarize. This demo showed how our AI task can be used in Universal Automation Center workflows with large language models from Azure OpenAI, AWS Bedrock, and Google Vertex AI. With these models, workflows can analyze, transform, and enrich data and make decisions based on context. All of this works with a simple markdown instructions. No coding needed.

AI is changing how enterprises process data, but most teams still run AI workloads outside their automation platform. That means manual handoffs, disconnected pipelines, and no governance over how models get called or what they produce. Stonebranch changes that with AI tasks built directly into Universal Automation Center. With native integrations for Azure OpenAI, AWS Bedrock, and Google Vertex AI, your workflows can analyze logs, classify data, generate content, transform unstructured inputs into structured formats, and make decisions based on context. Business rules are defined in simple markdown prompts. No coding needed. All model access runs through your existing cloud platform, so security, GDPR, and EU AI Act compliance stay intact.

AI Workflow Automation in Action

This guided tour walks through how AI tasks work inside UAC, from basic configuration to a full business workflow.

AI task configuration. See how to set up an AI task with authentication, model selection, system prompts, and conversational prompts. The setup works the same way across Azure OpenAI, AWS Bedrock, and Google Vertex AI.
Real business workflow. Watch a complete seller commission workflow that calculates commissions from sales data, runs sentiment analysis on customer feedback, anonymizes personal information, merges results, and generates personalized emails. All driven by markdown instructions.
Multi-model orchestration. Learn how different AI models handle different steps in the same workflow. The demo uses GPT-4 mini for commission calculations and GPT-5 nano for sentiment analysis, each selected based on cost and task fit.
End-to-end automation. See how the workflow connects to SharePoint, AWS S3, Slack approvals, and SAP, with AI tasks running alongside file triggers, data processing, and email generation in a single orchestrated flow.

This video is built for IT operations teams, automation engineers, and data leaders exploring how to bring AI into production workflows. If your organization wants to use large language models for data processing, content generation, or decision-making without building custom integrations, this demo shows how UAC makes it possible.