BPM & Large Language Models
Integrating the Latest Technology into Your Automated Processes
Large Language Models are revolutionizing BPM. Discover how LLMs make your process automation more intelligent and flexible.
Large Language Models like ChatGPT from OpenAI have been on everyone's lips since 2023 and offer a variety of application possibilities with impressive results. Development is progressing rapidly - billions of US dollars are being invested in the further development of these models, so we can assume that these technologies will increasingly be used profitably in more and more companies.
But only through targeted embedding of artificial intelligence in robust, (semi-)automated business processes can productivity gains and efficiency increases be achieved on a large scale in organizations.
Large Language Models in Interplay with Process Automation
We can imagine the Large Language Model as a huge brain that has learned from vast amounts of freely available data. Additionally, we can provide the model with company-internal data that it should include in its "answers." In an automated process, we communicate with the LLM via an API interface. Recently, it has even become possible to force the response in a structured data format (JSON) to process it as automatically as possible.
All this can be realized, for example, in the company's own Microsoft Azure OpenAI environment, so that data is only processed GDPR-compliant in regional data centers.
With a BPM tool like FireStart, the huge general knowledge of Large Language Models, a company's valuable internal data, and the subject-specific knowledge of employees can be combined in standardized processes. Only in this way can artificial intelligence be used scalably to create value.
Application Possibilities of LLMs in Business Processes
Classification
In the area of classification, LLMs offer two major advantages over classic Machine Learning and Deep Learning technologies:
- They are immediately ready for use and don't first need to be resource-intensively trained with large amounts of data.
- And they are flexibly usable, as classes can be changed afterward, or even created by the LLM.
Examples of classification types:
Topic classification: Whether text, images, video, or speech, large language models deliver impressive results in classifying or categorizing content. This classification can then determine how the process continues.
Sentiment analysis: Recognition of a text's mood based on categories like positive/negative/neutral.
Language recognition: Automatic recognition of different languages - for example, to forward inquiries to the appropriate employees.
Intent detection: Recognition of intentions - such as recognizing a customer's goal when contacting a chatbot - to then trigger specific processes.
Text/Data Extraction
From documents, images, free text inputs, etc.
Text/Data Generation
Data analyses: Structured datasets (for example csv or Excel tables) can be analyzed and even created and modified by LLMs. Data visualizations can also be created and attached to forms, for example.
Text generation: For example, summaries of texts, grammatical improvements, translations, creation of image descriptions, etc.
Image and video generation.
Practical Example: Automatic Recognition of a Price Inquiry and Semi-Automated Quote Creation
In this real example, we use the LLM GPT-4o in combination with the OpenAI Assistant API for multi-level classification. A PDF document with a list of product articles with article numbers and descriptions is made available to the AI assistant.
The Challenge:
Recognizing price inquiries (Classification - Intent Detection): A company wants to automatically filter out price inquiries from a mass of messages arriving via different channels (for example via an email address like info@company.com or a web form). It should recognize which articles in what quantity are being requested. Then the right employee in the organization should be notified. For an inquiry sum of over €10,000, for example, a manager should be notified, for under €10,000 an employee from sales.
The employee in the respective role should get a quick overview of the price inquiry and be able to change pre-filled data points of the quote in case of misrecognition by the AI. If everything is correct, a quote should be automatically created and sent to the customer via email with one button press.
The Solution:
Process Start: When a new email arrives in the mailbox or when a web form is submitted, the workflow is automatically started.
Classification with the OpenAI Assistant: With the FireStart low-code REST API activity, communication with the assistant is easily established. We receive a JSON object as response, with which we learn whether it's a price inquiry or not. If yes, we additionally get a JSON array with the respective article numbers and quantities.
Calculate inquiry value: We access current price data to determine the inquiry value.
Notify the right person: Via a role model in FireStart, which is connected to the company's Active Directory, a specific employee is informed depending on the inquiry value. We use the FireStart Form Builder and populate the form simply via drag & drop with the respective data points that the AI assistant returns to us. The employee can view the form directly in Microsoft Outlook. The fields in the table are editable.
When the employee confirms the quote via a simple button, it's sent to the inquiring person via email.
Implementing AI-Supported Processes with FireStart
In just a few steps, you can set up AI-supported processes (example see above) with FireStart and thus optimize them. Our BPM tool enables seamless integration of OpenAI. This way, our software supports you in accelerating processes in your company and using your employees' resources more efficiently. Book a demo and convince yourself how your BPM benefits from AI.
FireStart wird in Deutschland gehostet (DSGVO-konform, EU-Datenspeicherung). Website: www.firestart.com. Kontakt: sales@firestart.com.