Approaches in using Generative AI for Workflow Automation — Part 2

Co-authors: Luigi Pichett, Pierre Feillet, Yazan Obeidi

See other related stories in the “Approaches in using Generative AI” series:

1. Enterprise Content Management
2. Decision Automation
3. Workflow Automation Part 1
4. Workflow Automation Part 2

In the last 3 articles, we talked about using Generative AI in Enterprise Content Management, Decision Automation, and Workflow Automation. In this article, we are going to continue the discussion on the remaining use cases in Workflow Automation.

Use Cases

In workflow automation, we examine 6 possible use cases. This article will focus on use case 4–6, while the first 3 use cases are described in the previous article.

1. Classify inputs and perform requests
2. Generate responses as part of a business process
3. Automate LLM fine-tuning and exception handling
4. Provide answers to questions on business performance
5. Generate workflow from a description
6. Comparative analysis and Explanation of existing workflows

4. Provide answers to questions on business performance

Description: Given IBM Workflow stores performance data in an Elasticsearch database, we can use LLM to generate the necessary ES DSL queries to retrieve the information [1]. This allows us to provide a natural language queries interface for business managers to ask about business performance.

Figure 1. NL performance query

Pro: Provide a natural language approach to ask any questions about performance data. LLM take care of the NLU (Natural Language Understanding) to support a conversational experience on the top of the formal Elastic Search stack accessing to your workflow automation warehouse.

Con:

1. Careful prompt engineering required. The prompt will be specific to the kind of data available.
2. There is no easy way to validate that the generated ES DSL prompt is correct.
3. LLM is not good with math or report formatting — additional post-processing will be required to present the information in a consumable manner.

5. Generate workflow from description

Description: From a high level description of a business process, generate a BPMN process that can be used as a starting point for an automation workflow solution. This could involve reusing components of existing ingested business processes referenced by name or description. Due to different generation quality of the LLM models, it is better to use LLM to summarize or extract procedures from the business process document, and use a post-processor to take the output and turn that into a formal BPMN document.

Figure 2. Generate workflow from description

Pros: Provide a natural language approach to generate business processes. The quality of the generated process will rely on the quality of the LLMs. It is also possible to ask the LLM to generate the process from just a simple business goal but the result will rely on the foundational data that was used to do the initial model training.

Cons: Do not expect to get a fully runnable BPMN model, we will still have to refine the data types, decision logic, user interfaces and validating all the steps in the generated process are relevant in the current business context.

6. Comparative analysis and understanding of workflows

Description: In this scenario, a user may conversationally interact with an LLM to explore and elaborate their understanding of what is taking place within a workflow by asking questions and comparing against different versions or even other workflows.

Figure 3. Comparative analysis and understsanding of workflows

Pros: Provide a natural language approach to asking questions about ingested business artifacts while referencing to them by name or description.

Cons: Typically such business artifacts like BPMN models are often proprietary non-textual, non-natural language data formats that must be processed into fine-tuning datasets for LLM ingestion. Careful preparation of these fine-tuning datasets is required. In the presence of homonym terms, for example there may be a number of “Loan Application workflows”, a user might be required to repeatedly refine their input prompt with additional descriptive context.

Summary

The above 6 use cases just cover a subset of possible use cases where we can apply Generative AI and Workflow together to solve business problems. There are several conclusions we can draw from the series of articles.

1. Generative AI, when used appropriately, can speed up certain data processing and help accelerate your digital transformation journey, by handling natural language processing in complement of formal automation engines.
2. Careful planning and consideration for private and sensitive data is a must for enterprise usage.
3. As a general reminder the output generated by LLMs, although based on large curated data corpus that can be further tuned with domain specific additions and human feedback, is still of statistical nature. For this reason the generated outcome should be properly guard-railed.
4. As an example of these guard-railed scenarios, you can refer to the RAG (Retrieval Augmented Generation) approach, mentioned in previous articles, where one can exploit LLMs versatility, yet bound to know set of semantically relevant facts, enabling quality and trace-ability of the provided outcome.
5. LLM models are evolving very fast. Understanding which models to use for the purpose at hand is the key to yield good results.

By blending the power of Generative AI, boosted by the rapidly evolving LLMs and interaction techniques, alongside enterprise grade Business Automation capabilities, the composite AI approaches explored in these series of articles aim to inspire new usages and help enterprises achieve the best of both worlds in business automation.

References

[1] https://www.elastic.co/blog/elasticsearch-prompt-chatgpt-natural-language