Periodic Reporting for period 4 - PIX (The Process Improvement Explorer: Automated Discovery and Assessment of Business Process Improvement Opportunities)
Reporting period: 2024-03-01 to 2024-08-31
Every mid-to-large organization needs to continuously improve its business processes to remain efficient, to deliver excellent customer service, to comply with regulations, and to adapt to changes in customer expectations. Business Process Management (BPM) is the art and science of how to organize, coordinate, and continuously improve business processes by making optimal use of human and technology assets in an organization. One of the most recurring questions in the field of BPM is: How to discover and evaluate opportunities for improving one or more business processes?
Existing approaches to discovering business process improvement opportunities are expert-driven. In these approaches, managers identify improvement opportunities based on their intuition and experience. When an improvement opportunity is identified, a team of business analysts are tasked with validating and quantifying the improvement opportunity using data extracted from enterprise IT systems. In other words, the ideas for improvement are generated using expert intuition, and then these ideas are evaluated manually. This approach is sub-optimal: many ideas for improvement are left unexplored, because it would be too time-consuming and costly to consider them all.
The PIX project has built a foundational collection of data-driven methods to discover business process improvement opportunities that optimize one or multiple performance indicators. One of the groundbreaking achievements of the project is a method to automatically discover high-fidelity business process simulation models from execution data. These high-fidelity simulation models enable business stakeholders to obtain accurate and reliable answers to questions such as “By how much would we improve the performance of the process if we automate an activity using a generative AI agent?” or “What would be the effect of adding a new verification step in a process on our response times?”.
The project also led to an integrated tool chain that uses high-fidelity simulation models, in conjunction with optimization algorithms and generative AI techniques, to recommend combinations of process changes to optimize one or more performance indicators.
Existing approaches to discovering business process improvement opportunities are expert-driven. In these approaches, managers identify improvement opportunities based on their intuition and experience. When an improvement opportunity is identified, a team of business analysts are tasked with validating and quantifying the improvement opportunity using data extracted from enterprise IT systems. In other words, the ideas for improvement are generated using expert intuition, and then these ideas are evaluated manually. This approach is sub-optimal: many ideas for improvement are left unexplored, because it would be too time-consuming and costly to consider them all.
The PIX project has built a foundational collection of data-driven methods to discover business process improvement opportunities that optimize one or multiple performance indicators. One of the groundbreaking achievements of the project is a method to automatically discover high-fidelity business process simulation models from execution data. These high-fidelity simulation models enable business stakeholders to obtain accurate and reliable answers to questions such as “By how much would we improve the performance of the process if we automate an activity using a generative AI agent?” or “What would be the effect of adding a new verification step in a process on our response times?”.
The project also led to an integrated tool chain that uses high-fidelity simulation models, in conjunction with optimization algorithms and generative AI techniques, to recommend combinations of process changes to optimize one or more performance indicators.
The PIX project led to three major advances to the state of the art in the field of data-driven business process optimization.
The first advancement relates to the question of “How to automatically evaluate a business process improvement idea?” In other words, how to find out if and by how much a change to a business process will improve its performance? To address this question, we developed a method, called Simod, that automatically discovers a business process simulation model from a dataset containing historical traces of a business process. Simod incorporates several novel ideas, chiefly: (1) the use of machine learning techniques to optimize the fidelity of a simulation model; (2) algorithms to automatically discover “availability timetables” and “work habits” of workers in a business process entirely from historical data, including work habits such as multi-tasking, batching, and prioritization.
The second major advancement of the project is an optimization method, namely Optimos, that takes as input a historical execution dataset and generates all sorts of “candidate changes” that are likely to improve a process with respect to cost and time. This optimization engine relies on Simod to determine if and to what extent a given change improves (or degrades) the performance of the process. The optimization engine then builds a set of Pareto-optimal set of improvement opportunities.
A third major contribution of the project is a tool called Robidium, which analyzes interactions between office workers and software applications to detect repetitive data transfer routines that can be partially or fully automated. Robidium surgically separates tasks that can be automated and tasks that cannot be automated because they involve a human decision or human knowledge not present in the input data. Robidium then generates a script that allows a software bot to automatically replicate the repetitive tasks performed by humans. Each of the repetitive tasks identified by Robidium represents a potential improvement opportunity.
We have also made inroads in the direction of optimization of business decisions. Along this front, we developed a semi-automated method to discover condition-action rules to reduce the defect rate in a business process with respect to a given notion of “defect” (e.g. customer complaints, deadline violations, etc.). For example, we can automatically discover rules of the form “when a customer is from Southeast-Asia, then assigning activity A to worker X or executing activity A before activity B (rather than the other way around) increases the probability that this customer will be satisfied by 10%.” Such cause-effect relations help process managers to identify business process improvement opportunities.
The SIMOD method was transferred via a separate R&D project into a commercial tool for process optimization, namely Apromore. The methods are now part of the Apromore commercial platform. The methods are used to optimize business processes in several dozen large organizations across Europe, Australia, and USA.
The project has also made available an open-source implementation of the SIMOD, Optimos, and Robidium toolsets. The open-source implementation of SIMOD has attracted interest from other research groups and commercial vendors as evidenced by a dozen forks of the corresponding open-source code repositories.
The first advancement relates to the question of “How to automatically evaluate a business process improvement idea?” In other words, how to find out if and by how much a change to a business process will improve its performance? To address this question, we developed a method, called Simod, that automatically discovers a business process simulation model from a dataset containing historical traces of a business process. Simod incorporates several novel ideas, chiefly: (1) the use of machine learning techniques to optimize the fidelity of a simulation model; (2) algorithms to automatically discover “availability timetables” and “work habits” of workers in a business process entirely from historical data, including work habits such as multi-tasking, batching, and prioritization.
The second major advancement of the project is an optimization method, namely Optimos, that takes as input a historical execution dataset and generates all sorts of “candidate changes” that are likely to improve a process with respect to cost and time. This optimization engine relies on Simod to determine if and to what extent a given change improves (or degrades) the performance of the process. The optimization engine then builds a set of Pareto-optimal set of improvement opportunities.
A third major contribution of the project is a tool called Robidium, which analyzes interactions between office workers and software applications to detect repetitive data transfer routines that can be partially or fully automated. Robidium surgically separates tasks that can be automated and tasks that cannot be automated because they involve a human decision or human knowledge not present in the input data. Robidium then generates a script that allows a software bot to automatically replicate the repetitive tasks performed by humans. Each of the repetitive tasks identified by Robidium represents a potential improvement opportunity.
We have also made inroads in the direction of optimization of business decisions. Along this front, we developed a semi-automated method to discover condition-action rules to reduce the defect rate in a business process with respect to a given notion of “defect” (e.g. customer complaints, deadline violations, etc.). For example, we can automatically discover rules of the form “when a customer is from Southeast-Asia, then assigning activity A to worker X or executing activity A before activity B (rather than the other way around) increases the probability that this customer will be satisfied by 10%.” Such cause-effect relations help process managers to identify business process improvement opportunities.
The SIMOD method was transferred via a separate R&D project into a commercial tool for process optimization, namely Apromore. The methods are now part of the Apromore commercial platform. The methods are used to optimize business processes in several dozen large organizations across Europe, Australia, and USA.
The project has also made available an open-source implementation of the SIMOD, Optimos, and Robidium toolsets. The open-source implementation of SIMOD has attracted interest from other research groups and commercial vendors as evidenced by a dozen forks of the corresponding open-source code repositories.
A breakthrough idea developed in this project is that the availability of workers in a business process should be captured using “probabilistic timetables”. In other words, workers in modern knowledge-intensive organizations do not perform work Monday-Friday 9-5pm, but they are rather available to perform certain tasks during different periods of the day, week or month with varying probabilities. The project developed algorithms to discover such probabilistic timetables from execution data. The project demonstrated that simulation models of business processes that use probabilistic timetables are more accurate than those that assume that workers always work during fixed periods of time (e.g. Monday-Friday 9-5pm), especially when those availability timetables are inferred from execution data.
Another breakthrough of the project is a method for dissecting five types of waiting times entirely from historical traces of business process executions: (1) waiting time due to resources being busy; (2) waiting time due to resources being off-duty; (3) waiting to execute work in a single batch; (4) waiting due to prioritization of other work; (5) and waiting time due to external factors. The project demonstrated that this integrated approach to waiting time discovery from data leads to simulation models with higher levels of fidelity. The project also developed an optimization technique that takes into account all these distinct sources of waiting time. This is an advancement with respect to previous approaches that focus on resource contention (resources being busy or not) and resource availability (on-duty versus off-duty).
Another breakthrough of the project is a method for dissecting five types of waiting times entirely from historical traces of business process executions: (1) waiting time due to resources being busy; (2) waiting time due to resources being off-duty; (3) waiting to execute work in a single batch; (4) waiting due to prioritization of other work; (5) and waiting time due to external factors. The project demonstrated that this integrated approach to waiting time discovery from data leads to simulation models with higher levels of fidelity. The project also developed an optimization technique that takes into account all these distinct sources of waiting time. This is an advancement with respect to previous approaches that focus on resource contention (resources being busy or not) and resource availability (on-duty versus off-duty).