A4A: Process Intelligence Quantifier
Automation techniques had an accelerated improvement over past few years. These improvements have twofold effect. First, an already automated task can now be performed with much higher level of confidence. Automation models that were introduced in the previous decade have matured to such an extent that they are now dependable enough. Second, new type of tasks can now be automated. Tasks that were considered near to impossible cases for automation ten years ago, are already automated.
Here, we are focusing on this second effect i.e. how we can find new opportunity for automation.
Before we try to automate any task, we must identify it. We must find something that is repeatable as well as automatable. Tasks that are easy to identify from a complex business process have already been identified and may have been automated as well. Business analysts have gone through the processes time and again to find new opportunity to automate. What is left now is tasks that are difficult to identify.
Through data observation and process mining it is possible to solve the problem of identifying repetition of work in business process tasks. However, the degree of automatability of repetitive task is dependent on its complexity. Therefore, we must look at complexity of a task. It is to be noted that, a task that is difficult from users’ standpoint, may not be difficult for a machine and vice versa. For example, a complex screen design with multiple tabs and many options could be difficult for a user to manage but is not complex for a robotic process automation tool.
Some technique to automatically measure the complexity of tasks from automation point of view is important to avoid loss of effort. We introduce a term Process Intelligence Quantifier (PIQ) to score a task within a business process on complexity from automation standpoint. One can think of a PIQ scale of 1 to 10 (10 being most complex) and rate each unit of work there. With today’s technology we are possibly able to automate anything that is rated 4 or below. Tomorrow, we might be able to move further up in the scale.
A term that is often used in this context is ‘Cognitive load’. Loosely put, cognitive load is the amount of working memory or brain power used to solve a problem. In other words, it indicates the difficulty level of a task. It is further categorized in to three sub categories, depending on the root cause of the difficulty. To discuss, let us consider an example of a child solving a math homework problem.
- Intrinsic load – How hard the problem is. Is it addition / subtraction problem, or unitary method?
- Extraneous load – Is the environment around him calm and suitable for studies? Does he have a clear description of the problem? Does the problem description include a pictorial representation?
- Germane load – What is the expertise level of the child? Is he math champion or scared of mathematics?
Note that, each of these three types can individually affect the time he would take to complete the homework, and the amount of effort he must put in. The same concept can be applied to people working in office, performing steps of business process. Pause for a moment, and you can easily identify the three types of load in any office job.
Cognitive load therefore, determines the limit on possibility of automation. A task with high cognitive load may not be automated given the maturity of available tools and technology. On the other hand, tasks requiring low cognitive load can be performed by a machine. Measuring the cognitive load can be a good starting point to define such a complexity scale.
Several studies have been done by different research groups to estimate the cognitive load. Most of these methods are not suitable for determining cognitive load in an office environment. Three basic issues are
- In regular days, users do not work in a controlled/constrained environment
- Users do multitask, so the task steps are interleaved with other tasks
- A business process task is assigned to a role, thus multiple users working on same task.
While identifying repetitive tasks through data observation, these issues can be dealt with. This data includes task variant characteristics and time series analysis of user activities on the tasks of the process. It also includes information about task interaction within a process and process to process interactions. As these data are averaged over a period and for several users, anomalies are already excluded such as
- variations due to user’s state of mind/ well-being or other factors that influence such anomalies
- it also excludes machine performance variability related anomalies.
Once these anomalies are gone, we can start measuring the complexity of a task (PIQ value) and thereby determine the automatability. But before that, a little formal definition is required. Here we define the properties of PIQ –
PIQ does not depend on a single user’s ability to perform a task. It
does not indicate complexity from the users standpoint, but indicates the automatability
To determine PIQ, we first determine PIQ for a single step (PIQs). PIQs is a function of
- “Average normalized time spent (Tn)” on the step by user – positive impact on PIQ
- Length of input (Ni) for the step – U shape impact on PIQ
- Length of output (No) for the step – positive impact on PIQ
- Category of the step being performed (Cs) – variable impact on PIQ.
Some steps are more complex than others from a machine standpoint. For example, a step categorized as ‘copy paste’ does not have significant impact on length of input or output. On the contrary, a step categorized as ‘RCA’ has major impact on those parameters.
PIQs = f (Tn, Ni, No, Cs)
PIQ = g (max (PIQs), Nv, Ns) where
max (PIQs) is the maximum PIQs observed among all the steps within the task
Number of steps performed (Ns) – positive impact on PIQ
Number of variants (Nv), or number of decision points in the task – positive impact on PIQ
The following diagram may help.
The interaction data is analyzed to recognize the type of step, i.e. whether it is a copy paste type of step or at this step root cause analysis is being performed by the user. The parameters for the step that is currently analyzed are extracted, which will later be used to calculate the PIQs of the step. To find PIQs machine learning models can be used to rank the features (identified before) that will be most appropriate for the process data available. Over-fitting and multicollinearity challenges will be addressed by this action. The selected features become the independent variables and the PIQs of the step is derived using multiple regression mechanism. The decision to use supervised on unsupervised learning for f() and g() function will depend on availability of labelled data and volume/quality of data. Moving forward, parameters for task are obtained, but this time weights need to be applied to the parameter for the task. The max (PIQs) is used as it provides the lower bound on the complexity of PIQ.
Conclusion – This is a useful technique in Automation for Automation(A4A) category to recognize new automation opportunities without much involvement on human analysts. This is an essential mechanism that will enable 360 degrees of automation.
Authors
Caesar Chatterjee Data, Technology & Automation Architect
Priyankar Malakar Data Scientist & Automation Evangelist
