From Robotic Process Automation
to Intelligent Process Automation^††thanks: Authors are in alphabetical order.
– Emerging Trends –

Business Process Management (BPM) is a multi-disciplinary field that supports the management of business processes with some combination of modeling, automation, execution, control, measurement and optimization. BPM involves business activity flows (workflows), systems, and people such as employees, customers and partners within and beyond the enterprise boundaries.

Business processes in reality have a wide scope from the traditional rigid processes (modeled and running under the supervision of a strict workflow management system) to completely ad-hoc unstructured flows driven by humans over e-mail, chat and phone. Traditional BPM systems, at one end of this spectrum, demand a process model that can be completely defined in advance and typically include restrictions such as rigid control flow and context tunneling [39]. Case management is closer to the other end of this spectrum: a case consists of people, documents, and tasks [64]. Flexible ordering of task execution is enabled through Event-Condition-Action rules as well as the ability for a user to add new (ad-hoc) tasks. A task itself may be defined as a fully structured workflow, making case management a hybrid model.

Businesses seek to support growth while maintaining low costs by automating repetitive and time consuming tasks, especially seeking to eliminate costly and error-prone manual steps. Business Process Automation (BPA) seeks to improve the efficiency of business processes in terms of cost, resources and investment through automating the management of relevant information and data, the time spent by team members, and the execution logic.

RPA is an emerging technology in BPA that creates software robots that perform tasks previously done by humans. RPAs are one form of BPA implementation that is specifically focused on repetitive workflows. The overall goal of RPA is to provide the shortest route to automation by introducing a user interface automation layer rather than interacting deeply with the application code, system or database that are behind those applications.

Performance measurement in business processes is the first step for analyzing and monitoring the process health and progress for process automation. Identifying the right measures for process performance is extremely important. Performance of a process measures how well the process is doing with respect to chosen indicators. Examples of such performance indicators can be time to execute a task, cost per task in terms of employee head count or number of approved loans [46]. Numerous authors have proposed a fixed number of category classes for indicators in order to provide a structure. The majority of authors, including [58], proposed a process-oriented view of the indicators which resulted in four groups of performance indicators: quality, time, costs, and flexibility. As [1] indicates, better processes contribute to meeting the strategic objectives of an organization. Therefore, we specifically call out another association with respect to existing groups of indicators that focuses on the impact of all the indicators towards the business goals. Such indicators can be, for example, analyzing the process performance indicators with respect to profit and revenue of the organization.

In practice, most business process owners focus on productivity measures related to time and cost. The challenge with flexibility and quality measures is that they are difficult to standardize, optimize, implement and generalize. It is common to use indicators related to a process’s utilization and assignment of resources, such as repeated tasks or time to execute a task. Resources are usually aligned with the length of the task to avoid bottlenecks in the process. However, shortening the time of task execution or lowering the cost of resources does not necessarily yield better business outcomes. Therefore, it is important to consider methods for measuring process performance that can assess the impact of indicators on business goals and outcomes.

Adopting the right measures is crucial to the success of BPA. They would then be used to evaluate the performance of RPAs and other automation solutions, allowing business users to assess the effectiveness and return on investment of these solutions. Furthermore, these measures can be used by the RPAs themselves to iteratively improve their performance using machine learning models.

While the use of automation has been gaining traction across many industries, its incorporation into business processes poses several challenges. Automation capabilities such as RPAs can provide transformational benefits, however, it is unclear where their use can provide the highest value. Tools and analytic approaches for identifying high value automation opportunities in a process are still nascent. As we discussed previously in Section 1, recent innovations in machine intelligence stand to disrupt this field through the digital transformation of business processes. Throughout the rest of this paper, we will take the reader along this journey, starting from the state of the art in RPAs, through the vision and promise of IPAs, to the major challenges to be overcome to reach this goal. Finally, we will conclude with a quick overview of a recently concluded workshop on this topic.

Implementing an RPA requires a costly manual analysis of the tasks performed by the users either by observing their behavior, which does not scale when there are hundreds of processes, or through careful analysis of process-related documentation, which can be outdated. Finding opportunities to automate tasks that are more complex than routine repeatable tasks requires the use of structured and unstructured data from process logs. There have been several research efforts to identify candidate tasks for automation [38, 29] from textual descriptions, but they focus on particular business domains (e.g generating utility bills) and are still not implemented at scale. Identifying automatable tasks only solves part of the problem. The results should also be augmented with a recommendation of possible IPA templates or AI models that are suitable to automate these tasks, as in [35].

Unlike RPAs whose overall potential results in a significant increase in turnaround time and cost savings of up to 30% [32], there is a higher cost associated with developing IPAs. To build the next generation of IPAs requires data preparation (identifying relevant data, and cleaning and transforming it) and feature engineering (extracting appropriate features), before building and validating the AI capabilities. Similarly, there is a higher cost associated with maintaining IPAs in comparison with RPAs. In addition to the deployed code, the AI capabilities within have a lifecycle of their own.

The AI models must be retrained in response to changes in the business process (control flow drifts) or changes in the data (data drifts). These higher costs require larger return on investment for IPAs to be suitable. Some ways to mitigate the cost required to build and maintain IPAs include decreasing the effort to develop them, enabling them to be reused for different types of processes, or using them to replace/augment different customer tasks.

Adopting IPAs comes with an added risk of monetary or reputation loss. For example, data used for training may be manipulated or contain implicit racial, gender, or ideological biases [67]. In addition, business users are risk-averse and do not implicitly trust AI models. Mitigating the risk of deploying the AI models requires staged deployment techniques such as canary deployments, bandit services, and A/B testing [49]. Increasing business users’ trust will require a variety of solutions including maintaining action provenance for audits and providing explanations for any automated IPA decisions.

IPA research currently focuses on non-routine tasks. Handling more complex tasks will require the composition of multiple IPAs, as well as the collaboration and coordination of these IPAs. To achieve this, new frameworks need to be developed that enable IPA cooperation. Previous research efforts to use multi-agent systems in BPM [52, 12] need to be adjusted and revised for IPAs. Frameworks must now take into account the diversity of automation tasks and domains, and the fact that RPAs can be created by different developers without shared development guidelines. Maintaining compatibility between RPA and business process versions as each co-evolves is also crucial. Finally, a unified interface such as a conversational system may also be required to facilitate the interaction among IPAs and end users [57].

The BPM literature offers a variety of AI solutions to cluster process traces [50, 51] for better process discovery, predict business outcomes [9, 33], and provide decision support [42]. Deep learning models, including those in the NLP domain have also been applied [62, 15]. Recent efforts attempt to discover automatable routines from user interaction logs [8]. Unfortunately, due to the reasons mentioned above, very little of these innovations have been applied and adopted by enterprises [13], and those adopted are limited to narrow domains such as customer service, enterprise risk, and compliance [66]. Solutions need to take into account the structure of these highly regulated domains that require paper trails of all transactions and must adhere to privacy and security laws.

An area of AI that is readily applicable to business processes is automated planning, which concerns itself with generating sequential courses of actions or plans from its declarative components and thus provides a powerful framework for sequential decision making in a BPM system. [44] provides an overview of existing work and challenges at the intersection of planning and BPM. Perhaps the most important (and natural) among them is the specification and synthesis of business processes in the form of planning problems [54]: the planner composes workflows on demand automatically based on the components specified by the process author or workflow designer.

A particular area of interest here is that of the composition of automated services for the optimization of a business process [14, 5, 10]. This work is motivated by research on “web service composition” [53]. We refer to [56] for a comprehensive summary of work in this area, while [61, 60] provides an overview of many of the challenges involved.

Another key application of planning to business process management is in the prediction of how different process components will evolve over time, thereby anticipating possible risks. Generative model-based approaches such as planning are uniquely situated to do this, finding applications in the robustification and adaptation of processes to failures [27], validation, verification, and monitoring of processes [37], and so on. A particular useful tool towards achieving this is referred to as top-k and diverse planning [30] where a set of solutions are computed instead of a single one thereby allowing one to anticipate likely ways a process may evolve. Such approaches have found many applications¹¹1http://ibm.biz/ai-scenario-planning in enterprise risk management and scenario planning recently.

Reducing the need for direct human involvement with the business process is one of the main goals of automation. There is a very strong trend of automating people-driven processes to chatbot interactions throughout the industry [22]. According to Gartner ²²2https://www.gartner.com/imagesrv/summits/docs/na/customer-360/C360_2011_brochure_FINAL.pdf, by 2020 customers will manage 85% of their relationship with the enterprise without interacting with humans. Conversational interfaces apply not only to customer facing businesses but also to employee services such as help-desk and support bots which have been deployed within almost all enterprises. The focus has also expanded to carrying out a business process with a conversational agent [41], or automating tasks such as placing orders, paying and following up invoices, repetitive data base queries, external service inquiries and automatic analytics and reporting.

Chatbots bring ease of access to all these applications in one interactive mode. The natural language interaction helps democratize access to data, automation, and analytics for a broader range of business users enabling faster adaptation by users and greater personalization  [69]. Chat interactions also serve as a rich source of data to mine for additional automation candidates, closing the loop on bringing intelligence into process automation.

Introducing AI into mission-critical business applications can be a risky endeavor. The software engineering community has developed formal methods to verify the correctness of programs in critical systems [68], but these techniques are not applicable to learned AI models. Approaches to improve the interpretability and explainability of AI models are a more promising avenue [2]. For example, knowing why a model recommended denying a loan to an applicant is important to ensure adherence to anti-discrimination regulations [20]. These approaches, however, need to be expanded along at least two dimensions. First, existing interpretability techniques such as perturbation-based methods or interpretable proxy models [2] need to be augmented with domain knowledge of the business process, including the control flow semantics, decision rules, and business objects, thereby leading to more complete and accurate explanations [26]. Second, the explanations need to be targeted at non-technical subject matter experts. Statistical measures of feature importance or Shapley values are useful to data scientists but do not give actionable insights to a loan officer or process owner. The explanations need to be tailored to the business user, including using the business domain vocabulary and concepts as well as taking into account the context of the user’s needs and preexisting knowledge.

We posit that business process data should be considered a new modality in machine learning, similar to image, text, audio, or video. At the very least, it should be treated as a multi-modal domain [7]. A non-exhaustive list of the different types of data embodied in a business process includes the control flow (graph structure), the execution of a process trace (sequence of events), the metadata associated with an event (multi-dimension set of attributes), references to unstructured documents (images or text), interactions between participants (both graph and time series representations), and the social networks (also graphs).

Many existing techniques exploit one or more of these data structures to extract insights or build predictive models [50, 51, 42, 62, 15]. We believe that a more principled approach to unify these different sub-modalities of business processes will accelerate research in this area. Reifying business process data as a distinct modality opens up a number of research questions for the machine learning community including developing novel techniques for representation learning, explainability, and transfer learning for this new modality.