Why do business processes deviate? Results from a Delphi study

BPM is the science and practice of overseeing how cross-functional work is performed in organizations to ensure consistent outcomes and to seize improvement opportunities (Dumas et al. 2018). BPM strives for two objectives, i.e., improving business processes and developing the BPM capability itself (Rosemann and vom Brocke 2015). From a lifecycle perspective, BPM comprises activities such as the identification, definition, modeling, implementation, execution, monitoring, controlling, improvement, and innovation of business processes (Dumas et al. 2018). Having all business processes of an organization in scope, BPM provides an infrastructure for effective and efficient work (Harmon 2016). Business processes, which are the central unit of analysis of BPM, include events, tasks, and decisions that involve process participants and resources to create valuable outcomes for customers (Dumas et al. 2018). Despite contrary empirical evidence, BPM research has long time been working on the tacit assumption that business processes can be specified exhaustively prior to execution in terms of process models and that they are executed as specified. Only recently business processes have been recognized as drifting information structures with partly emerging behavior that cannot be entirely controlled by process managers, a paradigm shift that further stimulated research on process deviance (Beverungen 2014).

In the literature, business processes are classified based on different criteria. The most common classification is that into core, support, and management processes, splitting business processes according to their role in corporate value creation (Armistead 1999; vom Brocke et al. 2016). Core processes are business processes whose customers are from outside the organization and willing to pay for products and services. Support processes ensure that core processes continue to function, whereas management processes plan, organize, communicate, monitor, and control corporate activities (Harmon 2016). Another popular classification, which stems from organizational science, focuses on how business processes deal with variation (i.e., deviance from objectives) and variety (i.e., the number of process variants) (Lillrank 2003). Accordingly, processes are split into standard, routine, and nonroutine processes. This classification is similar to Johnston et al. (2012) distinction between runner, repeater, and stranger processes, the difference being that Johnston et al. also account for volume (i.e., the number of executions). Encompassing a single variant with a defined input and output, standard processes are very simple. As combinations of standard processes, routine processes feature an arbitrary, but fixed number of variants that can be specified prior to execution via imperative process models, e.g., following the Business Process Model and Notation (BPMN) or Event-driven Process Chain (EPC) standards. Routine processes cover the predictable part of organizational behavior. In routine processes, the search for errors and solutions can proceed analytically and logically. In contrast, nonroutine processes deal with semi- or unstructured problems. They entail high uncertainty and degrees of freedom, covering the unpredictable and emerging part of organizational behavior. With the input and output of nonroutine processes being at least partially undefined, they cannot be exhaustively specified prior to execution in terms of imperative process models. If at all, this can be done via declarative process models or process flexibility strategies such as flexibility-by-underspecifiation (Schonenberg et al. 2008; di Ciccio et al. 2017). Due to their novelty, however, declarative process models are neither as mature nor as broadly used in industry as imperative process models. Importantly, although nonroutine processes are hard-to-capture via process models, they can show different behavior than intended, and thus be deviant. As the distinction between routine and nonroutine processes is more closely related to process deviance than the classification into core, support, and management processes, we adopted it as analytical lens to answer our second operational research question. Further, the distinction between routine and nonroutine processes is relevant from an IS perspective, as the digital age entails a shift from routine to nonroutine processes (Swenson 2010; Bennett and Lemoine 2014).

The BPM literature considers process deviance from various perspectives (Mertens et al. 2016b). It also discusses closely related concepts such as exceptions, workarounds, or non-compliant processes (Rinderle and Reichert 2006; Alter 2015a, b). Below, we look at deviance from an organizational psychology perspective, define process deviance for our purposes, and discuss related concepts.

In organizational psychology, Robinson and Bennett (1995) define deviant behavior as actions that differ from norms and damage an organization. Spreitzer and Sonenshein (2004) specify deviant behavior as behavior that intentionally differs from organizational norms. Compared to Robinson and Bennett (1995), they conceptualize deviance by abstracting from the negative effects typically associated with deviance. Thereby, they coined the term positive deviance as opposed to negative deviance. We critically reflect on the conceptual distinction between deviance and its effects in more detail below. Spreitzer and Sonenshein (2004) further distinguish between supra-conformity as well as a reactive, statistical, and normative approach to deviance. Supra-conformity, also referred to as excessive conformity, means that behavior is deviant if it exceeds boundaries deemed appropriate. The reactive approach means that behavior is deviant if it is criticized by a reference group. In line with the theory of variation, the statistical approach defines deviance as deviation from an expected average or another predefined value (Deming 1994). Finally, the normative approach defines deviance as deviation from accepted norms (Spreitzer and Sonenshein 2004).

Many ideas about deviant behavior from organizational psychology have been adopted in the context of process deviance (Müllerleile et al. 2015; Mertens et al. 2016a, b). When reviewing the literature, we found criteria that help define process deviance, which currently features no commonly accepted definition, for our purposes. These criteria are frequency, scope, and intention. Frequency captures whether deviant behavior occurs in one process instance, various or all process instances (Dumas and Maggi 2015). Scope covers whether deviant behavior occurs in individual tasks, sub-processes, or the entire process (Alter 2014). Finally, intention reflects whether deviant behavior occurs intentionally or unintentionally (Depaire et al. 2013; Mertens et al. 2016b). On this foundation, we propose and used the following definition of process deviance in line with our understanding that process deviance is an umbrella concept as illustrated below:

Process deviance indicates that a business process shows different behavior than intended. It may occur in individual tasks, sub-processes, or the entire process (scope). Process deviance may occur in one process instance, various or all process instances (frequency). Finally, it may also occur intentionally or unintentionally (intention).

Process deviance can affect process performance positively or negatively (Spreitzer and Sonenshein 2004; Alter 2014). Cases where process deviance positively influences process performance (e.g., less compliance violations, less volatility, or better average values of performance indicators) are referred to as positive or constructive process deviance, while cases associated with negative effects are called negative deviance (Chakraborty 2013; Dumas and Maggi 2015; Andrade et al. 2016; Delias 2017). As process deviance should not be mixed up with related performance effects, we discuss both concepts and its effects separately as far as possible.

The BPM literature discusses concepts such as exceptions, workarounds, and non-compliant processes. Admittedly, these concepts cannot be separated accurately as they stem from different research streams. However, we classify them using the criteria from above to provide clarity as far as possible (Table 1). Exceptions are interruptions of tasks or process executions due to expected or unexpected events (Eder and Liebhart 1998; Mišic et al. 2010). Exceptions typically occur in one instance or various instances as well as in individual tasks or sub-processes. Further, they occur mostly unintentionally. Exceptions occurring in all process instances are referred to as errors. Captured and handled appropriately, exceptions enable runtime flexibility and avoid process deviance (Simon and Mendling 2006). Otherwise, they reflect different behavior than intended. Workarounds are intentional and goal-orientated adaptations, improvisations, or best practices in work systems that relate to single tasks or sub-processes (Alter 2014). They can be seen as responses to barriers during process execution (Röder et al. 2014). Workarounds can occur once, sometimes, and always. Finally, process compliance generally requires business processes to be executed in conformance with predefined specifications (e.g., process models, business rules, or external regulations) (Ferneley and Sobreperez 2006; Weidlich et al. 2010; Becker et al. 2011; Alter 2015b). Non-compliant processes violate such specifications. Non-compliant processes can occur once, sometimes, and always as well as intentionally and unintentionally (Alter 2015b). Non-compliance can also occur in individual tasks, sub-processes, or the entire process. Concluding, the term non-compliance is conceptually very similar to process deviance and can be used synonymously. Exceptions have a more technical background and workarounds have been introduced from a work systems perspective. They represent special forms of process deviance.

To answer our operational research questions, we conducted a Delphi study with international experts from industry and academia. Exploratory in nature, Delphi studies strive for consensus on a specific topic among a group of experts (Dalkey and Helmer 1963; Keeney et al. 2006). Thereby, experts provide their opinions on the topic in focus and comment on the opinions of other experts throughout multiple rounds (Dalkey and Helmer 1963). Our Delphi study strived for identifying and structuring reasons for process deviance from a process manager’s perspective. It also investigated the identified reasons’ importance for causing deviance in routine and nonroutine processes (Lillrank 2003).

In recent years, guidelines and rigor criteria for Delphi studies have been proposed (Okoli and Pawlowski 2004; Keeney et al. 2006; Paré et al. 2013; Skinner et al. 2015). We accounted for this knowledge and elaborate on this in the next sections. In line with Dalkey and Helmer (1963), Delphi studies include panelists, ensure end-to-end anonymity, reach consensus iteratively via several rounds, and incorporate the panelists’ feedback in a structured manner. Panelists are experts who provide their opinion and comment on the opinion of other experts (Okoli and Pawlowski 2004). Experts can be organized in one or more panels. Ensuring anonymity is essential to avoid bias as, for instance, panelists with a self-confident appearance may influence other panelists (Okoli and Pawlowski 2004; Skinner et al. 2015). In addition, Delphi studies take multiple rounds to capitalize on the creativity and expertise of all panelists. In each round, the panelists should be provided with structured feedback such that they can trace progress and assess whether their input has been reasonably integrated into the overall position of the panel. Panelists should also be allowed to change their opinion and to provide feedback on both (intermediate) results and the study at large. Thereby, it is vital that panelists are not forced to adjust their previous responses (Paré et al. 2013).

The exploratory and iterative nature of the Delphi method fits our research questions. As knowledge on reasons for process deviance is immature, an exploratory research method is appropriate (Andrade et al. 2016). Further, an iterative method enables experts to benefit from the feedback of other experts and to think about reasons for process deviance as well as their importance repeatedly (Okoli and Pawlowski 2004). Thus, we preferred the Delphi method over other methods such as exploratory interviews where experts are involved only once and sequentially. The perspective of process managers, which we chose in our Delphi study for the reasons outlined in the introduction, can be suitably addressed by the Delphi method (Paré et al. 2013). This is because experienced experts from industry and academia can put themselves in the position of process managers and judge reasons for process deviance accordingly. However, in case process deviance is studied from complementary perspectives (e.g., the perspective of process participants), other research methods (e.g., observations, interviews, or ethnographies) are more suitable. As we will discuss in the conclusion, confirmatory research methods (e.g., surveys) are most appropriate to validate the results of our exploratory Delphi study.

To structure our Delphi study, we followed the blueprint of ranking-type Delphi studies as proposed by Schmidt (1997), which is the most commonly used Delphi blueprint in IS research (Paré et al. 2013). Correspondingly, our study included three phases: brainstorming, narrowing down, and rating. As we did not rank, but rate reasons for process deviance with respect to their importance for causing process deviance in the third phase, we refer to our study as a rating-type Delphi study. While the brainstorming and narrowing down phases helped answer the first operational research question, the rating phase covered the second research question. We preferred rating (i.e., the assignment of reasons to predefined ordinally scaled importance categories) over ranking (i.e., the assignment of reasons to ordered ranks in line with their importance), as we were interested in the reasons’ importance for causing deviance in routine and nonroutine processes as well as in potential differences between routine and nonroutine processes. We were not interested in the reasons’ relative importance for causing deviance compared to one another for either process type (Keeney et al. 2006). The rating of reasons enables multiple reasons to be assigned to the same rating category. Further, ranking is only feasible for a small number of items. However, restricting our results to a few reasons for process deviance would not have done justice to the multi-faceted nature of business processes. Table 2 provides an end-to-end perspective on our Delphi study with all relevant statistics, which we explain below.

In the brainstorming phase, we collected an initial list of reasons for process deviance (round 1), created an initial and aggregated coding, and asked the panelists to comment on the resulting medium list (round 2). We also proposed categories for structuring the reasons. In the narrowing down phase, each panelist had to select the most important reasons from his or her perspective, independent from specific contexts (round 3). On this foundation, we compiled a shortlist of reasons for process deviance. In the rating phase, the shortlisted reasons were rated regarding their importance for causing deviance in routine and nonroutine processes (rounds 4–6). Whereas the number of Delphi rounds is predetermined for the brainstorming and narrowing down phases (i.e., two and one rounds, respectively), it depends on three termination conditions in the rating phase (Paré et al. 2013). The rating phase terminates if consensus has been reached among the panelists (Okoli and Pawlowski 2004), if the expert assessment does not change significantly between subsequent rounds, or after at least three rounds so as not to overstrain the panelists. It is sufficient that one of these termination rules holds true (Paré et al. 2013).

In our Delphi study, we conducted six rounds. The panelists had 1 week per round to provide their answers, either per email or an online questionnaire. In each round, the panelists could provide open-ended feedback, and we provided them with their input from the previous round, a summary of relevant changes, and detailed instructions. We asked for the panelists’ satisfaction with the study (overall satisfaction) as well as with the presented reasons and categories (coding satisfaction), using a Likert scale ranging from 1 (fully unsatisfied), 2 (strongly unsatisfied), 3 (unsatisfied), 4 (neutral), 5 (satisfied), and 6 (strongly satisfied) to 7 (fully satisfied) (Schmiedel et al. 2013).

Panelists should be motivated and have great expertise (Delbecq et al. 1975; Keeney et al. 2006). To recruit and manage experienced panelists, we prepared a knowledge resource nomination worksheet that captures selection criteria (Okoli and Pawlowski 2004). To ensure a broad view on BPM and process deviance, we included panelists from academia and industry as recommended and common practice (Okoli and Pawlowski 2004). We used the following selection criteria: Panelists from academia had to be specialized in BPM and at least hold a PhD. These criteria ensure that involved academic panelists had great domain expertise and a broad overview, which enabled them to complement the experience of experts from industry (Okoli and Pawlowski 2004). Grounded on their research activities, some academic experts also had valuable insights into the BPM practices of several organizations. Industry panelists had to hold a key role in their organization’s BPM function or in BPM consulting, and they should have at least 5 years of work experience (Okoli and Pawlowski 2004; Schmiedel et al. 2013). This ensures that they are familiar with the perspective of process owners. To identify panelists, we used our industrial and academic networks. We also invited authors from vom Brocke and Rosemann’s ‘Handbook on Business Process Management’ (vom Brocke and Rosemann 2015a, b), recognized for its comprehensive view on BPM (Schmiedel et al. 2013).

To identify the most suitable experts for our Delphi study, all prospective panelists were ranked according to the derived selection criteria (Okoli and Pawlowski 2004; Keeney et al. 2006). Overall, we invited 120 experts from 11 countries. These experts were asked to nominate further experts (Okoli and Pawlowski 2004). In the beginning, 40 experts committed themselves to participate in our study. This corresponds to a response rate of 33%. Due to no shows and drop-outs, between 22 and 33 panelists participated in the single rounds. This amounts to an initial no show ratio of 27.5% and a fraction of involved panelists between 82.5 and 66.7% (Table 1). The size of our panel complies with the guidance available for Delphi studies (Paré et al. 2013). With business processes and process deviance being multi-faceted phenomena, a large panel size also fit the domain in focus. The participating panelists were physically distributed with a geographical focus on Germany, the US, and Australia. All panelists met the selection criteria. Almost all panelists from industry even exceeded our requirements, having more than 10 years of work experience and holding leadership roles. Please have a look at Appendix A-1 for more information about our panel.

As recommended, we conducted a pilot study to ensure an appropriate study design (Skinner et al. 2015). The pilot study was done with test panelists that met the same requirements as for the main study. To offset potential bias, we excluded the test panelists from the main study. When preparing the pilot study, we identified two ways for setting up the brainstorming phase. The first option was a greenfield approach where reasons for process deviance were collected, providing the panelists without additional information (Schmidt 1997). The second option was to provide the panelists with an established framework to structure their brainstorming efforts (Kasiri et al. 2011). For this purpose, we identified Rosemann and vom Brocke’s (2015) BPM capability framework as suitable, as it is well known and takes a holistic broad perspective on BPM (Kerpedzhiev et al. 2016; Schmiedel et al. 2013). Although Rosemann and vom Brocke’s framework focuses on an organization’s BPM capability, its core elements (i.e., strategic alignment, governance, methods, IT, culture, and people) also apply to individual processes, if the ‘methods’ core element is re-conceptualized as ‘process design’.

In the pilot study, we prepared a questionnaire for both options. In the first questionnaire, the test panelists were asked to name and briefly describe at least fifteen reasons for process deviance. In the second questionnaire, the test panelists had to provide five reasons per core element of the BPM capability framework. Groups of three and two test panelists used the first and second questionnaires, respectively. The test panelists who used the first questionnaire were unable to list fifteen reasons, starting to repeat reasons after about ten items. The panelists who used the second questionnaire experienced problems in applying the framework. One industry panelist could not answer the second questionnaire as he was unfamiliar with the framework, a situation we would have been unable to avoid in the main study. In line with these results, we decided to design the brainstorming phase as a greenfield approach. Further, we asked each panelist to name only between six and ten reasons in the brainstorming phase to focus on the most important reasons for process deviance (Schmidt et al. 2001).

In the first brainstorming round, we provided the panelists with a description of the end-to-end structure of our Delphi study as well as with detailed definitions and instructions. In line with the results of the pilot study, the panelists had to name between six and ten reasons why processes show different behavior than intended (Schmidt et al. 2001; Okoli and Pawlowski 2004). To support the subsequent coding and consolidation, we also asked the panelists to provide a short description per reason. We also used the first questionnaire to gather further information about the panelists and to confirm that they meet the selection criteria. In this round, all panelists received identical questionnaires (Okoli and Pawlowski 2004). We received 192 responses from 29 panelists.

To consolidate the panelists’ individual input into an integrated position of the panel, we coded all responses using iterative coding (Schmidt 1997; Okoli and Pawlowski 2004; Krippendorff 2013). Before that, one co-author anonymized all responses. Thus, the panelists were anonymous not only for one another, but also for the research team. As recommended by existing Delphi guidelines, the other co-authors first coded the panelists’ responses independently (Paré et al. 2013). We initially identified identical reasons and merged explanations (Schmidt et al. 2001; Okoli and Pawlowski 2004). We then identified categories and merged reasons with similar descriptions (Delbecq et al. 1975; Schmidt et al. 2001; Okoli and Pawlowski 2004; Paré et al. 2013). In a final workshop series, all co-authors discussed and consolidated their individual coding results. All reasons were formulated in a domain-agnostic manner and with a positive polarity, meaning that reasons point to process deviance if they can be confirmed. After this initial coding, we had 91 reasons and 9 categories.

To validate these results, we performed a second brainstorming round (Schmidt et al. 2001; Keeney et al. 2006). This time, we asked the panelists to provide feedback on each reason and category. To provide the panelists with sufficient guidance, we presented them their input from the first round. To avoid bias, we implemented a randomized questionnaire, i.e., the order of categories and reasons per category were randomized (Paré et al. 2013). The panelists were also allowed to name new reasons based on their own or ideas and those of other panelists. We sent the questionnaire to 39 panelists, as 1 expert withdrew in the first round, and 27 panelists provided feedback. We distributed the questionnaire to all panelists who had initially agreed to participate in our study to leverage as much creativity and expertise as possible. As the average coding satisfaction was 5.63 on the seven-point Likert scale, the panelists were almost strongly satisfied with the consolidated reasons and categories (Table 1). Again, all co-authors evaluated the panelists’ anonymized feedback independently and then consolidated their individual evaluations. Not to get biased by individual panelists, reasons or categories were only changed or deleted if this was suggested by several panelists. Reasons were added, if they were not included in any existing reason. Importantly, our intention was not to keep the number of reasons for process deviance small in this round, as this was the objective of the narrowing down phase. Finally, we reached an agreement on 61 reasons on the medium list, structured into ten categories.

The narrowing down phase consisted of one round, aiming to identify the most important reasons for process deviance. Another purpose is to reduce the number of reasons to a manageable number. We asked the panelists to select those reasons that are both in accordance with their individual expert judgement and independent from distinct contexts (i.e., routine and nonroutine processes), most important for causing process deviance (Schmidt 1997). As typical for Delphi studies, we did not provide the experts with a formal definition of importance, but left the assessment up to their judgement. Each panelist had to select 20 reasons, as this number has been recognized as manageable and comprehensive in previous Delphi studies (Schmidt 1997). The idea behind narrowing down is that only reasons that exceed a minimum number of votes remain on the shortlist used as input for the rating phase (Schmidt 1997; Paré et al. 2013). In contrast to some other Delphi studies, we decided not to split the panel into an academic and an industry panel. We did so as we were interested in commonalities and differences between routine and nonroutine processes in terms of the reasons’ importance, which we analyzed in the rating phase, instead of commonalities and differences regarding the judgement of academics and practitioners. If we had split the overall panel into academics and practitioners, we also would have run the risk that the final panels became too small.

The randomized questionnaire of round three was distributed to 39 panelists. Analogous to the second round, we invited all initially committed panelists, as panelists did not require specific knowledge from prior rounds to select the most important reasons. In this round, 33 panelists provided feedback. The satisfaction with the coding and the panelists’ overall satisfaction with the study increased to 5.73 (Table 1). Based on the panelists’ responses, we shortlisted all reasons voted for by at least ten panelists, i.e., 30% of the panel (Paré et al. 2013). We shared this cut-off criterion and let it approve by the panelists. On this foundation, 34 reasons remained on the shortlist. One reason voted for by ten panelists (i.e., exactly the threshold) received qualitative comments about redundancies with another reason and was dropped after a discussion within the author team. Further, we dropped one category (‘IT changes’) as all related reasons received too few votes.

Consequently, the narrowing-down phase resulted in a shortlist of 33 reasons structured along nine categories. As we did not only identify reasons for process deviance, but also structured them into categories, the results of the narrowing down phase comply with other Delphi guidelines, which recommend the number of shortlisted items not to exceed 30 (Paré et al. 2013). Further, the reasons and categories answer our first operational research question, which strived for identifying and structuring reasons for process deviance independent from specific contexts.

The rating phase sought to answer our second operational research question, asking how important the identified reasons are for causing process deviance in different contexts. To do so, we let the experts rate all shortlisted reasons regarding their importance for causing deviance in routine and nonroutine processes (Schmidt 1997). The panelists had to indicate the reasons’ importance for both process types separately using the following ordinal scale: A (extremely important), B (very important), C (important), and D (unimportant). Just like in the narrowing down phase, we did not provide the experts with a definition of importance. As the rating phase included several rounds where panelists were confronted with their own rating as well as the aggregated rating distribution of the panel, the experts’ individual understanding of ‘important’ was calibrated throughout these rounds. We chose the mentioned ordinal scale to seamlessly connect to the narrowing-down phase, which resulted in a shortlist of generally important reasons for process deviance independent from specific contexts. Thus, the first three rating categories related to different degrees of importance. We included rating category D (unimportant) for cross-checking purposes because it may have been the case that individual panelists wanted to assess reasons as unimportant for causing process deviance in either routine or nonroutine processes. As can be seen in Table 3, this was indeed the case, but very seldom. If this circumstance had occurred often, we would have had to go back to the narrowing down phase. Further, if we had not included rating category D, the panelists would not have been able to declare reasons as unimportant, a circumstance that would have biased our rating results. We informed the panelists that rating category D was only included for cross-checking purposes, ensuring that they did not perceive the rating scale as non-equidistant. As outlined in Sect. 3.1, the key advantage of a rating of reasons is that multiple reasons can be assigned to the same rating category and that larger items sets can be handled.

The questionnaire of round four was distributed to those 33 panelists who participated in the narrowing down phase. We decided to invite only those panelists who had at least participated in the third round to avoid discussions about the shortlisted reasons in a late and structured Delphi phase, not designed to account for qualitative feedback on the identified reasons. Thereby, we ensured that all panelists were familiar with the identified reasons and categories and had approved the shortlist. The questionnaire included the shortlisted reasons, definitions, and examples of routine and nonroutine processes as provided by Lillrank (2003), as well as a description of the rating scale. In this round, 27 panelists provided feedback. The overall satisfaction and the coding satisfaction, which we collected for the last time in this round, increased to 5.81 and 5.78, respectively (Table 1). This coding satisfaction needs to be attributed to the shortlist of reasons, as the panelists were confronted with the entire shortlist in this round the first time. As the coding satisfaction was very close to strongly satisfied and had continuously increased during the previous rounds, we considered this as content-wise consensus regarding the shortlist of reasons for process deviance. That is, the shortlist has been approved by the panelists and constituted a solid foundation for the rating phase.

As a constitutive characteristic, rating-type Delphi studies do not only strive for content-wise consensus of shortlisted items in the narrowing down phase. Rather, in the rating phase, they also strive for quantitative consensus in terms of stable rating distributions regarding the shortlisted items’ importance for different contexts (Paré et al. 2013). However, as the Delphi method is exploratory in nature, the quantitative results of the rating phase should be interpreted as trend statements. They should be validated through both further exploratory (e.g., semi-structured interviews) and confirmatory research methods (e.g., surveys). We get back to this need for future research in the discussion.

To support the quantitative assessment of the rating distributions’ stability, we applied Kendall’s W, an accepted consensus measure typically applied in ranking-type Delphi studies (Schmidt 1997). In line with the nature of the Delphi method, even the rating phase must not be terminated only based on quantitative criteria. Rather, the termination decision should also account for qualitative insights such as provided by the panelists’ feedback. It is important to note that we used Kendall’s W for measuring changes in consensus, not for measuring absolute consensus. We did so by comparing the Kendall’s W values for subsequent rounds. The reason is that we did not rank the identified reasons in relation to one another, but rated them individually. However, Kendall’s W only measures absolute consensus for rankings. Looking at changes in consensus is reasonable when striving for consensus in terms of stable rating distributions at the end of a Delphi study, i.e., minor changes in the rating results of two subsequent rounds (von der Gracht 2012). In round four, Kendall’s W was 0.08 and 0.06 for routine and nonroutine processes.

The questionnaire of the fifth round included the shortlist of reasons for process deviance as well as the aggregated rating distributions of the fourth round in terms of bar charts highlighting the rating category (modus) that received the most votes (Table 3). We also repeated the definitions and examples of routine and nonroutine processes to remind the panelists of the characteristics of these process types. In addition, we provided each panelist with his or her individual rating from the previous round (Schmidt 1997; Paré et al. 2013). We informed the panelists that they can but need not adjust their prior voting. In this round, 22 panelists provided feedback. The overall satisfaction increased to 6.05, i.e., the panelists were strongly satisfied (Table 1). Kendall’s W doubled to 0.15 and 0.22 for routine and nonroutine processes, respectively. Due to this substantial change compared to the previous round, we performed a sixth round.

In round six, the panelists received the same questionnaire as in round five, with updated information. In this round, 25 panelists provided feedback. The panelists continued to be strongly satisfied, with an overall satisfaction of 6.08 (Table 1). Taking a value of 0.20, Kendall’s W increased only slightly for routine processes. With a value of 0.21, Kendall’s W slightly decreased for nonroutine processes. As the rating distributions changed only slightly between the fifth and the sixth rounds and as we had already conducted three rating rounds, we terminated our Delphi study in line with accepted termination conditions (Paré et al. 2013). Further, the panelists’ satisfaction continuously increased during the study and reached the highest value in this round. Complementing these quantitative assessments, the panelists’ qualitative feedback corroborated that the results had converged. In sum, the results reflect stable rating distributions per reason and process type. These results answer our second research question.

After the sixth round, we checked whether the differences and commonalities regarding the reasons’ importance for routine and nonroutine processes were statistically significant. To do so, we used the G test, as it is geared to small sample sizes (Holmes et al. 2011). The G test checks the rating distributions of routine and nonroutine processes for homogeneity. Significance means that both distributions most likely stem from heterogeneous populations. Although the size of our panel fitted the exploratory nature of Delphi studies, it is rather small for statistical purposes. Thus, as mentioned above, the differences in importance for routine and nonroutine processes should be seen as trend statements whose underlying causality needs to be substantiated in future research. We get back to this limitation in the conclusion. This in mind, we investigated commonalities and differences regarding the reasons’ importance for causing deviance in routine and nonroutine processes, because digitalization leads to a shift between both process types and because related insights indicate how process deviance needs to be managed. In our opinion, these insights complement the rating distributions and stimulate future research.

Finally, we checked for selection bias to ensure that satisfaction values had not risen because dissatisfied experts dropped out and only satisfied experts remained, but rather because the remaining panelists had become more satisfied during the study. This was important as the satisfaction values continuously increased during the study (Table 1). To that end, we looked at the final satisfaction values for all panelists who dropped out. An overall satisfaction of 5.56 and a coding satisfaction of 5.44 before dropout suggest that panelists did not drop out due to dissatisfaction and that our results do not suffer from selection bias.

Please refer to Appendix A-2 for detailed information about the medium list of reasons including the fraction of votes, votes from academics and practitioners, and ranking differences. The final rating distributions for the shortlisted reasons obtained in the rating phase are included in Appendix A-3. More information about the panelists’ satisfaction values can be found in Appendix A-4.

Despite the growing interest in academia and industry, process deviance is low on theoretical insights. Against this backdrop, our study makes two contributions to the descriptive knowledge on BPM in general and process deviance in particular: first, a compilation of 33 reasons for process deviance structured along nine categories, which are grounded on the input of academic as well as industry experts and have been approved by these experts; second, preliminary insights into the reasons’ importance for causing deviance in routine and nonroutine processes together with commonalities and differences.

Our findings extend current work on process deviance, most of which focuses on the detection (ex post view) and runtime prediction (ex nunc view) of process deviance based on process models and logs for business processes executed in automated workflow environments (Swinnen et al. 2012; Maggi et al. 2014; Nguyen et al. 2014; Delias 2017). While the value of these approaches is unquestioned, they focus on automated decision support and do not explain why processes show different behavior than intended. Existing approaches also neglect many processes from reality, namely those that are not or cannot be executed in automated workflow environments. Our findings strengthen the ex ante view on process deviance, and extend those works that already analyzed reasons for process deviance (Mertens et al. 2016b; Delias 2017). The reasons identified in our study apply to business processes in general, and only require the experience of process managers to be assessed. Finally, our findings extend research on success factors related to BPM and process reengineering projects. While Trkman (2010) proposed BPM success factors that address the fit between business processes with their business environment and IT, Al-Mashari and Zairi (1999) focused on success factors for process reengineering projects including effective communication, strong leadership, appropriate job descriptions, adequate resources, or effective IT. Bandara et al. (2005) investigated success factors specifically for process modeling initiatives. Although some success factors bear similarities with reasons and/or categories identified in our study, they are at best indirectly linked with process deviance. Nor do they focus on individual processes as unit of analysis or are operational enough to be assessed by process managers.

Our work entails the following further implications. As we already realized during the preparation of our Delphi study, process deviance must be conceptualized differently for routine and nonroutine processes to account for the properties of both process types. In general, process deviance can be specified operationally for routine processes, i.e., as non-compliance with process models. The reason is that, due to their predictable and transactional nature, most routine processes have been specified in terms of imperative process models prior to execution. Regarding nonroutine processes, the conceptualization of process deviance must account for their unpredictable and problem-solving nature. Despite recent advances in declarative process modeling, nonroutine processes remain hard to capture at design time. In case declarative models are used, the deviance of nonroutine processes can be assessed operationally, e.g., in terms of rule or constraint violations. To the best of our knowledge, however, declarative process models are hardly used in industry so far. Most importantly, the finding that process deviance needs to be conceptualized differently for routine and nonroutine processes is corroborated by our insights into the reasons’ importance for causing process deviance. Accordingly, routine and nonroutine processes call for distinct management practices related to process deviance, a requirement in line with research on context-aware BPM (vom Brocke et al. 2016).

Further, the identified reasons and categories suggest that process deviance is a complex multi-causal concept. Reasons for process deviance are not only closely associated with the process itself (e.g., included tasks, control flow, and documentation), but also with their context in terms of involved process participants, IT systems, and resources (i.e., material, equipment, employees). Further, reasons for process deviance are also associated with less operational and sometimes hardly tangible topics, known to be relevant for BPM as a corporate capability, such as governance and strategic alignment or process participants’ knowledge and skills as well as their attitude and behavior (Rosemann and vom Brocke 2015). This finding pinpoints the limitations of data-driven approaches when it comes to the analysis and management of process deviance. We also found that the relationship of most reasons with process deviance has a positive polarity, whereas some reasons appeared to have an inverse U-shaped relationship. A positive polarity means: the more one agrees with a reason, the stronger the process in focus tends toward deviance. An example is the reason that process participants do not have sufficient knowledge or skills. In contrast, an inverse U-shaped means that there can be a too much and a too little. Examples are the reasons that processes deal inappropriately with different contexts or that processes include an inappropriate control flow. Finally, our results inspire research on process deviance proneness, i.e., the tendency of business processes toward deviant behavior, and context-aware BPM, i.e., the design of appropriate management practices for routine and nonroutine processes.

Our study also offers practical implications for process managers. Analogous to the theoretical insights, our results sensitize process managers to process deviance being a complex multi-causal construct that needs to be assessed holistically. Moreover, using the identified reasons as a checklist, process managers can easily assess to which extent the processes in their area of responsibility tend toward process deviance. Despite the explorative nature of Delphi studies, it is justified to use the reasons as a checklist because the panelists agreed on them throughout multiple rounds. Using the identified reasons as a checklist offers a low-threshold assessment tool as process managers only require their own knowledge. Moreover, process managers can apply the reasons in any application domain. When assessing to which extent business processes tend toward deviance, process managers can derive custom weights for each reason based on the reasons’ importance for routine and nonroutine processes. Process managers can leverage such assessment results for multiple purposes: From a stand-alone perspective, process managers can use the results for redesigning existing or creating novel business processes. From a process portfolio perspective, they can leverage the assessment results to prioritize processes for improvement purposes if they contrast the processes’ tendency toward deviance against the expected impact of process deviance. Finally, our reasons make the case for not only checking for process deviance during or after process execution. Rather, process deviance should also be considered at design time and when redesigning existing business processes before it occurs.

To account for the ubiquity of and call for research on process deviance, this study identified, structured, and rated reasons for process deviance that can be assessed from a process manager’s perspective without requiring process models and logs as input. To do so, we conducted a rating-type Delphi study with more than 30 experts from academia and industry. This study resulted in 33 reasons for process deviance, structured along nine categories. Our study also resulted in preliminary insights into the reasons’ importance for causing deviance in routine and nonroutine processes.

Our study is beset with limitations. As with any Delphi study, our results are based on the input of a limited number of panelists. While we chose the panelists for their experience in BPM and their ability to take a managerial perspective on process deviance, we can make no formal claim about the representativeness of our study. Although we did not control for distinct industries or an experience regarding different process types, we are convinced that our Delphi panel is sufficiently diverse, because the panelists stem from multiple countries, have different academic backgrounds, and hold different positions in organizations. Although the composition and size of our Delphi panel fitted the exploratory nature of our research, the panel is comparatively small for statistical purposes. Therefore, the insights into the reasons’ importance for causing deviance in routine and nonroutine processes as well as related differences and commonalities should be treated as trend statements. During the study, a few panelists shared their opinion that our panel may have included too many highly qualified panelists with a strong managerial perspective. They suggested that the study be replicated only with process participants to cover an operational perspective on process deviance. With our study aiming to identify reasons for process deviance from a process manager’s perspective, we do not see an overly strong bias toward management experience within the panel. Nevertheless, we appreciate the idea of covering an operational perspective or further perspectives on process deviance in follow-up studies to complement our findings. As we strived for reasons that apply to business processes in general and can be assessed by process managers, some reasons tend to be abstract. However, this is in line with the setup of our study. As just mentioned, our reasons should be complemented by more operational reasons identified by taking other perspectives on process deviance and by domain-specific reasons. Finally, we treated the identified reasons as independent. In reality, process deviance is likely to depend on interactions among reasons, e.g., if some reasons occur simultaneously.

Both the limitations of our Delphi study and our results stimulate further research. Below, we overview the most prominent streams for future research, starting with the mitigation of limitations followed by ideas that take our results one step further.