Key takeaway: Optimizing BPMN processes starts not with simplifying the diagram, but with interpreting execution data. Measure lead time, waiting time and active work separately; identify the cause of congestion; then intervene through a versioned, controlled change. Improvement is demonstrated when results improve for comparable cases without compromising quality or compliance.
Within the topic of “BPMN 2.0 in Practice,” this is where the model meets operational reality. An approval process may be logically flawless yet regularly run late. Process improvement based on the xFLOWer approach therefore examines not only how a case progresses, but also where, why and for how long it stalls.
Practical focus: This series is not a complete BPMN 2.0 reference manual—the official notation set is far broader, and here we focus on the most essential elements and practical workflows used in everyday operations.
1. First, clarify what your measurements mean
A process instance is the execution of a specific case, such as a purchase requisition. Analysis is based on its event log, which identifies the process version, step, status change and timestamp. For user tasks, assignment, acceptance, the actual start of work and completion may be separate events.
The time between task creation and completion is not automatically active working time. If you store only these two timestamps, you cannot reliably separate waiting time from work. Measuring active effort requires separate time tracking or appropriate status management.
Also specify whether you calculate time in calendar or business hours, how you handle waiting for the customer, and which event constitutes business completion. Technical closure, for example, is not necessarily the same as sending out a purchase order. Without defined start and end points, two reports can produce different results for the same process.
Assign an owner to the measurement agreement: the business process owner should approve it, while operations should verify data completeness and timestamp consistency.
2. KPIs: measure speed alongside quality
A single average lead time is not enough to support decisions. Alongside the average, the median shows the typical case, while an upper percentile highlights slower cases. With few observations, however, percentiles can fluctuate, so always include the sample size and reporting period.
| KPI | What does it show? | Interpretation guidance |
|---|---|---|
| Total lead time | The time elapsed from business initiation to completion. | Compare using identical start and end points and the same time calculation method. |
| Step-level waiting time | Idle time before work starts or between two activities. | Separate internal queueing from waiting for an external response. |
| Active working time | The actual processing effort. | Calculate it only from properly recorded work periods. |
| On-time completion rate | The proportion of cases completed within the committed timeframe. | Specify the denominator, pauses and exceptions. |
| Return-for-correction rate | The frequency of cases returned for correction. | Track the number of affected cases and repeat returns separately. |
| Throughput | The number of cases closed in a given period, for example 80 cases/week. | Examine it alongside incoming volume: even with 80 closures per week, the queue can grow if 100 new cases arrive. |
| Open caseload | The number of open cases at a given point in time, for example 35 cases at close of business on Friday. | Always measure at the same time, and break the figures down by case type. |
The age of open cases is particularly important: a report that examines only closed cases can hide work that has been stuck for a long time. Break down the data by case type, priority, organizational unit and process version, but avoid unjustified rankings of individuals. The goal is to uncover operational causes, not assign blame.
3. How to distinguish a symptom from a bottleneck
The longest step is not necessarily the bottleneck. A lengthy but rarely used expert review may constrain the overall process less than a short approval that routes every case to a single role. A persistently growing queue is a stronger warning sign.
- Locate the congestion. Where are the number and age of open tasks increasing? Also examine how arrivals are distributed over time.
- Break down the time spent on each step. Is the delay caused by actual processing, waiting for assignment, requests for missing information, or a system error?
- Examine the paths. Which gateway branches, exceptions, and rework loops do slow cases follow?
- Check capacity and rules. Is progress blocked by a lack of backup coverage, permission restrictions, or overly broad approval requirements?
- Validate the findings with the people performing the work. The log can show where work stalls, but understanding the cause often requires employees’ practical experience.
In an illustrative procurement example, the financial check may appear slow. However, if cases are regularly returned because a cost center is missing, adding another reviewer is not necessarily the solution. Validating data at entry could eliminate repeated processing. Treat this initially as a testable hypothesis, not an established fact. If entering a cost center becomes a prerequisite for submission, an invalid request may never create a process instance. The log may therefore show fewer requests for missing information even if more users get stuck on the form. Track started requests and successfully submitted requests separately on the form as well, so you can see how many users reach submission after the new validation is introduced.
4. Intervene at the point in the BPMN model that addresses the cause
Translate the optimization proposal into a specific model element. An event can indicate an arrival or a deadline; an activity can validate data; a gateway controls branching and merging; a lane makes the responsible role visible. A lane alone does not provide backup coverage or task assignment logic.
Use an exclusive gateway when conditions allow only one path to continue. A parallel gateway can launch independent checks. A joining parallel gateway waits for an arrival from every incoming branch. For example, if the financial check starts only above a value threshold, a shared parallel join could block a case below that threshold; this requires conditional branching and an appropriate merge. An event-based gateway is appropriate when, for example, whichever occurs first— a customer response or a timer event—determines how the process continues.
Parallelization does not automatically eliminate capacity shortages. If the same overloaded employee performs two checks, the model may look faster while the overall lead time remains unchanged. Nor should the times for parallel branches simply be added together: the actual path and synchronization determine the total time for the case.
When introducing a timer, clarify whether you want a reminder, an escalation, or an interruption. These have different business consequences; the wrong choice can cause duplicated or interrupted work.
5. Treat the change as a versioned release
Optimizing BPMN processes in a live environment is also a change management task. Do not overwrite the model without a trace: every release should have a version identifier, a reason for the change, an owner, test results, and a deployment date. The event log should continue to show which version each case ran on.
What should happen to cases that have already started?
A safe starting point is often to let running instances finish on the old version while new instances start on the new one. However, this is not sufficient for every change. A mandatory rule change may require migration or a controlled manual transition.
Before migration, check the mapping of active steps, the data structure, permissions, pending events, and the state of parallel branches. Examine the capabilities supported by the process engine separately; BPMN 2.0 notation alone does not guarantee a safe version transition.
Test the normal path, returns for rework, timeouts, integration errors, and repeated messages. Have a rollback plan, but remember that rolling back the model does not reverse an order already sent or a payment already executed.
6. Follow-up measurement: did the intervention actually help?
Define the expected result before deployment: which KPI should improve, for which category of cases, and subject to what quality requirement? For example, you may expect mandatory data validation to reduce requests for missing information, while monitoring whether it causes an unjustified increase in request entry time or the number of abandoned cases.
Choose groups of cases with a similar composition for the before-and-after comparison. Account for seasonality, staffing changes, incoming workload, and priorities. Do not uncritically compare old cases that are still open with cases that close quickly under the new version.
Start with a limited pilot if possible. During evaluation, examine the distribution, exceptions, and open caseload alongside the average. If congestion has merely moved to a later step, a local improvement does not yet represent an improvement in the overall process. Measurement should therefore lead to another documented decision.
7. Summary and related practical articles
An effective improvement cycle: accurate measurement, root cause analysis, targeted model changes, controlled release, and follow-up measurement. Further articles in this series explore the foundations needed for this approach:
- BPMN 2.0 elements explained with examples: event, activity, gateway, and lane.
- Exclusive, parallel, and event-based gateways: which one should you use, and when?
- The 10 most common BPMN modeling mistakes: preventing stalled processes and ambiguous models.
- BPMN vs. flowcharts vs. UML activity diagrams: making an informed choice of notation.
- How does a BPMN diagram become an executable process? Data, rules, roles, and integrations.
Want to identify where your processes get stuck? Consult xFLOWer experts about a measurement and optimization plan for a specific process. As a starting point, bring your current model, the available execution data, and the business problem you want to address.
This article is part of our comprehensive guide: BPMN 2.0 in Practice: A Complete Guide from Modeling to a Working Process
xFLOWer.ai
Design the process behind this article – for free!
AI reads the processes hidden in this article; open any of them in xFLOWer.ai with one click.
