Purchase-to-Pay Conformance and Data Quality
A procurement group extracted 1.2 million events from its ERP, supplier portal, and e-invoicing system covering 18 months of purchase-to-pay. The reference model has 47 activities. Conformance checking shows 31% of cases follow the model, with 412 variants; the most frequent variant covers 18%. In 9% of cases, 'Change payment terms' is repeated three or more times. In 4% of cases, 'Three-way match' occurs before 'Goods receipt'. The average lead time is 19 days, but the 90th percentile is 71 days. The process owner concludes the process is 'out of control' and wants to enforce the reference model with workflow automation. The process mining team cautions that the log is incomplete for two legacy plants and that timestamps for goods receipt are entered manually at end of shift. The CFO wants a defensible interpretation before approving a redesign budget.
1. Interpret the 31% conformance figure. What can and cannot be concluded, and what data-quality issues must be checked first?
Read the worked answer
31% conformance means that 31% of cases followed every step and transition in the reference model as recorded in the log. It does not mean 69% of cases are wrong or uncontrolled. A reference model is a simplification; variants can be legitimate responses to customer, supplier, or plant differences.
Before drawing conclusions, check log completeness and timestamp quality. Two legacy plants are missing, so the log is not a representative sample of the whole process; variant and performance figures are biased. Goods receipt timestamps are entered manually at end of shift, which distorts waiting times and can create false 'three-way match before goods receipt' sequences. Check event deduplication, case ID mapping across the three systems, and whether activities are recorded at start and complete. Also distinguish conformance by variant: the top variant covers 18%, so the remaining 13% conforming cases are fragmented. A low conformance rate with a long tail of rare variants may indicate genuine process diversity rather than misconduct. Do not enforce the reference model with workflow automation until the log has been validated and the variants have been explained with process owners.
2. Propose a prioritized analysis plan to distinguish genuine process problems from model or data artifacts, and advise on the process owner's automation proposal.
Read the worked answer
First, validate the log: reconcile case counts against ERP document numbers, confirm the two missing plants, and correct or exclude manually entered goods-receipt timestamps. Re-run conformance and compare results. Second, analyze variants by plant, supplier category, and material type. If variants cluster by organizational unit, they may be legitimate local procedures; if they cluster by value or buyer, they may indicate control failures.
Third, investigate the specific anomalies. Repeated 'Change payment terms' in 9% of cases may be a legitimate renegotiation pattern or a workaround; check who changes terms, when, and whether approvals exist. 'Three-way match before goods receipt' may be a timestamp artifact; verify with invoice and receipt documents. Fourth, analyze performance by variant: compare lead time and the 90th percentile for the top variants to find redesign candidates.
The automation proposal is premature. Enforcing the current model would hard-code a model that may not reflect reality and could block legitimate variants. Instead, use the validated analysis to redesign the target model, then automate selectively. Set up continuous monitoring with conformance and cycle-time alerts after the redesign.