A Declarative Process Mining Framework Evaluated on Real Sales Data

Customer Relationship Management (CRM) automation is usually built from procedural workflows that prescribe a fixed path, although real sales processes are highly variable. This study proposes and empirically evaluates a framework for declarative automation in CRM that mines constraint-based (Declare) rules from event data, refines them, validates them on held-out cases, and separates hard rules suitable for enforcement from soft rules suitable only for monitoring. The framework is evaluated on a real-life sales-funnel event log (BPI Challenge 2012) comprising 12,688 closed cases, 4,040 process variants and 68 employees, using a temporal train/test split. Discovery produced 1,310 constraints. Four logic-based refinement steps reduced them by 67.3% to 429 rules, of which 380 hard rules held in 100% of unseen test cases and map directly onto CRM flows and validation rules. In a discrimination test against Inductive Miner and Heuristics Miner models, the hard declarative rules achieved the highest balanced accuracy (0.856 versus 0.70–0.73), accepting all legitimate cases while rejecting 71% of invalid variants. Multi-perspective analysis showed that following up 1–7 days after a proposal was associated with substantially higher win rates (about 60% versus 39% after more than seven days; p < 0.001), whereas speed-to-lead was not associated with conversion. Cases that deviated from soft rules were won more often than compliant cases (29.4% versus 14.8%), which indicates that soft rules should guide rather than constrain salespeople. The study contributes a reproducible method and empirical evidence for combining enforced and monitored declarative rules in CRM automation. This is a preprint that has not been peer reviewed. It has been submitted to Process Science (Springer).