Synthesis: The Prescriptive Process Monitoring Lineage (2014–2026)

A consolidated timeline of how PrPM crystallised as a distinct BPM sub-discipline, organised around four pivot points and three architectural traditions.

The four pivots

Year	Source	Pivot
2014	Gröger, Schwarz & Mitschang — Recommendation-based BPO	First explicit recommendation-from-process-data framing. Process warehouse + association-rule mining → operator recommendations. The "descriptive → predictive → prescriptive" pyramid is named.
2015	Krumeich, Werth & Loos — Prescriptive Control of Business Processes	First enterprise architecture for PrPM. Coins ED-BPM + EDPA. 4-component concept + 5-layer system architecture + 7 requirements. German steel-industry case study. Built on Complex Event Processing rather than the data-warehouse tradition.
2021	Dumas — Process Mining 2.0: From Insights to Action	Keynote that consolidates the vision: PM 2.0 = action-oriented PM, of which PrPM is the core technical instrument. Sets the stage for the SLR consolidation phase.
2022	Kubrak, Milani, Nolte & Dumas — PrPM SLR	The field's canonical SLR. 36 included papers. 6-dimensional characterisation framework (performance objective × metric × intervention type × modeling technique × data inputs × intervention policy). Identifies 6 research gaps.

Three architectural traditions

The 2014 and 2015 papers establish two distinct PrPM traditions that subsequent literature elaborates rather than reconciles. The 2026 LLM-PPM line is opening a third.

1. Data-warehouse / classifier tradition

Originates with Gröger et al. 2014. Architectural assumptions:

Cleaned, aggregated process data in a warehouse.
Periodic (batch-window) analytics rather than continuous streams.
Association-rule mining or supervised classifiers driving recommendations.
Operator-in-the-loop intervention.

Most of the literature in Kubrak's SLR inherits this tradition — even when the modeling technique changes to LSTM, RL, or causal trees.

2. CEP / event-driven tradition

Originates with Krumeich, Werth & Loos 2015. Architectural assumptions:

Real-time event streams (sensor data, transactional events, sub-second latency).
Complex Event Processing engine as substrate.
Bayesian networks / rule induction for probabilistic prediction.
Control feedback loop into process engine — closer to industrial automation than to managerial recommendation.

This tradition is conceptually closer to process control in chemical engineering and to manufacturing-MES integration than to traditional BPM tooling. Strongly motivated by process manufacturing (steel, chemicals) — less obviously useful for transactional white-collar processes.

3. LLM-based tradition (emerging, 2026)

Padella, de Leoni & Dumas 2026 flag prescriptive extension of LLM-PPM as future work. Distinct architectural assumptions:

Pre-trained foundation model (Gemini / Claude / GPT class).
In-context learning rather than supervised retraining.
Recommendations as natural-language reasoning + structured action proposal.
Particularly suited to data-scarce settings where the warehouse tradition requires data the organisation lacks.

No concrete LLM-PrPM paper exists yet. Open empirical question: can an LLM combine the warehouse tradition's interpretive flexibility with the CEP tradition's latency budget?

Kubrak's 6-dimensional framework as a unifying vocabulary

Where the three traditions disagree on infrastructure, they overlap on the decision content. Kubrak et al. 2022's six dimensions provide a comparison vocabulary regardless of architectural tradition:

Dimension	Krumeich 2015 (CEP)	Gröger 2014 (warehouse)	Padella 2026 future LLM-PrPM
Performance objective	KPI optimisation (time / quality / resources)	Process-instance outcome improvement	Both — context-dependent
Performance metric	Cycle time, quality units (QU), throughput	Various — case-level	Both KPIs in Padella et al. (Total Time + Activity Occurrence)
Intervention type	Control-flow (production-step choice)	Resource + process-flow recommendations	Open (paper does not specify)
Modeling technique	Bayesian network + rule induction over CEP	Association-rule mining over warehouse	LLM in-context learning
Data inputs	Sensor + transactional event streams	Process warehouse aggregates	Trace + global attributes + reasoning examples
Intervention policy	Continuous (every event)	Periodic (warehouse update cadence)	Per-query (LLM call)

Research gaps from Kubrak's SLR, still open

In-vivo validation — most evaluations are historical back-tests; field deployments remain rare.
Intervention discovery — no principled method to mine candidate interventions from logs; they are hand-specified.
Causal policies — accounting for cause-effect and second-order effects is the emerging frontier (concepts/causal-process-discovery).
Explainability and feedback loops between PrPM systems and end-users.
Dimension coverage — temporal metrics dominate; quality, compliance, revenue under-served.
Terminology — proactive adaptation, next-step recommendation, next-best action, on-the-fly resource allocation coexist; a common vocabulary is needed.

Practical entry points

For a manager wanting recommendation-style PrPM today: read Gröger 2014 for vision, Kubrak 2022 for tool selection.
For a manufacturing-operations engineer: read Krumeich 2015 for architecture, then go to the underlying CEP literature (Etzion & Niblett 2011 not yet in raw/).
For an APM-adjacent researcher: read Dumas 2021 for the vision, then the APM Manifesto for how PrPM fits into the agent-centric paradigm.
For someone building LLM-PrPM: there is no ingested precedent. Closest is Padella 2026 for the predictive half + the open prescriptive-extension flag.

Concept hub: concepts/prescriptive-process-monitoring.
Adjacent concept: concepts/event-driven-bpm.
LLM-PPM hub: concepts/llm-based-ppm.
Sibling synthesis: syntheses/ppm-landscape for the predictive layer beneath PrPM.
Reading list: syntheses/llm-bpm-reading-list.