From Process Mining to Augmented Process Execution — Chapela-Campa & Dumas 2023

Short Expert Voice article (10 pages) in Software and Systems Modeling that positions the trajectory from process mining to AI-augmented execution as a four-level analytics pyramid. Co-authored with entities/marlon-dumas by his postdoc David Chapela-Campa at the University of Tartu Information Systems group.

Summary

Business process management has traditionally relied on manual data collection and analysis, leading to slow and narrow decisions. Over the past two decades, data-driven techniques have emerged to automate activities across the BPM lifecycle. The authors synthesise this progression as a four-layer pyramid:

           [ Augmented Process Execution ]       (operational)
         [ Prescriptive Process Optimization ]
       [    Predictive Process Analytics     ]
     [    Descriptive Process Analytics        ] (tactical)
           [ Augmented Process Execution ]       (operational)
         [ Prescriptive Process Optimization ]
       [    Predictive Process Analytics     ]
     [    Descriptive Process Analytics        ] (tactical)

Each layer builds on the ones below. The left–right axis of the pyramid splits each layer into tactical use-cases (weeks-to-months horizon, aimed at managers) and operational use-cases (day-to-day, per-running-case). The chronology of research in data-driven BPM follows this pyramid from bottom to top.

Descriptive Process Analytics — the mature base. Automated process discovery, conformance checking, performance mining, and variant analysis, typically applied to event logs extracted from ERP/CRM/IoT systems. Consolidated in the 2000s and early 2010s.

Predictive Process Analytics — two sub-families. What-if Digital Process Twins (DPTs) build simulation models from event logs (a more honest term is digital process shadow since the model does not affect the real process) to answer tactical what-if questions. Predictive Process Monitoring (PPM) operates at the case level: predict outcome, remaining time, or next activity for a running case. A footnote observes that maintaining model performance post-deployment under concept drift is a persistent unsolved challenge.

Prescriptive Process Optimization — predictions become recommendations. Automated process optimization produces design-time redesign suggestions (e.g., resource re-allocation to relieve a bottleneck) using search-based algorithms + DPTs. Prescriptive Process Monitoring (PrPM) recommends runtime actions per case, e.g., sending a second loan offer to prevent rejection. Case-level techniques dominate; macro-level PrPM (recommendations across many cases) is identified as a gap.

Augmented Process Execution — the nascent top. Where prescriptive systems recommend to humans who execute, augmented systems execute the process themselves while humans assist/supervise. Roles invert: the system drives, humans intervene at the frame boundary. Augmented systems are of two types:

Autonomic — the system acts inside the frame; when a situation out-of-frame or high-uncertainty is detected, it escalates to a human and records the human's decision for future autonomous handling.
Fully autonomous — the system has complete control and may even change the frame, provided the change advances the business goals. Humans are supervisors.

The authors explicitly align these with levels 3–5 of van der Aalst's six levels of autonomous process execution management (APEM, arXiv:2204.11328, 2022) — level 0 maps to predictive analytics, levels 1–2 to prescriptive, levels 3–5 to augmented.

Conversational Process Management is introduced in a separate final section as an orthogonal dimension: LLM-based conversational systems translate natural-language user questions into executable analytical, predictive, prescriptive, or frame-specification actions at any level of the pyramid. This is the first explicit articulation in this wiki's corpus of conversational actionability at BPM-analytics granularity — a precursor of what the APM Manifesto later elevates to a first-class APM agent capability.

Key claims

The four-layer pyramid — descriptive → predictive → prescriptive → augmented — is the organising schema for data-driven BPM research.
Autonomy grows monotonically up the pyramid. Predictive = level 0 (assist only). Prescriptive = levels 1–2 (actively assist). Augmented = levels 3–5 (system drives; human intervenes at the frame).
Two sub-families per layer, split by scope (tactical / macro vs operational / per-case). Researchers tend to work in one quadrant at a time.
Augmented execution is nascent. Few deployed systems; challenge of reliable short-term impact evaluation across many cases.
Macro-level PrPM is a gap. Existing techniques recommend for a single case; cross-case recommendations (e.g., "too many resources on task A, too few on task B, redeploy") are open.
Conversational systems are orthogonal to the pyramid, not a new layer. Their role is translation, not new analytical capability.
Risk: frame-compliant, goal-divergent action. A key open challenge is ensuring augmented systems do not technically obey the frame while producing undesirable business outcomes — essentially a specification-gaming problem for BPM frames.
Composability under frequent change is identified as a safe-transition requirement for hyper-automated BPM.

Framing distinctions introduced

Digital Process Twin vs Digital Process Shadow. Since changes to a simulation model do not affect the real process, shadow is the arguably more accurate term — an uncommon (and useful) qualification in a field that uses twin loosely.
Tactical vs operational use-cases inside each pyramid layer. Makes explicit that every layer contains both design-time and runtime instances.
Autonomic vs fully autonomous augmented systems. Autonomic systems escalate; fully autonomous systems may modify the frame. This maps onto autonomy levels and anticipates the meta-framing distinction from ABPMS.
Augmented execution ≠ prescriptive optimization. The role inversion — human assists the system rather than the reverse — is definitional, not a matter of degree.

Complement to AI-Augmented BPMS (Dumas et al. 2023). ABPMS is the system/architecture-side manifesto; this paper is the analytics/capability-side synthesis published the same year by one of the ABPMS co-authors. The two are explicitly pointed to by the APM Manifesto as complementary lenses on the same transition.
Complement to APM Manifesto (2026). APM takes the agent-centric, management-oriented lens; Chapela-Campa & Dumas take the intelligence/analytics lens. The four-layer pyramid is the data-driven backdrop against which the APM capability stack (framed autonomy, explainability, conversational actionability, self-modification) is enacted.
Extension of Dumas 2021 — Process Mining 2.0. The "insights → action" thesis is precisely the descriptive → predictive → prescriptive progression, extended here by one more layer (augmented) and orthogonally by conversational interfaces.
Extension of Van der Aalst Process Mining (2011/2016). The book's operational support (detect / predict / recommend) seeds layers 1–3 of this pyramid; this paper extends the ladder with a fourth rung (augmented) above recommend.

Connections

Concepts: concepts/process-mining-spectrum · concepts/operational-support · concepts/predictive-process-monitoring · concepts/prescriptive-process-monitoring · concepts/conformance-checking · concepts/process-discovery · concepts/business-process-simulation · concepts/conversational-actionability · concepts/framed-autonomy · concepts/abps-autonomy-levels · concepts/concept-drift · concepts/agentic-bpm

Methods: methods/process-mining-basics

Authors (entities): entities/marlon-dumas · entities/david-chapela-campa · entities/university-of-tartu

Related sources: sources/2023-dumas-ai-augmented-bpms · sources/2026-calvanese-agentic-bpm-manifesto · sources/2021-dumas-process-mining-2-from-insights-to-action · sources/2024-kampik-large-process-models · sources/2022-kubrak-prescriptive-ppm-slr · sources/2011-vanderaalst-process-mining-book · sources/2012-vanderaalst-process-mining-manifesto · sources/2026-dumas-agentic-bpms-pyramid

Cited by

sources/2026-calvanese-agentic-bpm-manifesto (ref [80]) — complementary analytics-side viewpoint.
sources/2026-dumas-agentic-bpms-pyramid (ref [8]) — the 2026 keynote explicitly adapts this paper's four-level pyramid into the concepts/agentic-bpm-pyramid, replacing the apex Augmented Process Execution layer with Agentic BPM.