From process evidence
to enterprise value.

Three fields are mandatory: a case identifier, an activity name and a timestamp. Everything else — resource, org unit, document type, value, cost centre — is enrichment, and it is the enrichment that makes an analysis actionable rather than merely correct.

The interchange standard is IEEE 1849-2023 for XES, which superseded IEEE 1849-2016 in September 2023 and is active through 2033. For object-centric event data the format is OCEL 2.0, released October 2023, with SQLite, XML and JSON exchange formats. Note that OCEL is the exchange format for object-centric event data — OCED itself remains an in-progress IEEE Task Force standardisation effort, not a ratified standard, and anyone describing it as one is ahead of the facts.

The practical constraints are timestamp granularity (same-second events destroy ordering), timezone consistency across systems, and case-ID stability where a document is reissued or a case is split.

The academic literature is unusually blunt about this. Berti, Park, Rafiei and van der Aalst put it plainly: "The extraction, transformation, and loading of event logs from information systems is the first and the most expensive step in process mining." SAP specifically "poses major challenges, given the size and the structure of the data," with scarce open-source ETL support.

The tables that matter are well known: EKKO / EKPO and EBAN, EKET, EKBE, EKPA, RBKP / RSEG for procure-to-pay; VBAK / VBAP and VBFA for order-to-cash; BKPF / BSEG for finance and payment; and CDHDR / CDPOS change documents across everything.

Change documents are where the real work is. There are three ways to derive activities from them — map transaction codes to activities, turn changed field names into activity names, or compare old and new field values to derive a semantic activity such as "Postpone Delivery" or "Reduce Price". The third produces genuinely useful activities and consumes most of the transformation effort. Budget for it explicitly.

On replication: Celonis uses an installed extraction client with RFC-based extraction and a real-time extension; SAP Signavio Process Intelligence documents SAP Datasphere replication flows as a route. SLT, ODP, CDS views and OData are the standard SAP-side delta mechanisms. We will tell you what your licence and landscape actually support rather than what a datasheet implies.

Conventional process mining assumes each event refers to exactly one object, usually called the case. Real operations do not behave that way: one sales order becomes three deliveries and two invoices, against one customer and several materials.

Forcing that into a single case notion produces two documented distortions, named by van der Aalst. Convergence — events referring to multiple objects of the selected type are replicated, causing unintentional duplication and misleading diagnostics. Divergence — where different non-selected object types participate in the same activity they become indistinguishable, and the causal relationships between objects are lost.

In practice: an order-to-cash analysis flattened to a sales-order case ID will double-count invoice events across order lines and lose the delivery-invoice interleaving entirely. Object-centric process mining models events against multiple related object types so the process you analyse is the process that happened. Celonis implements this through Objects, Events and Perspectives, on licence tiers that include it. It is more work to model and it is usually the right answer for anything genuinely end to end.

Process mining reads event logs from transactional systems; task mining records user interactions with the desktop into a UI log which is then preprocessed and analysed. Different data, different unit of analysis, different privacy posture — and they answer different questions.

Process mining tells you that an invoice sat for six days between receipt and approval. Task mining tells you that during those six days it was manually re-keyed into a spreadsheet, checked against a portal, and emailed twice. The automation business case usually lives in the second. ARIS names its module Robotic Process Discovery.

Task mining carries real workforce-surveillance and privacy considerations, particularly under Australian privacy obligations and enterprise agreements. We design consent, aggregation and anonymisation into the deployment rather than retrofitting them after the first complaint.

Token replay replays each trace against a Petri net and counts produced, consumed, missing and remaining tokens. It is fast and can mislead on unsound or highly concurrent models. Alignment-based conformance computes an optimal alignment between the trace and the closest model execution, giving a provably optimal explanation of the deviation at materially higher computational cost — the academic gold standard, with active research into making it tractable at enterprise volume.

The four quality dimensions — fitness, precision, generalisation and simplicity — were formally defined by Buijs, van Dongen and van der Aalst. Fitness and precision are computable against a log; generalisation and simplicity are the trade-off dimensions that stop a discovered model over-fitting into an unreadable spaghetti diagram that technically explains every case.

Commercial platforms generally do not publish which conformance technique they implement. We will tell you what the tool documents and what it does not, rather than filling the gap with a confident guess.

Event data is operational data with personal information in it — user IDs, approvers, sometimes customer detail. Role-level access, pseudonymisation of resource fields, retention limits and a documented lawful basis are design inputs, not afterthoughts, and they interact directly with the Australian Privacy Principles and with the automated-decision transparency obligation commencing 10 December 2026.

Architecturally, process intelligence sits beside rather than inside your data platform. It needs its own event model, but it should read from governed sources and publish its measures back into the same semantic layer your reporting uses — otherwise you have created a second version of the truth to argue with the first.

For SAP estates, the modelling repository is the durable asset: ARIS or SAP Signavio holding the designed process, integrated to SAP Cloud ALM, with mining supplying the observed reality against it. Solution Manager still appears in older integration material; treat Cloud ALM as the forward path.