Process Mining as Evidence for AI Redesign

01 · The evidence boundaryThe graph shows recorded work, not the process

The useful unit is a defensible claim about a defined population, interval and perspective, made for a named redesign decision.

An event log is produced by operational work and by the systems chosen to record it. An event might mark when a person acted, when a field was saved, when a workflow was submitted or when an interface synchronised. Activity names reflect configuration and data engineering. Case identifiers determine which events appear to belong together. Filters decide whether open, rejected and failed cases remain visible. A discovered model is therefore conditional evidence, not a neutral digital twin.

This distinction does not diminish process mining. Workshops struggle to recall thousands of cases, quantify waiting distributions or expose rare sequences. Event data can reveal concurrency, repeated transitions, queues, seasonal shifts and interactions across objects that participants cannot aggregate unaided. The discipline begins when the team stops asking whether the graph is accurate in general and asks whether a particular trace claim is valid for a particular decision.

Figure 1. Conceptual trace. Blank space is part of the evidence statement, not an inconvenience to conceal.

Thought experiment: a perfect log of the wrong system

Imagine a log with exact timestamps, no duplicates and complete coverage of every action in the claims platform. It still excludes the broker portal, telephone negotiation and the repair team’s shared mailbox. Perfect technical quality cannot make the platform boundary equal the claims process. Now imagine a sparse log containing order creation, dispatch and confirmed delivery. It may still support a narrow claim about elapsed fulfilment time if exclusions are explicit. Evidence quality belongs to a question, not to a dataset in isolation.

The negative control

Before proposing AI for a mined bottleneck, run the same representative cases after changing no model capability. Clarify authority, repair a timestamp, remove a redundant approval or expose a missing record. If the target outcome improves, the log identified a process condition but not a need for intelligence. This negative control prevents the visible queue from becoming automatic proof that classification, generation or an agentic loop is required.

02 · Process Mining Evidence ArchitectureNine layers discipline the journey from trace to action

Analysis sits in the middle of the architecture. The redesign question, population, semantics, perspective and coverage come first; triangulation, inference and revalidation follow.

The first layer names the decision, alternatives, owner, consequence and evidence needed. The second defines eligible cases or objects, exclusions, interval, versions and affected populations. The third governs what each event and timestamp means. The fourth states how events are correlated into cases or object relationships. The fifth measures missing, duplicated, reordered, inferred or selectively recorded material. Only then does the sixth layer choose discovery, performance, conformance, organisational, predictive or object-centric analysis.

The last three layers keep the result honest. Triangulation compares the trace with procedures, outcomes, incidents, observation and affected-party accounts. Redesign inference separates what was observed from a causal hypothesis and proposed intervention. Action and revalidation give the claim an owner, experiment, protection, expiry and deletion decision.

Figure 2. This article’s synthesis. A technically advanced analysis cannot compensate for an undefined population or ambiguous event.

Layer	Governing question	Evidence	Characteristic failure	Owner	Decision effect
Redesign question	What decision will this evidence change?	Options, consequence, deadline	Interesting graph without action	Process owner	Sets permissible claim
Population	Which cases, objects and periods count?	Inclusions, exclusions, versions	Selection bias	Service owner	Defines denominator
Semantics	What business event does each record represent?	Creation logic, lifecycle, timestamp	Interface action mistaken for work	Domain and data owners	Supports or blocks timing claims
Perspective	How are events related?	Case keys, objects, qualifiers	False loops or broken traces	Process architect	Chooses defensible view
Coverage	What is missing or transformed?	Lineage, reconciliation, samples	Missing work treated as no work	Data steward	Limits scope or requires repair
Analysis	Which method answers the question?	Parameters and reproducible run	Complexity without necessity	Analysis lead	Selects lowest sufficient rung
Triangulation	What other evidence tests the pattern?	Observation, outcomes, incidents	Trace treated as causal proof	Redesign lead	Strengthens or weakens inference
Inference	Which cause and intervention remain plausible?	Rivals, uncertainty, controls	Queue equated with missing AI	Accountable sponsor	Authorises experiment only
Action	What changes, closes and expires?	Docket, monitoring, deletion	Permanent analysis without owner	Operational owner	Creates a reversible decision

03 · Event Log Evidence ContractA CSV is not an event log until meaning is governed

Write the contract before visual filters, mining algorithms and vendor defaults harden into hidden assumptions.

The contract names the redesign alternatives and non-action consequence, then defines the population, start, end, unresolved cases and observation interval. It inventories source systems, fields, interfaces, extraction time and transformations. Every event receives a business definition, lifecycle meaning, time basis and source owner. Every identifier states what it joins, splits or excludes. Missing, duplicated, aggregated, inferred and corrected records remain visible as evidence conditions.

Two owners must sign different claims. The technical extraction owner explains how rows were created, joined and changed. The business meaning owner confirms what those rows can say about work. Neither can substitute for the other. A technically perfect join can connect the wrong business episodes; a plausible domain story can ignore how a batch job actually writes time.

Figure 3. This article’s contract. Rows become evidence only after source behaviour is bound to business meaning and permitted use.

Field	Claimed meaning	Possible source behaviour	Defect	Repair	Invalid use until repaired
Activity	Work step completed	Screen opened or status saved	Interface action substituted for work	Inspect creation logic and sample cases	Counting completed steps
Timestamp	Time work occurred	Overnight synchronisation time	Queue and order distortion	Use event time or bounded interval	Worker delay attribution
Lifecycle	Start and completion	Only completion is written	Active time unavailable	Add start evidence or avoid duration claim	Utilisation inference
Case key	One service episode	New key per facility or channel	One episode split or many merged	Test alternatives and object relations	Variant frequency claim
Closure	Outcome delivered	Administrative state only	Customer outcome absent	Join downstream acceptance evidence	End-to-end cycle claim
Actor	Person performing work	Shared account or automated writer	Misattribution and privacy risk	Aggregate to necessary role or team	Individual ranking
Missing event	No activity occurred	Offline, vendor or failed write	Absence confused with zero	Coverage sample and explicit blind spot	Compliance conclusion
Corrected event	Observed fact	Imputed or reconstructed row	Inference presented as observation	Flag method and uncertainty	Causal or disciplinary use

The overnight approval

An invoice log appears to show that most approvals occur at 02:00. A performance view treats midnight queues as evidence of slow managers. The semantics gate finds that approval happens during the day, while the enterprise system receives a consolidated batch overnight. The correct timestamp supports a statement about synchronisation, not human waiting. Repairing the time field changes the queue distribution before any redesign option is considered.

04 · Case-Notion and Object-Perspective TestThe identifier redraws the process

The same events produce different paths when grouped around an application, customer, order, item, payment or shipment. None is automatically the master view.

List every plausible case notion and business object. For each, show included events, excluded work and the redesign claim it can support. Test convergence, where several cases share a payment, review or resource. Test divergence, where one case branches into items, fulfilments or outcomes. Flattening many-to-many relationships can create false loops, duplicate activity and impossible ordering.

Object-centric event logs offer a richer representation. OCEL 2.0 formalises events, objects, object types, changing attributes and qualified relationships between events and objects or between objects. That expressiveness reduces distortions caused by forcing everything into one case key. It does not remove judgement. The analyst still chooses object types, relationships, qualifiers, scope and a decision question.

Figure 4. Rotating the perspective can change sequence, variant count and apparent rework while the underlying events remain fixed.

Perspective	Included events	Excluded work	Likely distortion	Supported claim	Material rival
Loan application	Intake, checks, recommendation, application decision	Shared customer review and facility effects	Review duplicated across applications	Application administration time	Customer-level dependency
Customer	Interactions across applications and facilities	Item-specific fulfilment detail	Concurrent work appears sequential	Customer burden and repeated request	Legitimate separate products
Order	Creation, payment, allocation, dispatch	Item splits and partial delivery	False loop at repeated allocation	Order-level fulfilment	Item shortage drives path
Item	Pick, pack, substitution, shipment	Shared payment and customer contact	Common events copied per item	Item handling variation	Order-level consolidation
Payment	Authorisation, capture, refund	Delivery and service recovery	Commercial outcome truncated	Payment processing interval	Dispute originates elsewhere
Object-centric	Qualified relationships across selected objects	Anything outside modelled sources and types	Complexity mistaken for completeness	Cross-object interaction	Omitted offline or external object

The bank with several facilities

A bank groups events by loan application. One customer may have several facilities, and one financial review can affect several applications. The case-centric view duplicates review work and suggests repeated assessment. A customer perspective reveals shared evidence but obscures facility-specific approvals. An object-centric view relates customer, application, facility and review events. The redesign docket retains all three because the question changes: reduce customer evidence requests, shorten an application decision or govern shared reviews.

05 · Process Evidence LadderClimb only as far as the question requires

More sophisticated mining is not automatically stronger evidence. Each rung adds prerequisites, privacy exposure and possible misinterpretation.

Begin at rung zero with existing service, quality, safety, error, cost and affected-party outcomes. A descriptive event profile then establishes sources, population, completeness, frequencies and time coverage. A bounded path and performance view can quantify directly-follows relations, waits and rework for a declared population. Conformance compares recorded behaviour with a reference model. Object-centric analysis becomes useful when one case notion distorts many-to-many work. Protected organisational analysis examines roles or handoffs only when necessary. Predictive or prescriptive analysis comes last because leakage, intervention, uncertainty and action rights must be governed.

Figure 5. This article’s ladder. The apparent summit has the largest action and privacy burden, not automatic evidential superiority.

Rung	Prerequisite	Output	Privacy control	Valid inference	Invalid inference	Escalate when
0 · Outcomes	Comparable outcome definition	Distribution by case type	Minimum affected-party fields	Where service differs	Which step caused it	Path relation is needed
1 · Profile	Population and lineage	Coverage, frequency, time span	Remove unnecessary identity	What the log contains	What the process contains	Sequence matters
2 · Path	Reliable order and case view	Variants, waits, loops	Process-level aggregation	Recorded pattern prevalence	Cause or waste	Reference comparison matters
3 · Conformance	Legitimate reference model	Alignments and deviations	Small-cell suppression	Difference from model	Harm or misconduct	Single case view distorts
4 · Object-centric	Governed object relations	Cross-object interactions	Relationship minimisation	Many-to-many pattern	Complete process truth	Role interaction is essential
5 · Organisation	Necessary purpose and consultation	Role and handoff pattern	Protected setting and thresholds	System-level coordination issue	Worker intent or rank	Prospective action is justified
6 · Predictive	Valid target, no leakage, action right	Bounded prospective alert	Purpose-specific access and expiry	Estimated risk under conditions	Causal benefit	Trial can test intervention

The retrospective alert that knew the future

A predictive model flags service cases likely to breach a deadline with impressive retrospective performance. Inspection shows that one feature is a “specialist escalation” event created after the moment when the alert would need to fire. The model has learned a downstream consequence, not an actionable early signal. Leakage testing blocks deployment. A prospective design must freeze features at the intervention point, name an authorised action and evaluate whether that action improves outcomes.

06 · Variant and Exception TopologyFrequency is not meaning

Classify a path only after connecting its population, trigger, work, authority and outcome. Common does not mean desirable; rare does not mean disposable.

For each material path, record absolute volume and population share, case mix, consequence, predecessor, successor, elapsed time, active time, queue, rework, evidence, decision, authority, control, channel and downstream effect. Then test meaning: intended, permitted, adaptive, harmful, unresolved or a logging artefact. Keep a plausible rival beside the label.

Figure 6. This article’s topology. The same loop shape can represent correction, authorised reconsideration or duplicate logging.

The accessibility route mistaken for inefficiency

A service team finds a long variant used by a small population. It contains two extra contacts and an apparent return to verification. Observation shows an assisted-service route that obtains consent carefully and produces fewer repeat contacts for people using accessibility support. Removing the path would improve average cycle time while worsening the outcome for those cases. The topology records it as an intended adaptive route and asks whether tools can reduce customer effort without deleting professional judgement.

Conformance measures difference, not goodness

A reference model might represent policy, an expected technical sequence, a control obligation or merely historical practice. Before interpreting deviations, ask whether the model permits legitimate discretion and affected-party routes. An expert may depart from a simplified sequence to prevent harm. Strict conformance could suppress that judgement. Conversely, a frequent deviation can reveal a control failure. Domain adjudication and outcome evidence decide whether the model, logging or work should change.

07 · AI Redesign Inference ChainA pattern is the start of an explanation

Mark each link as observed, triangulated, inferred, tested or disproved. Authority should not travel farther than the evidence.

The chain runs from recorded pattern to semantic finding, triangulated condition, process-cause hypothesis, AI or non-AI option, bounded experiment, end-to-end outcome and owner decision. A long queue is observed only if timestamp semantics support waiting. That queue may be caused by capacity, missing evidence, policy, authority, integration or demand mix. Missing intelligence is one rival among several. The experiment should intervene on the suspected cause and measure the full outcome rather than reward a local speed improvement.

Figure 7. This article’s inference chain. Every jump has a rival and a route back when evidence fails.

The approval queue with no authority

A visible queue attracts a proposal to summarise documents for approvers. Case observation finds that documents are already short and complete. The wait begins because no role owns decisions above a recently changed threshold. Policy clarification and delegated authority dissolve the queue. The mining evidence correctly located recorded waiting, but the causal test selects process repair rather than AI.

A rival explanation must survive beside the favourite

For an apparent rework loop, plausible rivals include duplicate logging, legitimate reconsideration, an upstream data defect, novice learning, policy ambiguity and an interface that forces repeated saves. Record which evidence would distinguish them. This discipline is especially important when a process graph seems to confirm an executive’s prior belief. The redesign docket should preserve material rivals even after an experiment begins.

08 · Privacy and purposeMine the process without turning people into a ranking

Availability is not necessity. Worker identifiers, free text and fine-grained traces require a specific purpose, proportionality and a protected use boundary.

The privacy and purpose contract names the redesign question and lawful or organisational basis, then minimises events, attributes, granularity and interval. It defines whether analysis occurs at case, object, role, team or affected-party level. It sets aggregation thresholds, pseudonymisation, small-cell handling, access, retention and deletion. It prohibits silent reuse for individual ranking, discipline, sentiment or productivity inference. Worker notice, consultation, representation and challenge are operating controls, not communications added after extraction.

Process analysis can expose a person even when names are removed. A rare role, unusual schedule or small team may be identifiable from its pattern. Prompt, message and document content should not be collected by default merely because a platform can ingest it. Prefer process states and outcome evidence. Protected organisational analysis belongs on a higher evidence rung because it changes the consequence of a trace.

The call-centre league table

A call centre groups cases by agent and ranks path length. Longer paths are treated as low performance. The purpose contract rejects the analysis: path length reflects case mix, transfers, accessibility needs and system failure as well as agent action. The redesign question is re-framed around repeated customer contact and handoff friction. Necessary evidence is aggregated by process condition, rare cells are suppressed, and individual performance reuse is prohibited. This is consistent with current UK worker-monitoring guidance on lawfulness, fairness, transparency, necessity and proportionality, while its legal scope remains the United Kingdom.

The strongest rival: the trace can be faster than talk

Process mining can discover patterns at a scale and precision interviews cannot match. Demanding extensive qualitative work may delay obvious improvements, introduce anecdotal bias and let process owners explain away inconvenient evidence. In highly digitised operations, logs can have excellent coverage. Object-centric methods can reduce distortion, while automated conformance and predictive monitoring can create immediate value.

The answer is proportional evidence, not reflexive scepticism. A well-governed log can support a narrow descriptive claim quickly. Triangulation effort should rise with the inferential distance from the trace and the consequence of being wrong. Use representative observations to resolve material ambiguity, not to recreate the entire dataset by hand. Escalate methods when the question needs them and semantics, privacy and action rights are ready.

Explicit failure boundary. Stop when the redesign question has no accountable owner, the population cannot be defined, material events have unstable meaning, the chosen perspective changes the conclusion without resolution, missing work is correlated with the outcome, protected groups would be exposed, a reference model omits legitimate discretion, predictive features leak future information, or no ethical and operational experiment can distinguish the proposed cause from its rivals.

09 · Ten named casesEvidence can quantify, redirect, repair or stop

Each case begins with a redesign question and ends with an evidence decision, not an automatic technology recommendation.

1 · Connected bank facilities

Application IDs duplicate a shared customer review across several facilities.

Route: compare application, customer and object-centric views before making any redesign claim.

2 · Accessible service route

A longer path serves a small population needing assisted communication and produces fewer repeat contacts.

Route: add observation and affected-party outcomes before changing the path.

3 · Overnight invoice batch

Approval timestamps record synchronisation rather than the time work occurred.

Route: repair time semantics and lineage before measuring queues.

4 · Claims reconsideration loops

Repeated assessment mixes legitimate reconsideration, duplicate events and avoidable rework.

Route: run a bounded variant analysis with sampled case adjudication.

5 · Expert conformance deviation

A normative model marks authorised professional discretion as failure.

Route: repair the reference model before evaluating the work.

6 · Agent path-length ranking

Fine-grained traces are repurposed to rank staff despite confounded case mix.

Route: stop the use, minimise identity and redirect to process-level handoffs.

7 · Ownerless approval queue

Waiting is visible, but documents are complete and no role owns the changed threshold.

Route: repair authority through a non-AI intervention and remeasure outcomes.

8 · Rare fraud variants

Consequential cases are too small for an open dashboard and reveal specialist roles.

Route: use protected analysis and independent adjudication with strict access.

9 · Future-aware alert

A retrospective risk model uses an escalation event written after the proposed alert point.

Route: reject the model, remove leakage and require prospective validation.

10 · Sparse small-business orders

Only order, dispatch and outcome times exist, but five observed cases explain the gaps.

Route: quantify a narrow elapsed-time pattern without buying a permanent platform.

10 · Process Mining Evidence DesignerChoose what the log is allowed to support

Load a scenario or describe the evidence condition. The instrument chooses an analysis route, not an automation answer.

Figure 8. This article’s live instrument. The route is constrained by semantics, perspective, coverage, privacy and inferential distance.

Redesign questionEvent semanticsCase or object perspectiveCoverageAnalysis needPrivacy conditionTriangulationAccountability

Use logs to quantify a known process pattern

Redesign questionMeasure a defined recorded pattern for an owned decision.

PopulationState eligible cases or objects, exclusions, interval and versions.

Event semanticsUse only domain-confirmed event and time meanings.

PerspectiveUse the case or object view aligned to the question.

CoverageKeep known blind spots beside the result.

Analysis rungDescriptive profile or bounded path view.

Privacy boundaryMinimise identity and prohibit performance reuse.

Supported claimA recorded pattern exists in the declared population.

RivalsRetain source, perspective and case-mix explanations.

TriangulationMatch effort to inference and consequence.

AI alternativeDefine only after the process cause is tested.

Non-AI alternativeTest policy, authority, data and workflow repair.

Responsible ownerNamed process owner decides.

ExperimentChange one cause and measure end-to-end outcome.

Closure and deletionClose the question and delete unnecessary extracts.

ReauthorisationReopen when question, source, meaning or purpose changes.

How to use the instrument

Complete the Event Log Evidence Contract first, then set the controls to the weakest material condition. Load several named scenarios to test the routes. A positive result permits only the stated analysis, not deployment of an AI component. A repair result authorises data and semantic work. An observation result authorises bounded inquiry. A stop result blocks the proposed use. The designer cannot establish causality, infer worker intent or determine whether an intervention will create value. Those claims require experiments and accountable outcome evidence.

11 · Redesign authorityClose the docket or reauthorise the evidence

A process-mining project is complete when an owner makes a bounded decision and the extract, derived models and permissions receive an explicit future.

The Process Mining Redesign Docket records the redesign alternatives, event contract, semantics and coverage findings, case or object perspective, analysis rung, reproducible parameters, path or conformance claim, triangulating and disconfirming evidence, privacy treatment, supported inference, rivals and invalid uses. It compares AI, deterministic, policy, workflow and no-change options. It names the experiment, interim protection, end-to-end outcome, deadline, closure and deletion.

Apply the Event Evidence Reauthorisation Rule when the question, population, event definition, source, extraction, timestamp, identifier, object relationship, system version, workflow, vendor, workforce, access, purpose, retention or affected consequence changes materially. Reauthorise after a logging incident, unexplained distribution shift, false conformance finding, privacy concern, worker chilling, new shadow work or divergence between mined performance and the end-to-end outcome.

A thought experiment: the graph stays still while reality changes

Suppose a discovered path remains visually identical for six months. During that period, a vendor changes when events are written, a policy creates a new discretionary check and workers move recovery into a protected channel. The stable graph could indicate a stable process or an increasingly blind sensor. Reauthorisation tests the evidence-generating system, not only the path distribution. Visual continuity is not proof that meaning held.

Process evidence earns, and can lose, authority

Process mining deserves a central place in AI redesign because it can force broad claims into observable populations and variants. Its precision becomes dangerous only when extraction choices disappear behind the picture. The disciplined move is neither to worship the log nor retreat to workshops. It is to bind traces to semantics, perspectives, missing work, outcomes and an owned decision.

When that boundary is visible, teams can quantify a known pattern quickly, repair the event foundation, add observation where meaning is contested, run a bounded conformance or variant study, or stop. The changed decision is simple: select evidence authority before selecting automation authority.

Evidence baseSources and foundations

These sources ground the article’s synthesis in process-mining methods, event standards, data quality, task fit, routine formation and current governance. Each source remains bounded by its version, jurisdiction and method.