Layer 3: Workaround Intelligence — The Evidence on the Hidden Work of Enterprise Systems
v0.1 · 14 November 2026 · First issued 12 November 2026 · Custodian: BusinessGPS · Version index and changelog
Enterprise systems record only part of the work that makes them succeed, and a transformation that treats the recorded part as the whole will automate the wrong things. This paper sets out the published evidence behind that claim and the method that follows from it. It is the evidence base for its companion, Layer 3: The Workaround Layer, which defines Layer 3 and the record each workaround needs. Read that paper for the definition; read this one for the research underneath it.
Abstract
<!-- vale BGPS.Denylist = NO --><!-- reason: taught here: O2O is defined at first use in the abstract (Opportunity to Operations, 27 steps in three tiles of nine), and section 7.3 gives its tiles as the example -->
Large enterprise systems formalise a subset of the work needed to deliver operational outcomes. In complex implementations, employees bridge the gap between the prescribed workflow and real demands. They use manual reconciliations, spreadsheets, decisions taken off the system, informal coordination, exception handling and local routines. These practices can compensate for poor design, integration or data. They can also express valuable expertise and the capacity to adapt.
Process mining reveals only the events for which a record exists. On its own, it cannot explain the purpose, the reasoning or the organisational context of hidden work. This paper synthesises research on workarounds, invisible labour, cognitive task analysis, process mining and psychological safety. It sets out a discovery method that combines several sources of evidence, and a governed Workaround Intelligence capability. O2O, short for Opportunity to Operations, is BusinessGPS’s 27-step method for taking change from intent to running operations. The capability sits within the Five-Layer Activity Model, the O2O method and Domain 6 of the Transformation Capability Management standard, ‘Social licence and absorption trust’. The paper also examines how AI agents may discover or propose new workflows, and the new risks they bring: unauthorised shortcuts and undocumented work. The outcome is an evidence-led approach that separates work to eliminate from expertise to preserve, augment or codify, before any automation decision is made. [1–7, 14–16]
This is a synthesis of published research and documentation, not original empirical research. The illustrative case is not field-study evidence. Sources were checked on 12 November 2026.
Keywords: enterprise systems; SAP; Salesforce; workarounds; invisible work; shadow IT; process mining; task mining; tacit knowledge; AI agents; social licence; business transformation.
Executive findings
- The intended process, the recorded system events and the actual work are three different forms of evidence. No single source is a complete record of operations.
- A workaround is not automatically a defect. It may be a legitimate adaptation, a sign of hidden expertise or a control risk.
- The best discovery combines process and task data with exception tracing, inspection of the artefacts people use, direct observation and structured elicitation of expert judgement.
- Employees reveal informal practice only when they trust the exercise. That trust depends on credible participation and on being told the truth about possible consequences for their jobs.
- Agents may find recurring patterns and test alternatives, but must not be allowed to bypass controls merely because an alternative appears to complete a task.
- The method uses existing detection technology where it exists and adds human-led discovery.
Research questions and scope
The paper addresses six questions:
- Why do large implementations generate workarounds?
- Which forms of invisible and manual work matter most?
- Which methods best reveal and explain them?
- What could a technically feasible scanner detect?
- Can agents produce new, potentially better workarounds?
- How do these findings shape the BusinessGPS method?
The focus is large-scale enterprise implementations, with application to SAP, Salesforce and mixed landscapes. The paper does not quantify how often each cause occurs by sector or implementation, because the cited literature does not establish a universal distribution of root causes.
1. A workaround is a purposeful adaptation, and hidden work is wider still
A workaround is a purposeful adaptation made when a person or team meets an obstacle on the prescribed route to a work objective. Alter’s theory treats workarounds as a process shaped by actors, constraints, incentives and the alternatives people believe are open to them. Longitudinal research on enterprise systems shows that workarounds may be developed collectively. They can become unofficial local rules and shape how a global system is institutionalised. [1, 2]
‘Hidden work’ is broader. It includes the effort needed to coordinate, reconcile, repair, negotiate, interpret and complete work that formal process descriptions and transaction records do not capture. Star and Strauss show that visibility depends on the observer, the indicators used and the organisational context. Work can be essential and still be weakly represented in management information. [3]
1.1 Three kinds of hidden work need three different responses
| Category | Defining characteristic | Default response |
|---|---|---|
| Compensatory work | Informal steps correct for limits in the system, data, integration or process. | Investigate the root cause; redesign or remove the need. |
| Necessary discretionary work | Expertise is applied to ambiguous, exceptional or high-consequence circumstances. | Capture the decision logic; preserve or augment the judgement. |
| Unsafe or unauthorised bypass | A workaround circumvents controls, duties, permissions or required evidence. | Remove or constrain the bypass and solve the underlying cause. |
This is BusinessGPS’s taxonomy. It is not a prevalence ranking. A single practice can contain more than one category. For example, a credit-control spreadsheet might compensate for poor reporting, embed sound credit judgement and bypass formal approval checks, all at once.
1.2 Six families of root cause, each with its own diagnostic question
| Root-cause family | Illustrative manifestation | Diagnostic question |
|---|---|---|
| Functional or configuration fit | A missing exception path or an unsuitable configured rule. | Is the needed outcome within the system’s intended scope? |
| Data and integration | Recurring corrections or manual re-keying. | What information fails to flow correctly? |
| Process and control design | Shadow approvals, duplicate steps or late escalations. | Is the prescribed sequence workable in operation? |
| Usability and capability | Private checklists, macros and step-by-step guides. | Is the barrier skill, design or access? |
| Local operational context | Site-specific scheduling or supplier practices. | What constraint is absent from the template? |
| Incentives and authority | Dependence on informal gatekeepers or local discretion. | What risk or benefit does the arrangement create? |
These families are not a ranking by prevalence. A diagnosis needs evidence. A spreadsheet is a signal to investigate, not proof that the system has failed.
2. People hold the knowledge, so discovery depends on their trust
Experienced employees are often the last safeguard that stops a technically compliant transaction producing a poor business result. A planner may know that a supplier is unreliable. A finance specialist may recognise a false reconciliation. A customer-service representative may read an exception through knowledge that is absent from the master data. None of these is a sequence of clicks. Each involves cues, mental models, trade-offs and discretionary judgement.
2.1 “The system still needs me” is evidence to investigate, not resistance
The claim that “the system still needs me” can express several things. It may be genuine responsibility, valued expertise, personal identity, concern for job security, a preference for autonomy, or a mix of these. It must not be read as resistance or deliberate obstruction without investigation. Learning from workarounds means understanding which practices help an organisation reach its outcomes, and which persist only because the formal environment does not work. [2, 3]
Psychological safety matters because revealing a workaround can expose an employee to reputational and compliance risk. Edmondson’s research found an association between team psychological safety and learning behaviour. For discovery, the implication is direct. A programme framed as surveillance or as automatic headcount reduction may undermine the quality of the evidence it seeks. [4]
2.2 Five terms make a workable social contract for discovery
- State the purpose: to understand and improve operational outcomes, not to rate individuals covertly.
- Involve respected practitioners in scope, diagnosis and the evaluation of solutions.
- Distinguish good-faith adaptation from deliberate circumvention of controls. Do not promise immunity from legitimate investigation.
- Give employees a route to challenge or correct how their work has been interpreted.
- Explain possible role changes and automation consequences accurately. Do not imply guarantees that management cannot honour.
This is a direct application of Domain 6 of the Transformation Capability Management standard, ‘Social licence and absorption trust’. Trust is not a communications accessory to implementation. It is a precondition for obtaining reliable accounts of how work actually happens.
3. Mixed methods find what any single method misses
Each method has a different boundary on what it can observe, so a mixed approach is preferred. Transaction records identify unusual flows. Desktop observation reveals manual activity. Interviews and cognitive techniques explain purpose and judgement. Used alone, each can misclassify legitimate practice as waste or miss important work done off the system.
| Method | Strongest evidence | Principal limitation |
|---|---|---|
| Process mining | Recorded variants, delays, rework, deviations and case paths. | Invisible steps do not appear in the event data. |
| Task mining | Desktop actions, transfers and switching between applications. | Purpose and judgement cannot be reliably inferred. |
| Exception-led tracing | How a difficult case was actually resolved. | Can over-sample unusual cases. |
| Contextual inquiry and shadowing | Live artefacts, informal interactions and recovery work. | Needs a large sample; the observer changes what is observed. |
| Cognitive task analysis | Expert cues, uncertainty, reasoning and trade-offs. | Needs expert elicitation and validation. |
| Artefact and dependency mapping | Shadow spreadsheets, local tools and knowledge held by one person. | Ownership of artefacts may be incomplete. |
| Comparative case analysis | Different successful approaches across sites or teams. | Differences in context can confound the comparison. |
3.1 Event and process mining shows what was recorded
Event logs allow the variants of a business process, and its patterns of rework, to be reconstructed. The SWORD study demonstrates semi-automated identification of workarounds using 22 patterns, 13 of which can be applied without prior domain knowledge, followed by expert validation. Its evaluation found 11 distinct types of workaround across 2 hospital departments. This is good evidence of technical feasibility. It is not a benchmark of sensitivity or precision that transfers to a multi-country enterprise resource planning (ERP) estate. [5]
Object-centric process mining matters when orders, invoices, deliveries and customers take part in overlapping processes. It avoids the distortion of forcing every operational event into one isolated case. [7]
3.2 Observation and exception reconstruction show what was actually done
For selected cases, reconstruct the whole journey. That means the first signal, the applications involved, local files, communications, decision points, corrections, handoffs and the final outcome. Ask employees to show the artefacts they actually use, rather than relying on a description of the official process. Assess a representative sample of successful, delayed and failed cases.
One effective interview prompt is: “Show me the last time the standard process did not work. What told you something was wrong, what did you use instead, what did you decide, and how did you know it was safe?”
3.3 Cognitive task analysis reaches the judgement at Layer 4
Applied cognitive task analysis gives structured techniques for eliciting the mental demands of a task and the expertise behind it, rather than only documenting visible actions. Its interview methods can surface how experts recognise cues, handle uncertainty and weigh trade-offs, and the uncommon situations that matter most. [6]
| Elicitation prompt | Knowledge sought |
|---|---|
| What first suggested the transaction was wrong? | Recognition cues and thresholds |
| Which source was trusted, and why? | Hierarchy of evidence and reliability of data |
| Which options were rejected? | Decision rules and unacceptable risks |
| What would a new colleague miss? | Tacit competence and ways of failing |
| How was success confirmed? | Outcome criteria and independent controls |
The output is a decision model with its source confidence, trigger conditions, alternatives, constraints and validation requirements. A macro or a spreadsheet is not the same as the knowledge that decides when its results can be trusted.
4. The Workaround Intelligence scanner raises hypotheses, then builds confirmed knowledge
The scanner is a discovery service that draws on several sources and links every finding to its evidence. It is not a general tool for monitoring employees. Its first purpose is to raise plausible hypotheses for investigation. Its second is to build structured operational knowledge once findings are confirmed.
4.1 Each source gives signals, and each signal carries a caution
| Source | Signals to investigate | Important caution |
|---|---|---|
| ERP, customer relationship management and workflow systems | Corrections, reversals, reassignments, repeated overrides and unusual variants. | These may be normal, authorised controls. |
| Integration and batch monitoring | Recurring interface errors, reprocessing, restarts and retries. | Some recovery patterns are designed. |
| Desktop tasks | Repeated copy and paste, re-keying, downloads and uploads. | Activity is not intent or value. |
| Shared artefacts | Critical spreadsheets, linked workbooks, macros and local databases. | Content may be commercially sensitive. |
| Tickets and approved communications | Frequent escalations, status chases and manual approvals. | Collection needs lawful, proportionate scoping. |
| Agent execution traces | Tool substitutions, retries, unsanctioned sequences and bypass attempts. | A successful result does not mean acceptable conduct. |
4.2 An operational graph connects evidence that a process map keeps apart
The Workaround Intelligence graph links business objects, process events, artefacts, roles, decisions, constraints, controls, evidence and outcomes. The same spreadsheet may support several processes. One issue may span ordering, logistics and billing. One expert decision may affect several cases. A linear process map alone does not represent these relationships well. The graph is worked in six stages:
- Discover: take in only approved sources and normalise the identifiers of events and objects.
- Detect: run statistical, rule-based and AI-assisted identification of candidate patterns.
- Explain: generate hypotheses and retrieve the provenance of every inference.
- Validate: confirm purpose, cause and actual consequence with operational experts.
- Decide: classify the intervention and set its owner, approval and evidence requirements.
- Monitor: compare performance after the change, control outcomes and any new adaptations that emerge.
4.3 Every confirmed workaround needs a minimum record
The companion paper defines the Workaround Record as the Layer 3 object. The register is the set of those records. At minimum, each holds:
| Data group | Required content |
|---|---|
| Identity and context | Identifier; value stream; related capabilities and processes; system and site. |
| Trigger and execution | Observed trigger; actions taken; artefacts; who or what performed them. |
| Interpretation | Cause, hypothesised or validated; the human judgement embedded; confidence. |
| Business consequence | Frequency, capacity, delay, quality, risk and dependency. |
| Governance | Policy and control implications; source provenance; approval and ownership. |
| Disposition | Retain, repair, standardise, redesign, augment, automate or retire. |
| Lifecycle | Detected, investigated, verified, approved, implemented and monitored. |
4.4 Illustration: a delivery promise that looked like rework was judgement
This example is hypothetical. It does not represent a measured result.
A scanner detects repeated changes to delivery commitments in an SAP system, correlated with spreadsheet updates and operational emails. Rather than assume redundant data entry, the analyst reconstructs the affected orders. A planner explains that the workbook holds supplier reliability information that the planning data lacks, and judgement about which customer commitments are critical. The finding is reclassified. It moves from probable low-value rework to a mix of data deficiency, a weakness in the planning model and high-value discretionary judgement. The preferred intervention improves the source data and the planning logic, documents the decision model, and leaves ambiguous cases with an accountable expert.
5. AI agents can find better routes, and can also create new hidden work
Agents can detect patterns across systems, compare practices, generate explanations, propose new workflows and execute authorised tasks. Yet a technical path an agent proposes is not necessarily legitimate or economically sound. The important question is not whether an agent can complete a task. It is whether the agent can reach the intended outcome within business, security, legal and control boundaries.
5.1 Agent authority comes in three levels, each with its own gate
| Mode | Permitted activity | Required decision gate |
|---|---|---|
| A. Investigate | Correlate evidence; propose a candidate explanation or workaround. | Expert confirmation of meaning and root cause. |
| B. Design and simulate | Propose alternative workflows; simulate and test them in approved environments. | Approval on control, value and performance. |
| C. Execute | Perform bounded, approved actions with audit and escalation. | Production authorisation, runtime limits and monitoring. |
The likely value lies in two places. One is recognising common root causes among local adaptations that look unrelated. The other is testing whether a workflow native to the system, or integrated with it, removes the need for manual compensation. An agent could also create fresh shadow work through unapproved tool sequences, temporary files, excess retries or bypassing a control. The Open Worldwide Application Security Project (OWASP) lists excessive agency as a material risk, entry LLM03:2026 in its Top 10 for LLM Applications 2026, published on 3 August 2026. Authority must therefore be limited to the actions and data needed for the assigned outcome. [15]
5.2 An authority-and-evidence contract sets the agent’s limits
At Layer 5 of the Five-Layer Activity Model, this contract is the authority face, Trusted Agency: what an agent is trusted to do, and on what evidence. It sets out:
- the explicit business objective, the objects in scope and the success criteria
- permitted tools, identity, data boundaries and operating limits
- prohibited actions, including unapproved changes to controls or master data
- mandatory human approval for irreversible, high-value or policy-sensitive actions
- structured records of actions and outcomes, with useful provenance and proper privacy protection
- escalation and safe-stop behaviour for unknown exceptions, material uncertainty or conflict with policy
- simulation before deployment, regression tests and monitored review after deployment
The National Institute of Standards and Technology (NIST) AI Risk Management Framework gives a useful governance pattern: govern, map, measure and manage the risks. Such frameworks complement validation by operational experts. They do not replace it. [16]
6. Software alone does not reconstruct why a workaround exists
No vendor documentation reviewed shows that software alone can reliably reconstruct the tacit reasons for a local workaround, or guarantee that employees will accept a new workflow. [8–13] That gap is why the method combines tools with human-led discovery.
7. Workaround Intelligence works across the layers, the method and the standard
7.1 The Five-Layer Activity Model: Layers 1 to 5
The Five-Layer Activity Model places every activity in an enterprise on one of five layers, by where the work is performed. Workaround Intelligence has a different job at each layer.
| Layer | Workaround Intelligence responsibility |
|---|---|
| Layer 1: Standard Software Process | Discover formal transactions, controls and documented system evidence. |
| Layer 2: Standard Operating Procedures | Compare prescribed workflows with recorded and observed execution. |
| Layer 3: Contextual Workarounds | Identify shadow artefacts, manual compensation and informal coordination. |
| Layer 4: Managerial Judgment | Elicit expert cues, decisions, constraints and exception logic. |
| Layer 5: Agentic Potential | Apply approved automation; monitor exceptions and adaptations that agents create. Its authority face is Trusted Agency. |
The central design rule is to keep observing a Layer 3 artefact separate from understanding the Layer 4 judgement that directs its use. An agent must not be allowed to reproduce consequential manual work solely because the sequence can be observed.
7.2 Core, Context and Signature classify the capability, not the transaction
Capabilities do not simply decompose into processes. A capability describes what an organisation can do; a process describes how activities are coordinated. The two relate many to many. That matters when one workaround spans several processes, or when one capability supports several value streams. Core, Context and Signature classify capabilities first. Processes and activities carry the classification of the capability they realise, and a process can be Signature in its own right.
| Capability classification | Treatment of the activity discovered |
|---|---|
| Context | Rent or delegate to platforms. No custom investment. |
| Core | Decide consciously how to own it: rent, buy, borrow or build. |
| Signature | Articulate it before any AI is pointed at it, and protect it from leaking into vendor models. |
7.3 The 27-step O2O method: a cross-cutting method, not a new sequence
O2O, short for Opportunity to Operations, is BusinessGPS’s method for taking change from strategic intent through delivery into running operations. It has 27 steps in three tiles of nine. Workaround Intelligence runs across all three tiles. It adds no steps.
| O2O tile | Discovery and decision contribution | Output |
|---|---|---|
| Strategy (steps 1 to 9) | Expose hidden dependencies, value leakage and constraints on capability. | An assessment of opportunity and exposure. |
| Portfolio / Execution (steps 10 to 18) | Prioritise interventions by value, risk, feasibility and capacity for change. Within step 18, its delivery cells 18.1 to 18.9 elicit knowledge and redesign, test and implement changes. | An approved portfolio of work and investment, and a validated operational and technical solution. |
| Operations (steps 19 to 27) | Monitor exceptions, agent actions and new adaptations. Within step 19, its absorption cells 19.1 to 19.9 carry the change into operations. | Continuing learning and evidence of benefits. |
7.4 Domain 6: a readiness gate on trust comes before discovery
A readiness gate establishes four things. Do operational experts understand the purpose? Do they trust the process enough to take part? Can they challenge the conclusions? Do they have the capacity to adopt the changes? Discovery and transformation then proceed in turns. Employee evidence improves the design of the system, and each demonstrated improvement makes the next round of discovery more credible. This is how Domain 6, ‘Social licence and absorption trust’, applies to operational discovery.
8. A workaround’s value is measured in outcomes, not hours removed
8.1 Measures and the factors of priority
Operational measures include observed manual minutes per case, exception frequency, rework, elapsed recovery time, rates of error and correction, dependencies on one person, coverage of documented decisions, control issues and the improvement realised after changes. A business case built only on hours removed is inadequate. A manual control may prevent a costlier loss. A fix to the system may remove recurring effort more reliably than automating its correction.
A priority score for interventions can be built from six factors:
- verified impact
- confidence in the root cause
- control risk
- feasibility
- how critical the knowledge is
- readiness for change
Weightings are agreed locally and validated against observed outcomes. This paper recommends no universal weights.
8.2 The method starts from read-only data and voluntary evidence
The method starts from read-only process and exception data and targeted, voluntary evidence collection. It builds the register and the expert-validation workflow before any investment in comprehensive desktop monitoring or autonomous remediation in production. It works alongside SAP, Salesforce and existing process-mining and task-mining products. The first proof is better decisions, not a large-scale scanning deployment.
9. Governance, limits and the research agenda
9.1 Privacy, labour and control set the boundary of collection
The UK Information Commissioner’s Office (ICO) guidance states that monitoring workers must have a lawful basis and be fair, transparent, necessary and proportionate. As at 12 November 2026, the ICO notes that this guidance is under review following the Data (Use and Access) Act 2025. High-risk monitoring requires a data protection impact assessment. Employers should use the least intrusive effective method and avoid function creep, the slow widening of a system’s use beyond its stated purpose. This matters most for desktop recording and the analysis of communications, which may collect personal or sensitive content. [14]
The architecture therefore applies purpose restriction, access controls, aggregation and pseudonymisation where appropriate, documented retention, consultation with employees, routes to challenge and human review. Technical availability does not authorise indiscriminate data collection. This paper is not legal advice. Any implementation needs a privacy and employment-law review for its jurisdiction.
9.2 What the evidence does not yet show
- The literature gives conceptual and case-study support. It gives no universal estimate of how common workarounds are in ERP implementations.
- The SWORD findings come from a bounded healthcare study. They must be validated independently for SAP and Salesforce data.
- Human motives cannot be reliably inferred from logs, communications or observed work alone.
- The scanner, the register and the readiness gate are designed. Section 9.3 sets out how they are tested in the field.
9.3 How the method is tested, and what follows
The method is tested one process at a time. A candidate process has adequate event data, a high burden of exceptions, experts who can be reached and modest privacy risk. Purchase-to-pay, order-to-cash, financial close, inventory reconciliation and resolution of service exceptions are all candidates. Each test asks four things: is the method useful, can it be measured, will employees take part, and what does it cost?
Detection is evaluated on a small labelled sample of cases, recording both the false-positive rate and the rate of missed workarounds. The ground truth may be incomplete, so evaluation includes structured adjudication by experts, not agreement with the event log alone.
Further research should:
- test the precision and recall of detection across different types of process
- compare discovery from events alone with mixed-method discovery
- study employee participation and perceived surveillance
- assess whether alternatives proposed by agents improve outcomes without eroding controls
- develop a reusable way of representing knowledge that links Layer 3 artefacts to Layer 4 decisions and Layer 5 execution
10. Conclusion: understand why the work exists before deciding what to do with it
Enterprise transformation misunderstands the business when it equates system transactions with work. A workaround can reveal faults in the architecture, gaps in the global template, missing integration, needless bureaucracy, adaptive local knowledge or a dangerous bypass of a control. Its significance cannot be judged from the fact that it occurs. The method triangulates system evidence, manual activity, expert judgement and organisational context before choosing an intervention.
Workaround Intelligence is a cross-cutting capability joining the Five-Layer Activity Model, Domain 6 and O2O. What sets it apart is the disciplined move from spotting an unusual activity to understanding why it exists. That means knowing what business outcome it protects, what knowledge it holds, and whether repair, redesign, preservation, augmentation or automation is the right response.
Design principle: Discover the real work. Understand the human judgement. Fix the underlying cause. Automate only what has been validated and authorised.
References
Academic research
- Alter, S. (2014). “Theory of Workarounds.” Communications of the Association for Information Systems, 34, Article 55. DOI: 10.17705/1CAIS.03455.
- Malaurent, J., & Karanasios, S. (2020). “Learning from workaround practices: The challenge of enterprise system implementations in multinational corporations.” Information Systems Journal, 30(4), 639–663. DOI: 10.1111/isj.12272.
- Star, S. L., & Strauss, A. (1999). “Layers of silence, arenas of voice: The ecology of visible and invisible work.” Computer Supported Cooperative Work, 8, 9–30. DOI: 10.1023/A:1008651105359.
- Edmondson, A. (1999). “Psychological safety and learning behavior in work teams.” Administrative Science Quarterly, 44(2), 350–383. DOI: 10.2307/2666999.
- van der Waal, W., van de Weerd, I., Beerepoot, I., Lu, X., Kappen, T., Haitjema, S., & Reijers, H. A. (2025). “Putting the SWORD to the Test: Finding Workarounds with Process Mining.” Business & Information Systems Engineering, 67(2), 171–190. DOI: 10.1007/s12599-023-00846-3.
- Militello, L. G., & Hutton, R. J. B. (1998). “Applied cognitive task analysis (ACTA): A practitioner’s toolkit for understanding cognitive task demands.” Ergonomics, 41(11), 1618–1641. DOI: 10.1080/001401398186108.
- van der Aalst, W. M. P. (2023). “Object-Centric Process Mining: Unraveling the Fabric of Real Processes.” Mathematics, 11(12), 2691. DOI: 10.3390/math11122691.
Official product, regulatory and risk-management sources
- SAP Learning. “Learning SAP Signavio Process Insights.” Product capability and improvement-recommendation documentation. https://learning.sap.com/courses/analyzing-business-processes-with-sap-signavio-solutions/learning-sap-signavio-process-insights_de1f24b8-b1a1-4eb2-8324-768ffe40a778
- SAP Help Portal. “SAP Cloud ALM: Integration & Exception Monitoring.” Product documentation. https://help.sap.com/docs/cloud-alm/applicationhelp/im-monitoring?locale=en-US
- SAP Signavio. “What is AI Agent Mining?” Concept and product-direction overview. https://www.signavio.com/wiki/process-discovery/ai-agent-mining/
- Salesforce Help. “Agentforce Testing Center.” Product documentation. https://help.salesforce.com/s/articleView?id=ai.agent_testing_center.htm&language=en_US&type=5
- Microsoft Learn. “Overview of process mining and task mining in Power Automate.” Product documentation. https://learn.microsoft.com/en-us/power-automate/process-advisor-overview
- UiPath. “Automation discovery (Preview).” UiPath for Coding Agents user guide; feature identified as preview. https://docs.uipath.com/coding-agents/standalone/latest/user-guide/automation-discovery
- Information Commissioner’s Office (UK). “Data protection and monitoring workers.” Regulatory guidance. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/data-protection-and-monitoring-workers/
- OWASP. (3 August 2026). “LLM03:2026 Excessive Agency.” OWASP Top 10 for LLM Applications 2026, PDF v1.0. https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/
- National Institute of Standards and Technology. “AI Risk Management Framework Core.” Governance guidance. https://airc.nist.gov/airmf-resources/airmf/5-sec-core/
The interpretations of research are BusinessGPS’s synthesis. Vendor materials establish stated capabilities, not independently tested efficacy. Links and product descriptions were checked on 12 November 2026.
Appendix A. Operational interview prompts
- Walk through the last genuinely difficult case, using the actual record and artefacts.
- What normally makes the official procedure inadequate, slow or unsafe?
- Which spreadsheet, local database, template, paper record or message is indispensable?
- What tells an experienced person that the system’s recommendation is untrustworthy?
- Who is contacted before a high-consequence override, and why?
- Which exceptions are easy, which are dangerous, and how are they told apart?
- Which practices would continue if the system were replaced tomorrow?
- What would break if the most experienced person were unavailable?
- Which one change to a rule, data source or interface would improve the work most?
- What would a well-designed agent need to know before taking the same action?
Appendix B. Example disposition logic
| Finding | First question | Possible treatment |
|---|---|---|
| Repeated re-entry of data | Why are the systems not exchanging trusted information? | Repair the integration or the master data. |
| Manual approval of exceptions | Is this judgement, a control or avoidable delay? | Retain, simplify or augment. |
| Shadow capacity spreadsheet | What constraint or signal is absent from planning? | Improve the planning model and capture the knowledge. |
| Monthly close that depends on one person | Which rules and evidence exist only in that person’s practice? | Document the decisions; build resilience; automate under supervision. |
| Agent bypasses the intended workflow | Was the shortcut authorised, and was the outcome independently verified? | Suspend, investigate, revise permissions and test. |
Appendix C. Candidate scanning rules and validation gates
The rules below are candidates for detection. They are not evidence of wrongdoing or measured potential for automation. Each needs an independent explanation and a control review before any intervention.
| Candidate pattern | Screening signal | Mandatory validation |
|---|---|---|
| Correction loop | The same business object is modified repeatedly after an error or rejection. | Is the correction expected, required or avoidable? |
| Reconciliation off the system | Frequent exports from ERP followed by recurring workbook calculations. | What logic is missing, and which results are trusted? |
| Integration repair | Error and retry events correlated with manual changes and later success. | Is the intervention necessary, approved and repeatable? |
| Informal approval | A decision or release follows an unrecorded external communication. | Was the approval authorised and evidenced? |
| Expert bottleneck | Cases systematically pause for one specialist’s intervention. | Is the knowledge tacit, the control necessary, or access constrained? |
| Agent detour | The tool chain or update sequence differs from the approved path. | Was it authorised, and were the business and control outcomes sound? |
A staged validation protocol
- Confirm the source event and the business object with traceable, permitted evidence.
- Establish the operational purpose by inspecting artefacts or reconstructing with an expert.
- Test at least one alternative explanation, so that normal work is not misclassified.
- Assess business impact, privacy, authority, control and the consequences of failure.
- Approve a disposition, an owner, a measure after the change and a review date.
<!-- vale BGPS.Denylist = YES -->
Companion papers
- Layer 3: The Workaround Layer: defines Layer 3, the Workaround Record, the six evidence channels and the legitimacy gate between Layers 3, 4 and 5. This paper is its evidence base.
- Layer 4: Judgement Intelligence: the judgement behind the work, and how it is articulated, tested and governed before anything is automated. It takes up where Section 3.3 leaves off.
- The Five-Layer Activity Model: the parent framework, which places every enterprise activity on one of five layers.
Version index and changelog
- v0.1 (14 November 2026): first issued.
From BusinessGPS source R&D.