Layer 4: Judgement Intelligence — Articulating Judgement Before Automation
v0.1 · 14 November 2026 · First issued 12 November 2026 · Custodian: BusinessGPS · Version index and changelog
<!-- vale BGPS.Denylist = NO --><!-- reason: taught here: O2O is defined at first use in the executive summary (Opportunity to Operations, a 27-step method), and section 12.2 gives its three tiles as the example -->
Position. Articulate judgement before you automate it. A system records the decision, but seldom the cues, options, doubts and trade-offs behind it. An enterprise that hands decisions to AI before it knows how those decisions are made loses the judgement that protects its results. Judgement Intelligence makes that judgement visible, testable and governable first. The aim is better decisions, not an artificial personality.
Executive summary
Judgement is the part of an enterprise’s capability its systems record least, and the part most at risk when work moves to AI. An enterprise performs through its systems, processes and data. It also performs through its people’s capacity to read a situation, weigh a trade-off, make a legitimate decision and learn from the result. That capacity sits at Layer 4 of the Five-Layer Activity Model, Managerial Judgment.
This paper is the companion to Layer 3: Workaround Intelligence. That paper sets out the evidence on the hidden work people do around their systems. Layer 3 shows what people do that the system cannot see. This paper asks why they do it, whether the reasoning is sound and who has the right to act on it.
The paper reaches eight conclusions.
- Judgement is a distinct organisational asset. It is not simply undocumented process content. The system records a decision, but seldom the cues, options, uncertainty, mental simulation and trade-offs behind it. Tacit and explicit knowledge interact continuously; one cannot simply be swapped for the other [1–3].
- Experience is not a reliable proxy for expertise. Expert intuition is most defensible where the task has learnable regularities and the person has had timely, accurate feedback. Seniority, eloquence and confidence are not enough [3, 5].
- Discovery must reconstruct specific decisions. Cognitive task analysis, the Critical Decision Method, knowledge audits, realistic scenarios, decision diaries and peer comparison reveal what routine system logs do not [6–9].
- An organisation’s judgement is distributed. CorporateMind, one of the two products on our roadmap that apply Judgement Intelligence, represents expertise networks, decision rights, competing viewpoints and organisational memory. It is not the average personality of the executive team [10–12].
- Human–AI teams are not automatically better. A 2024 meta-analysis covered 106 experiments and 370 effect sizes [13]. On average, human–AI combinations did worse than the best human or AI working alone, especially on decision tasks. The division of labour has to be designed and tested.
- A cognitive twin is a design proposition, not a copy of a mind. Research on cognitive digital twins is promising. It does not show reliable replication of an executive’s judgement in new situations [14].
- Personal and organisational judgement need different governance. ExecMind, the other roadmap product, represents an individual’s permissioned preferences and decision learning. CorporateMind holds authorised, evidence-linked collective knowledge and enforces the enterprise’s decision rights. Anything that moves from one to the other is reviewed first.
- The opportunity is decision quality, not artificial personality. BusinessGPS connects Layer 3 evidence, Layer 4 judgement and Layer 5 agent execution. It works inside the trust conditions of Domain 6 of the Transformation Capability Management standard. It also works inside O2O, short for Opportunity to Operations: our 27-step method for taking change from intent to running operations.
Autonomy follows a decision-quality gate. No decision class is handed to an agent until the quality of the judgement behind it, and the governance around it, have been measured. It is the Layer 4 stage of the legitimacy gate between Layers 3, 4 and 5 (Section 2.3).
1. What this paper covers, and how far its evidence reaches
1.1 The questions
The paper asks seven questions:
- What does the judgement at Layer 4 consist of?
- How is valuable knowledge told apart from intuition, habit or error?
- How is it elicited, recorded, tested and kept up to date?
- How do ExecMind, CorporateMind and Symbiotic Cognitive Twins use it?
- Where does AI help, and where does it introduce dangerous feedback?
- What do current standards and technology provide?
- How do these findings shape the BusinessGPS method?
1.2 The method
This is a targeted narrative synthesis of research. It is not a systematic review, a controlled experiment or a vendor certification. Sources span academic research, standards, peer-reviewed articles, public law, regulator guidance and vendor technical documentation. They were checked against publicly accessible sources as at 12 November 2026. Primary academic sources carry the theoretical claims. Vendor documentation supports descriptions of what a product does, not claims about its performance. The artefacts, dimensions and scores we propose are original designs, and Section 13 sets out how they are tested.
The strongest quantitative evidence cited here concerns human–AI team performance [13], expert elicitation [15] and the cognitive effects of AI use [16–17]. Those findings are useful. They do not automatically carry over to enterprise programmes such as ERP, or to senior executives. Several promising developments rest on small applied evaluations, simulations or working papers, and they are labelled as such.
1.3 Description, evaluation and authority are three different claims
Three statements are often confused:
- “This is how an expert normally decides” is a description.
- “This is a good way to decide” is an evaluation.
- “This person or agent may decide” is a statement of authority.
We treat them as three separate types of information, each with its own evidence. A manager’s right to decide over their own operations is a Layer 4 matter. The right to delegate a decision or an action to an agent is Trusted Agency, the authority face of Layer 5. The observed outcome is a fourth dimension. It can support or challenge an assessment, but it does not settle on its own whether the earlier decision was sound. A prudent decision can produce a bad result under uncertainty.
2. Layer 4 is the judgement behind the action
2.1 A working definition
Judgement at Layer 4 is the capacity of individuals and groups to do six things in a real situation:
- interpret incomplete or conflicting information
- recognise patterns and anomalies
- reason about causes and consequences
- weigh objectives and constraints
- choose and justify an action
- update their understanding when the outcome is known
It also includes knowing when not to act and when to seek another view. Judgement Intelligence is our name for the method that makes this judgement discoverable, testable and governable. Layer 4, Managerial Judgment, is where it applies.
Layer 4 judgement is related to knowledge management, business rules, behavioural profiles and decision records, but it is none of them. Some of it is articulated; some is tacit, embodied or relational. Some choices are formal; others emerge through coordination. Some expertise is reliable in familiar conditions and fails in unusual or fast-changing ones.
2.2 Eleven dimensions of judgement
We propose the taxonomy below. Each dimension has a question an observer can ask and an example of how it shows up in an enterprise.
| Dimension | Observable question | Illustrative enterprise example |
|---|---|---|
| Situation awareness | What is happening now? | A planner recognises a delayed supplier signal |
| Pattern and anomaly recognition | What does not fit? | A controller spots an unusual closing entry |
| Causal understanding | Why might this be happening? | A service lead finds the root cause behind an error code |
| Mental simulation | What happens if we do this? | A plant manager anticipates a downstream bottleneck |
| Uncertainty assessment | What do we know, and how sure are we? | A CFO qualifies a cash-flow scenario |
| Trade-off judgement | Which objectives must be balanced? | Procurement balances resilience against price |
| Contextual and social judgement | Which relationships and circumstances matter? | An account lead reads a fragile customer relationship |
| Ethical and control judgement | Is the action legitimate and acceptable? | A manager rejects a shortcut that defeats segregation of duties |
| Escalation and restraint | When should we ask, or stop? | An expert recognises unfamiliar boundary conditions |
| Metacognition | How might our own judgement be wrong? | An executive names their own confirmation bias |
| Collective coordination | Whose expertise and authority are needed? | A cross-functional team resolves conflicting priorities |
This is a proposed taxonomy, not an established psychometric instrument. Endsley’s model of situation awareness explains perception, comprehension and projection [7]. The other dimensions draw on several research traditions. They need operational testing before anyone uses them to score people.
2.3 How Layer 4 relates to the other four layers
The Five-Layer Activity Model places enterprise work by where it is performed:
- Layer 1, Standard Software Process, holds data, states and recorded transactions.
- Layer 2, Standard Operating Procedures, holds the expected workflows and policies.
- Layer 3, Contextual Workarounds, holds the informal corrections, spreadsheets, shadow work and deviations.
- Layer 4, Managerial Judgment, holds the knowledge and decisions that explain why people acted as they did. That includes cases where the work is neither a workaround nor a system defect.
- Layer 5, Agentic Potential, is where agents are orchestrated. Its authority face, Trusted Agency, decides what they may do. Agents there use approved knowledge and report outcomes back for learning.
A workaround at Layer 3 does not prove expertise at Layer 4. A user may override the system because they noticed a real constraint. They may equally have learned an obsolete habit or made a mistake. Layer 3 provides leads. Layer 4 supplies hypotheses about the reasons. Independent validation decides which explanation holds. What Layer 4 validates and authorises is what Layer 5 may act on. This is the legitimacy gate set out in Layer 3: The Workaround Layer. Layer 3 supplies observational truth. Layer 4 supplies organisational legitimacy. Only then can Layer 5 receive authority.
2.4 Core, Context and Signature classify the capability, and the judgement inherits it
Capabilities and processes are different views of enterprise work, related many-to-many. Neither is a simple decomposition of the other. A value stream gives the outcome. Core, Context and Signature classify capabilities first; the judgement at stake carries the classification of the capability it serves. Context judgement may be standardised where that is defensible. Signature judgement needs protecting and deliberate differentiation. A workaround or an expert insight is understood in its value stream and capability before it is classified.
3. What the research supports, and what it does not
3.1 Tacit knowledge can be articulated in part, never in full
Polanyi’s account of tacit knowing shows that people can recognise or do things they struggle to explain [1]. Nonaka’s theory of organisational knowledge creation describes the continuous interplay between tacit and explicit knowledge [2]. It also sets out the conditions under which individual knowledge can be articulated and amplified. Neither supports the claim that every tacit capability can be fully translated into text, rules or model weights. A transcript or a knowledge graph represents selected expertise. It is not the whole capability.
3.2 Intuition is trustworthy only where the environment teaches it
Kahneman and Klein’s joint 2009 paper gives a decisive test [3]. Credible intuition depends on regularities in the environment that can be learned, and on enough feedback and practice to learn them. A planner who meets the same observable supply exceptions again and again may develop useful intuition. A senior executive making a handful of one-off acquisitions under great uncertainty may not get learning signals of the same quality. So ExecMind keeps a person’s demonstrated areas of expertise separate from their preferences and their general confidence.
The deliberate-practice literature stresses meaningful feedback [5]. Later replication research has challenged how much practice volume alone explains the differences between experts [18]. The responsible conclusion is that relevant practice and feedback matter. Hours or tenure alone do not demonstrate decision quality.
3.3 Reconstructing real decisions reaches what generic interviews miss
The Critical Decision Method was designed to reconstruct real decisions made under time pressure, with incomplete information and changing conditions [6]. It draws out critical cues, recognition patterns, surprises, alternatives and how one case was told from another. Applied cognitive task analysis offers a more streamlined set of interview methods and knowledge audits [8]. Brown, Power and Gore show how these methods uncover complex cognition in context in management research [9]. They give better access to how experts think than generic requirements interviews. They remain open to recall errors and rationalisation after the fact, so we triangulate them with operational records, observation and independent examples.
3.4 Judgement is spread across people, tools and teams
Hollan, Hutchins and Kirsh argue that cognition is distributed across people, artefacts and technology [10]. Research on transactive memory describes how teams know who holds which expertise and coordinate around that knowledge [11]. These are the conceptual foundations for CorporateMind. Neither shows that a database, a knowledge graph or a large language model creates collective intelligence on its own. The design also has to preserve roles, who knows what, dissent, ownership and the ways people coordinate.
Edmondson’s field research links psychological safety to team learning [12]. That is a strong reason to treat trust and participation as foundations of Layer 4 discovery. It does not mean a reassuring interview script creates psychological safety by itself.
3.5 Calibration tests expertise against known outcomes
Cooke’s structured approach to expert judgement uses calibration questions with known answers, and suitable probability scoring, to evaluate judgements within a domain [15]. It is most useful where knowledge is subjective and observed outcomes are few. It needs careful design: unsuitable seed questions, selective data, inappropriate pooling of experts or too few cases can all create false precision. So the method records uncertainty and calibration evidence rather than reducing complex expertise to one confidence score.
3.6 The human–AI research is a warning, not a licence
Bainbridge’s classic analysis of the ironies of automation made a simple point [4]. Moving routine work to machines can leave people handling the rare, difficult exceptions, without the practice they used to get. Expertise has to be preserved and interventions rehearsed. A passive “oversight” role is not enough.
Vaccaro, Almaatouq and Malone ran a preregistered meta-analysis of 106 experiments and 370 effect sizes [13]. On average, human–AI combinations did better than humans alone, but worse than the stronger of the human or the AI working alone. The pooled synergy effect was g = −0.23 (95% confidence interval −0.39 to −0.07). Decision tasks showed a more negative result than creation tasks. This does not prove every human–AI decision system is worse. It means complementarity has to be designed and tested, not assumed.
Glickman and Sharot’s experiments on human–AI feedback loops, with 1,401 participants, found that bias was amplified under certain conditions of interaction [16]. Lee and colleagues surveyed 319 knowledge workers about 936 examples of their use of generative AI [17]. Higher confidence in the AI went with less self-reported critical-thinking effort, while the work shifted towards verification and stewardship. That survey is self-reported and is not direct evidence of lost skill. Together, the studies support explicit challenge functions, counterfactual testing and protection against a twin that simply agrees with its user.
An NBER working paper, first issued in 2025 and revised in August 2026, tests selective delegation between people and AI [19]. It does so in an online classification experiment. Its results support assigning work by relative confidence and human effort. They hold within that experiment and are not a universal rule for delegation.
4. Real judgement is uncovered from evidence, not from accounts alone
4.1 Every source of evidence is partial
A system event shows the action. An employee may later give a plausible explanation. Neither proves the explanation describes what actually caused the action. Observation shows behaviour, but not everything the person noticed or considered. The answer is triangulation across real events, documents, system traces, expert interviews, comparisons and outcomes.
4.2 Ten discovery techniques compared
| Technique | Main contribution | Best used for | Main limitation |
|---|---|---|---|
| Critical Decision Method | Reconstructs cues, surprises and options | Consequential, non-routine cases | Recall and rationalisation bias |
| Applied cognitive task analysis and knowledge audit | Exposes the hard cognitive demands of a role | High-value operational roles | Depends on interview skill and context |
| Shadowing and contextual inquiry | Shows work as it is actually done | The boundary between Layers 3 and 4 | Does not expose all the thinking |
| Scenario-based expert comparison | Compares responses under controlled conditions | Twin tests and calibration | Scenario realism, and outcome leakage |
| Decision diary and pre-mortem | Captures assumptions before the outcome | Strategic and executive decisions | Reporting effort and selectivity |
| After-action review | Links decisions to observed results | Repeated operational decisions | Hindsight bias |
| Structured expert elicitation | Captures uncertainty as distributions | Forecasting and judgement under uncertainty | Calibration needs relevant questions |
| Decision and process mining | Finds decision branches and predictors | Repeatable, logged decisions | Logs omit tacit context |
| AI-guided elicitation | Scales structured interviewing | Knowledge capture and follow-up | Hallucination and leading questions |
| Expert disagreement workshop | Surfaces hidden assumptions and competing priorities | Collective decisions | Hierarchy and groupthink can undermine it |
4.3 Eight stages of Layer 4 discovery
- Select a decision class. Define the business outcome, value stream, capabilities, decisions and boundaries. Separate recurring operational choices from strategic, one-off or legally sensitive ones.
- Establish the conditions for trust. This is Domain 6 of the Transformation Capability Management standard, social licence and absorption trust. Agree the purpose, participation, lawful access, permissions, limits on monitoring and employees’ right to review. Agree how any intention to automate will be communicated. Avoid individual performance surveillance by default.
- Collect genuine cases. Include routine successes, difficult exceptions, failures and counterexamples. Use records made at the time wherever possible. Seek several experts, not only the most senior.
- Reconstruct the timeline. Use the Critical Decision Method. Establish what was known at each moment, what was noticed and which options were considered. Then establish why the action changed and what was not yet known.
- Extract the decision knowledge. Capture the trigger, cues, evidence, assumptions, mental models, constraints, alternatives, uncertainty, escalation and acceptable outcomes. Record conflicts and open questions.
- Challenge and validate. Compare accounts with actual transactions, records and independently selected cases. Use blinded scenarios and peer comparison to test whether the knowledge generalises.
- Decide the disposition. Preserve it, teach it, augment it, formalise it as rules, redesign the process upstream, or automate a bounded part of it. Decide whether it may be shared with the organisation’s collective record.
- Monitor and learn. Track outcomes, overrides, exceptions, model drift, calibration and new workarounds. Keep versioned records, each with a named human owner.
4.4 AI interviewers help structure capture; they do not verify judgement
Roudiere and Lento published the Assistant–Scribe–Knowledge Checker (ASK) framework in 2026 [20]. It uses separate agents to run adaptive interviews, structure the knowledge and find gaps. It was evaluated in a consulting firm’s lessons-learned setting. Its results concern the quality and acceptability of the documents produced, not full extraction or verification of tacit judgement. Our method adds two steps to that pattern. A case-verification step links each claim to primary evidence. A challenge step proposes counterexamples. Those two steps are ours; they are not validated parts of ASK.
4.5 Questions that reach real judgement
Instead of asking only “What is your process?”, ask:
- “Recall a case where the system’s recommendation seemed wrong.”
- “What first made you suspicious?”
- “What would a novice have missed?”
- “Which evidence did you set aside, and why?”
- “What alternatives looked possible at the time?”
- “What would have made you stop or escalate?”
- “How confident were you then, before you knew the outcome?”
These questions work best when tied to a real case and its timeline.
5. The Decision Passport records judgement as evidence
5.1 A document repository is not enough
Unstructured interview notes are useful evidence. They are hard to evaluate, reuse, govern or connect to a decision workflow. The Decision Passport, our proposal, records individual cases separately from general heuristics and from policy rules. It also separates observed facts, expert assertions, analysts’ interpretations, AI-generated hypotheses and approved rules.
5.2 Four linked objects, and a fifth that holds the evidence
- Decision class: for example, “reallocate supply when a confirmed delivery is threatened”. It holds the outcome, process, capabilities, risk category, applicable policies and authorised roles.
- Decision instance: one event. It holds what was known at the time, the options, the decision, the confidence and the people involved. It also holds the consequences and the period over which the outcome is observed.
- Judgement pattern: a candidate generalisation, supported by one or more instances and challenged by counterexamples. It may be a cue-to-action heuristic, an exception rule, a risk indicator or a tacit dependency.
- Expertise claim: a claim, scoped to a domain, that a person or team can reliably make a certain kind of judgement. It carries evidence, limitations, consent and a date for review.
A fifth supporting object, the evidence item, links each assertion to its source records without copying sensitive content unnecessarily.
5.3 The minimum contents of a Decision Passport
| Group | Fields | What it controls |
|---|---|---|
| Identity | Passport ID; decision class; version; status | A stable lifecycle and ownership |
| Business context | Value stream; capabilities; process; objective | The link to the O2O method |
| Observed case | Trigger; timeline; information available; action | Fact kept separate from narrative |
| Reasoning | Cues noticed; assumptions; alternatives; trade-offs | Uncertainty preserved, not a fabricated chain of thought |
| Authority | Authorised role; limits of delegation; mandatory approvals | Governance, not personal preference |
| Outcomes | Intended result; realised result; measurement horizon | Protection against hindsight bias |
| Evaluation | Peer review; calibration; conflicts; confidence | Validation within conditions, not universal certification |
| Provenance | Evidence links; method; people; agent version | Reproducibility and audit |
| Privacy and intellectual property | Licence; consent or other legal basis; access rights; retention | The boundary between personal and corporate knowledge |
| Reuse | Decision service; prompt context; decision-model mapping; escalation | Permitted, versioned use |
No single numeric “judgement score” is issued to individuals without a validated construct and a justified use. It is safer to track coverage within a domain, factual support, predictive calibration and process compliance separately.
5.4 Example: a planning override (hypothetical, not a field finding)
A planner changes a supplier-dependent availability date in the ERP system. Layer 1 shows the changed planning value and its timestamp. Layer 3 shows a supporting spreadsheet. Layer 4 investigation finds that the planner recognised a seasonal delivery risk that the master data did not hold.
The Decision Passport records the original evidence and the planner’s assumptions at the time. It records the alternatives, the expected consequences and the actual result. It also records a question for other planners: “Would the same cue have justified intervening with a different supplier?”
The right response is not necessarily to automate the override. It may be to improve the master data, approve a formal planning rule, provide decision support or keep a human escalation route for rare exceptions.
5.5 How the objects connect
A knowledge graph can hold the links between the objects. Useful relationships include:
- a decision class requires a capability
- a decision instance occurred in a process, and uses evidence
- a judgement pattern is inferred from decision instances
- an expertise claim is demonstrated in a domain
- a role is authorised for a decision class
- an agent executed an approved action
- an outcome evaluates a decision instance
6. ExecMind applies the method to one person’s judgement
6.1 Decision support, not impersonation
ExecMind and CorporateMind are the two products on the BusinessGPS roadmap that apply Judgement Intelligence: ExecMind to the individual, CorporateMind to the organisation.
ExecMind is a personal, permissioned decision-support and learning system. It does not impersonate the user. It remembers the goals, preferences, operating principles, domain experience and past decisions the user chooses to give it. It also remembers how the user reacted to its advice. It makes no claim to know a person’s private thoughts. It does not infer competence from personality tests or predict all future behaviour from past conversations.
The distinction matters. Preference is a way of working. Expertise is a capability tested within a domain. Authority is a legitimate right to act. A person can be decisive without being accurate, knowledgeable without being authorised, and authoritative without having the best specialist expertise.
6.2 ExecMind’s seven modules
- Personal doctrine and preferences: declared principles, priorities, risk tolerance, preferred style of challenge, communication and limits.
- Domain knowledge and evidence: explicit professional knowledge, with its sources and checks that it is current.
- Episodic decision memory: selected decisions, the assumptions made at the time, the alternatives and the later lessons.
- Uncertainty and calibration: confidence estimates, forecasting record, recurring error patterns and the edges of the user’s expertise.
- Reasoning support: structured options, opposing perspectives, pre-mortems, reversibility and decision criteria.
- Learning and development: after-action reviews and deliberately varied scenarios that keep the user’s own judgement in practice.
- Permission and identity boundary: the user controls what is shared, and content the user wrote is always distinguished from content the AI inferred.
6.3 ExecMind challenges its user rather than agreeing with them
ExecMind separates “what you have tended to prefer” from “what the evidence supports”. When the user’s view conflicts with independent evidence, ExecMind says so and offers a reasoned alternative. It tracks disconfirming examples, uncertainty and whether its recommendations improve the quality of decisions. This is the design’s answer to the risks of amplified bias and reduced critical engagement found in the research [16–17].
6.4 How ExecMind’s judgement is tested
ExecMind starts from the user’s declared doctrine and selected past decisions. Case-based elicitation then draws out the reasoning behind them. Its responses are tested against a blinded set of scenarios, and its scope and uncertainty are set from the results. It advises before it acts. Any increase in autonomy needs a separate assessment of consequences, competence, security and authorisation, and that delegation is a Trusted Agency decision at Layer 5. ExecMind answers one question for its user: “What do I know about this decision, how do I know it, and what could I be missing?”
7. CorporateMind applies the method to the organisation’s judgement
7.1 An organisation’s judgement is a decision system, not an average of its executives
CorporateMind is not built by collecting executives’ personal twins and averaging their conclusions. Organisations have distributed expertise, formal authority, historical routines, contractual obligations and objectives that can conflict. The unit of analysis is the decision system, not a synthetic corporate personality. Distributed cognition and transactive memory are the conceptual starting points [10–11].
7.2 CorporateMind’s components
| Component | Content | Design safeguard |
|---|---|---|
| Enterprise intent | Strategy, objectives, value drivers | Traceable to approved strategy |
| Decision catalogue | Major recurring and strategic decision classes | Each linked to a business owner |
| Expertise map | Who knows what, and on what evidence | No ranking by status alone |
| Decision-rights map | Who may recommend, approve, act, contest or escalate | Enforced where tools act |
| Corporate decision memory | Evidence-linked cases, rationales and outcomes | Versioned, with its time context |
| Policy and controls | Risk appetite, legal, ethical and business rules | Policy kept separate from informal practice |
| Plural perspectives | Finance, operations, sales, customers and affected staff | Dissent preserved, not averaged away |
| Learning loop | Results, error detection, revisions and audit | Independent review of consequential cases |
7.3 Example: competing executive objectives (hypothetical)
A CFO argues for preserving cash while a sales director argues for an aggressive customer acquisition programme. These are not necessarily conflicting facts. They may be legitimate differences in priorities and assumptions. CorporateMind shows the shared evidence, the forecast ranges and the constraints that cannot be broken. It shows the consequences of each choice and who holds the authority to decide. It may propose options or criteria for deciding. The decision stays with the accountable executive unless a bounded delegation explicitly applies.
7.4 The knowledge-sharing airlock moves private experience into corporate knowledge deliberately
The airlock, our proposal, has six steps: select, distil, review, authorise, publish and revalidate periodically. It stops private ExecMind content being disclosed by accident. It separates the employer’s operational learning from a person’s own preferences and reflection. CorporateMind receives business-relevant generalisations with their provenance, not a verbatim copy of anyone’s private conversations or personal history. Ownership, confidentiality, lawful processing, intellectual property and redaction are assessed case by case.
7.5 Five measures against groupthink
- Experts submit independent views before the group discusses them.
- Minority interpretations are stored, with the conditions under which each could be right.
- For contested decisions, the evidence and assumptions are shown, not a single averaged “corporate consensus”.
- The record shows whether disagreement led to further investigation.
- Psychological safety [12] comes from how leaders behave and how the decision process is designed, not from an anonymous feedback form.
8. Symbiotic Cognitive Twins: people and AI decide together, without pretending to be equal
8.1 What a twin is, and the limits of its fidelity
A Symbiotic Cognitive Twin, as we propose it, is a decision-support relationship between a person and an AI system. The person can also be an organisation’s decision system. The relationship is permissioned and continually updated. It uses selected histories, domain knowledge, decision patterns, evidence and feedback. It assists, challenges and sometimes carries out well-bounded tasks. It is not a verified simulation of a person’s brain, a substitute legal decision-maker or proof of anyone’s identity or competence.
CogTwin, presented at IJCAI 2025, is one example of research into hybrid cognitive digital twin architectures [14]. Its findings include simulated use cases and early feasibility claims. They do not show high-fidelity modelling of human executives. So we present “cognitive twin” as a design concept with explicit fidelity limits.
8.2 Five modes of working, each with its own release test
| Mode | What the twin does | The human’s role | Typical release test |
|---|---|---|---|
| Reflect | Recalls earlier decisions and assumptions | Reviews and corrects | Provenance and permission checks |
| Challenge | Surfaces errors, contrary evidence and missing options | Evaluates independently | Reliability under adversarial tests |
| Recommend | Compares scenarios and trade-offs | Decides | Accuracy and calibration in the domain |
| Co-decide | Structures deliberation between several parties | The accountable authority resolves | Audit of evidence and governance |
| Delegate | Carries out bounded, low-risk classes of work | Oversees and can revoke | Verified policy, testing and runtime controls, under Trusted Agency |
Moving down the table is never automatic. An agent may be safe to carry out a narrow, reversible data-preparation task and still be unfit to recommend a high-value acquisition.
8.3 The learning loop
- The person sets the objective and authorises the interaction.
- The twin retrieves approved evidence and the relevant Decision Passports.
- The twin builds alternatives, with confidence levels and points of challenge.
- The person accepts, changes, rejects or escalates. For delegated actions, a policy engine checks the authority.
- The action and its evidence are recorded. Uncertain or unusual cases are reviewed.
- Outcomes are observed over a suitable period. Independent evaluators assess the quality of prediction, any harm and the business value.
- Changes to general patterns or to an agent’s permissions pass a change-control gate.
The design must avoid recursive self-endorsement. An agent’s earlier statement is not independent evidence because another agent retrieved it. Source lineage is stored, and results verified outside the system are kept apart from AI-generated summaries.
8.4 Eight red-team questions for a cognitive twin
- Would the twin reach the same recommendation with the user’s name removed?
- Does it change its position appropriately when the objective evidence changes?
- Can it explain what was observed and what was inferred?
- Does it preserve disagreement?
- Does it detect conflicts with policy?
- Can it say “not enough information”?
- Does its confidence improve or deteriorate with repeated use?
- Would a different expert spot a consequence it missed?
9. Who should decide depends on the kind of decision
9.1 Four kinds of decision work
- Deterministic decisions rest on stable rules and known data. Standard workflow and decision tables suit them.
- Statistical judgements are predictions under uncertainty. Calibrated models with uncertainty intervals suit them.
- Contextual expert judgements depend on experience, relationships and exceptions. Expert augmentation with verification suits them.
- Governance or value judgements involve real conflicts between interests, legal duties and risk appetite. They need clear, accountable human decision rights.
The categories overlap. A large decision may include routine rule checks, predictive modelling, human interpretation and formal approval. So the decision is broken into parts rather than handed whole to one person or one agent.
9.2 A decision allocation matrix
| Conditions | Suggested arrangement | Why |
|---|---|---|
| High repeatability, low consequence, stable rules | Automated deterministic workflow | Strong control and auditability |
| Repeatable judgement, reliable outcomes, reversible | Model-assisted or bounded agent action | Testable and monitorable |
| Valuable specialist cues, ambiguous evidence | Human lead, with AI analysis and challenge | Preserves contextual knowledge |
| New strategy, sparse feedback, hard to reverse | Accountable human authority; AI explores options | A weak basis for learned automation |
| Regulated, legal or rights-affecting action | Specific legal and policy assessment | Legal safeguards and accountability |
| Weak data, high uncertainty, unknown distribution | Gather evidence, simulate, abstain or escalate | Avoids false precision |
Shrestha and colleagues identify several structures for human–AI decisions, including delegation, sequential and aggregated decision making [21]. Raisch and Krakowski argue that automation and augmentation cannot always be cleanly separated [22]. The matrix is a design heuristic, not a certified legal or safety classification.
10. Standards and tools model decisions; none of them recovers judgement on its own
Decision Model and Notation (DMN) is a formal way to represent repeatable decisions, their dependencies and their rules. The Object Management Group lists DMN 1.6 as the formal version published in September 2026 [23]. DMN suits structured decisions. It is not enough to model the whole of a senior executive’s judgement or their tacit reading of a situation. Knowledge graphs, case stores, retrieval-augmented generation, workflow, process mining, agent traces and scenario evaluation are complementary parts. None of them replaces expert discovery.
No vendor documentation we reviewed, as at 12 November 2026, shows that decision modelling or agent telemetry alone elicits and validates tacit expertise [24–27]. So the method concentrates on decision evidence, validation, permissions, learning and the link to business outcomes. It uses existing mining, rule and observability tools where they fit.
11. Governance, human factors and the law are part of the design
11.1 Social licence is a technical requirement
Layer 4 discovery can feel like an attempt to extract what makes an employee valuable. People may reasonably protect their job security, dignity, autonomy, reputation and professional discretion. A trust programme cannot promise that automation will not affect jobs unless that is genuinely the policy. It states the purpose, the benefits, the possible consequences, who sees what, the right to review and the routes for escalation. Domain 6 of the Transformation Capability Management standard, social licence and absorption trust, is a precondition for broad capture of employee knowledge.
11.2 Workplace monitoring and personal data
The UK Information Commissioner’s Office (ICO) requires worker monitoring to be proportionate, transparent and lawful. As at 12 November 2026, the ICO notes that its guidance is under review following the Data (Use and Access) Act 2025 [28]. The method prefers consented interviews, records people volunteer and evidence aggregated by role over blanket interception of employees’ conversations. Where personal data is processed, the method sets an appropriate legal basis. It assesses necessity, proportionality, security, retention and confidentiality, and whether a data protection impact assessment is needed. Consent at work may not always be freely given, and it does not settle every legal question.
11.3 Automated decisions need meaningful human involvement
The Data (Use and Access) Act 2025 revised the UK rules for significant decisions made solely by automated means using personal data [29]. It kept safeguards such as information, the right to make representations, human intervention and the right to contest. These rules depend on the facts. A human approval click is not necessarily meaningful human judgement. High-stakes decisions about people need their own legal review and clear accountability. This paper is research and product-design analysis, not legal advice.
11.4 Security, integrity and lifecycle governance
The NIST AI Risk Management Framework sets out how to govern, map, measure and manage AI risks [30]. The OWASP GenAI guidance for 2026 identifies the risks of large language model and agentic systems [31]. They include unauthorised tool use, compromised context and loss of control. The controls that follow include:
- least privilege and scope-limited access
- sandboxing and trusted-data boundaries
- source provenance and revision history
- approval checkpoints and step-by-step logs
- policy written as code
- incident response and rollback
Sensitive, unrestricted internal reasoning is not stored. The method keeps an auditable decision rationale and evidence trail, and makes no claim to see a model’s hidden cognition.
11.5 Eight ways it goes wrong
- Personality mistaken for competence: the twin sounds like the executive but is systematically wrong.
- Authority laundering: an agent takes an unauthorised action because an influential user asked it to.
- Consensus laundering: CorporateMind hides real disagreement behind a single, misleading answer.
- Knowledge leakage: personal ExecMind memories or confidential interview material reach shared search indexes.
- Hindsight contamination: explanations written after the outcome are presented as the reasoning before it.
- Feedback loops: the AI and the user reinforce each other’s bias instead of improving calibration [16].
- Deskilling: users stop checking the AI’s suggestions for themselves. A 2025 survey gives early grounds for concern, not proof that skills decline everywhere [17].
- Drift: changes in policy, people or markets invalidate judgement patterns that used to be useful.
12. Where Judgement Intelligence sits in the BusinessGPS method
12.1 Layer 4 holds the judgement; the products draw on it and add to it
Layer 4 is the layer of judgement and expertise. ExecMind and CorporateMind draw on it and contribute to it, each within its own boundary. At Layer 5, agents apply approved decision support or action only within explicit authorisation, and Trusted Agency governs that authorisation. Workaround Intelligence at Layer 3 and Judgement Intelligence at Layer 4 share their evidence and feedback models and keep their different purposes. Layer 3 asks what people do around the system. Layer 4 asks why, and whether the reasoning should be trusted.
12.2 Judgement Intelligence runs across all 27 steps of the O2O method
O2O, short for Opportunity to Operations, is our method for taking change from intent to running operations. It has 27 steps in three tiles of nine. Judgement Intelligence adds no steps. It runs through the existing ones.
| O2O tile | Layer 4 work | Evidence it produces |
|---|---|---|
| Strategy (steps 1 to 9) | Identify material decisions, strategic uncertainty and dependence on particular experts | A decision landscape and a map of judgement risk |
| Portfolio / Execution (steps 10 to 18) | Prioritise decision classes by value, criticality, learnability and risk. Within step 18, its cells 18.1 to 18.9 elicit, validate, test, redesign and deploy | A portfolio of judgement improvements and its business case; Decision Passports; tested decision services |
| Operations (steps 19 to 27) | Monitor the outcomes of judgement and the exceptions agents raise; within step 19, its cells 19.1 to 19.9 where they apply | Reviews of drift and calibration; validated learning |
The work stays cross-cutting. Value streams frame the outcomes. Processes show the sequence. Capabilities define what the business can do. Core, Context and Signature classify capabilities, and processes and activities carry that classification. Domain 6 has explicit gates throughout. It is not a one-off communications exercise.
12.3 Prioritise decision classes by consequence, not by time saved
Decision classes are prioritised by eight criteria: business consequence, frequency, uncertainty, concentration of expertise, available evidence, learnability, control criticality and strategic differentiation. Value comes from avoided error, faster response, resilience, better customer or operational outcomes, preserved expertise and reduced dependency, as well as labour saved. Estimated time saved multiplied by salary is not realised benefit. Added controls, verification work and the cost of change are counted too.
13. Research agenda: how the method is tested
13.1 The questions only field testing can answer
- Which interview techniques uncover the most useful knowledge for each hour of participants’ effort?
- How much of the judgement discovered generalises to cases not yet seen?
- How does an agent’s style of challenge affect calibration and acceptance?
- Can CorporateMind preserve minority expertise while resolving conflicts between decisions?
- Which privacy mechanisms most increase people’s genuine willingness to share?
- When does an AI twin produce better outcomes than a plain decision checklist or a reliable rules engine?
- How do decision models degrade when the operating environment changes?
13.2 Two contrasting tests: operational and executive judgement
The method is tested on two kinds of judgement, because they fail in different ways.
- Operational judgement is tested on a repeatable ERP decision, such as supply-planning exceptions, credit holds or purchase-order priority. Historical cases are reconstructed with several experts, the cues are identified, and recommendations are compared with recorded outcomes.
- Executive judgement is tested on decisions such as investment priorities, transformation portfolio choices or governance of material change. Carefully designed scenarios vary the evidence and define the decision criteria. Each decision-maker’s view is captured before any AI suggestion is shown. Objective ground truth may be late or disputed. So the test looks first at the quality of the process, the use of evidence and calibrated forecasts. It also asks whether important alternatives come to light.
13.3 The comparison is blinded and reports failures as well as gains
Where it is feasible and ethical, people alone, AI alone and people with AI are compared on a common, blinded set of cases. Historical and test data are kept apart, and no outcome information leaks into prompts. Case order is randomised where possible. Errors and their costs are defined in advance. Cases from outside the training distribution are included. Gains and failures are both reported.
The test measures absolute quality, calibration, decision time, quality of escalation, compliance, effects on trust and downstream outcomes. Participant satisfaction alone is not a result. The expert panel stays independent of the team that built the system.
13.4 Release thresholds are set per domain, and “do not automate” is a valid result
There is no universal numeric pass mark. Release thresholds are set for the specific domain and the cost of its errors. A low-risk suggestion may need only accurate retrieval and a user’s review. A high-consequence automation needs stringent tests of correctness, policy adherence, robustness, abstention, security and rollback. Autonomy always needs the policy owner’s approval and an explicit way to revoke it. At every stage, the right decision may be not to automate.
14. Conclusion: capability sovereignty is knowing where your judgement lives
Systems vendors can increasingly observe process deviations, model deterministic decisions and evaluate what AI agents do. BusinessGPS connects actual work (Layer 3), validated judgement (Layer 4) and governed automation (Layer 5), in the context of strategic value and trust.
An executive buys better decisions, resilience, faster learning and less dependence on irreplaceable people. They do not buy the claim that an AI has copied their mind. Our roadmap follows from that. CorporateMind is for organisational continuity that keeps the people whose expertise it holds. ExecMind is for private, evidence-linked support and an independent challenger. Symbiotic Cognitive Twins are a way of working together that can be evaluated continually.
Layer 4 is where the enterprise discovers and strengthens how its people and groups interpret, decide and learn. It is not an exercise in collecting personality profiles or turning all tacit knowledge into text. The strongest findings from psychology and organisational research call for close observation, elicitation in context, calibration, several perspectives and feedback. The research on human–AI performance carries a warning. A twin that seems helpful can add bias or lower quality if the division of labour is badly designed.
So Judgement Intelligence works with Workaround Intelligence at Layer 3, with ExecMind and CorporateMind, with Domain 6 and with the O2O method. It records decisions and their evidence, tests competence separately from confidence, enforces the boundaries of authority and measures whether decisions made together improve. Capability sovereignty is not achieved when a model appears to know everything. It is achieved when the enterprise knows what it knows and where its judgement depends on people. It also knows how reliable that judgement is and who is accountable for using it.
Appendix A: A field guide to interviews and case reconstruction
Opening statement. “We are studying how important decisions are actually made, including where systems and procedures do not capture all the context. We want to learn from successes as well as difficulties. We will keep your account separate from verified facts, let you review our interpretation and explain how any authorised learning might be used.”
Case selection. “Choose a recent, consequential and unusual decision where your experience made a difference. Please avoid a hypothetical example.”
Timeline. “When did you first realise a decision was needed? What information did you have at that point? What had not yet happened?”
Recognition. “What caught your attention? Was something inconsistent? What would a less experienced colleague have missed?”
Options. “What were the plausible alternatives at the time? Which option looked attractive first? What ruled the others out?”
Uncertainty. “What were you guessing about? How confident were you before you saw the result? What extra information would have changed your view?”
Boundaries. “Which constraints could not be broken? Who had the authority to approve or reject this action? When would you have stopped or escalated?”
Outcomes. “What happened? Which consequences became visible later? What would you do differently next time?”
Disconfirmation. “Tell me about a similar-looking case where the same approach would have been wrong. Could an expert colleague reasonably disagree with your judgement?”
Provenance. “Which documents, transaction records, emails or data could independently support this account? Which parts should stay private?”
Validation. Summarise the reconstructed case and ask the participant to point out anything inaccurate. Then test how general it is against an independent case and another expert.
Appendix B: An illustrative Decision Passport (hypothetical)
| Field | Hypothetical planning example |
|---|---|
| Decision class | Change the promised date for an at-risk order |
| Business objective | Improve reliable delivery without creating uncontrolled inventory |
| Trigger | Supplier lead-time risk conflicts with the availability the system calculated |
| Evidence available at the time | Supplier confirmation, past delivery performance, stock and customer priority |
| Expert cues | Recent decline in the supplier’s performance; short notice of a material shortage |
| Options | Keep the date; reallocate stock; negotiate delivery; escalate |
| Judgement | The planner predicts a high risk of shortfall despite the system showing availability |
| Uncertainty | How reliable the supplier’s revised estimate is, unknown |
| Constraints | Allocation policy, customer commitments, approval limits |
| Action | Re-plan a limited quantity after authorised approval |
| Outcome | Observe the actual supplier delivery and customer fulfilment |
| Status | Hypothesis awaiting independent validation |
| Potential Layer 5 role | Recommend alternatives; never bypass allocation approval |
Appendix C: Evaluation and release checklist
| Evaluation question | Evidence needed | What it means for release |
|---|---|---|
| Were the cues and assumptions captured accurately? | Signed-off cases and source records | Do not use a pattern the evidence does not support |
| Does the pattern generalise? | Held-out cases; counterexamples | Limit the domain if the pattern is unstable |
| Is the confidence meaningful? | Forecast calibration and measures of uncertainty | Require abstention if it is weak |
| Do people with AI beat the right baseline? | Tests of people alone, AI alone and both together | Assume no synergy by default |
| Are authority and policy respected? | Policy tests and negative scenarios | No execution if there are critical failures |
| Is private information protected? | Access controls, review, a data protection impact assessment where required | Block inappropriate sharing |
| Does it preserve independent human judgement? | Interaction experiments and analysis of overrides | Redesign if biased agreement rises |
| Does it improve real business outcomes? | Business measures tracked forward in time | Scale only with credible benefit |
Appendix D: How strong the evidence is for each claim
| Claim | Evidence status | Key references |
|---|---|---|
| Expert intuition depends on the environment and on feedback | Strong theoretical and empirical synthesis | [3], [5] |
| Cognitive task analysis and the Critical Decision Method can elicit otherwise hidden cues and strategies | Established applied methods; depend on context | [6], [8], [9] |
| Collective judgement is distributed across people and artefacts | Established conceptual and field literature | [10], [11], [12] |
| Synergy between people and AI in decisions is not guaranteed | Strong warning from a multi-study meta-analysis | [13] |
| Interaction with AI may amplify bias and affect critical-thinking effort | Controlled experiments plus a self-report survey | [16], [17] |
| AI-assisted interviewing can improve structured capture | Emerging evaluation, specific to its context | [20] |
| Digital twins accurately reproduce executives’ cognition | Not established by the research cited | [14] |
| The BusinessGPS Judgement Intelligence model and the Decision Passport | Designed; tested as Section 13 sets out | This paper |
<!-- vale BGPS.Denylist = YES -->
References
- Polanyi, M. (2009; first published 1966 by Doubleday). The Tacit Dimension. University of Chicago Press. https://press.uchicago.edu/ucp/books/book/chicago/T/bo6035368.html
- Nonaka, I. (1994). A Dynamic Theory of Organizational Knowledge Creation. Organization Science, 5(1), 14–37. https://doi.org/10.1287/orsc.5.1.14
- Kahneman, D., & Klein, G. (2009). Conditions for Intuitive Expertise: A Failure to Disagree. American Psychologist, 64(6), 515–526. https://doi.org/10.1037/a0016755
- Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6), 775–779. https://doi.org/10.1016/0005-1098(83)90046-8
- Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The Role of Deliberate Practice in the Acquisition of Expert Performance. Psychological Review, 100(3), 363–406. https://doi.org/10.1037/0033-295X.100.3.363
- Klein, G., Calderwood, R., & Macgregor, D. (1989). Critical Decision Method for Eliciting Knowledge. IEEE Transactions on Systems, Man, and Cybernetics, 19(3), 462–472. https://doi.org/10.1109/21.31053
- Endsley, M. R. (1995). Toward a Theory of Situation Awareness in Dynamic Systems. Human Factors, 37(1), 32–64. https://doi.org/10.1518/001872095779049543
- Militello, L. G., & Hutton, R. J. B. (1998). Applied Cognitive Task Analysis: A Practitioner’s Toolkit. Ergonomics, 41(11), 1618–1641. https://doi.org/10.1080/001401398186108
- Brown, O., Power, N., & Gore, J. (2025; online 2024). Cognitive Task Analysis: Eliciting Expert Cognition in Context. Organizational Research Methods, 28(3), 375–404. https://doi.org/10.1177/10944281241271216
- Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed Cognition. ACM TOCHI, 7(2), 174–196. https://doi.org/10.1145/353485.353487
- Lewis, K. (2003). Measuring Transactive Memory Systems in the Field. Journal of Applied Psychology, 88(4), 587–604. https://doi.org/10.1037/0021-9010.88.4.587
- Edmondson, A. (1999). Psychological Safety and Learning Behavior in Work Teams. Administrative Science Quarterly, 44(2), 350–383. https://doi.org/10.2307/2666999
- Vaccaro, M., Almaatouq, A., & Malone, T. W. (2024). When Combinations of Humans and AI Are Useful: A Systematic Review and Meta-analysis. Nature Human Behaviour, 8, 2293–2303. https://doi.org/10.1038/s41562-024-02024-1
- Mandal, S., & O’Connor, N. E. (2025). CogTwin: A Hybrid Cognitive Architecture Framework for Adaptable and Cognitive Digital Twins. IJCAI 2025, 9286–9295. https://doi.org/10.24963/ijcai.2025/1032
- Colson, A. R., & Cooke, R. M. (2018). Expert Elicitation: Using the Classical Model to Validate Experts’ Judgments. Review of Environmental Economics and Policy, 12(1), 113–132. https://doi.org/10.1093/reep/rex022
- Glickman, M., & Sharot, T. (2025; online 2024). How Human–AI Feedback Loops Alter Human Perceptual, Emotional and Social Judgements. Nature Human Behaviour, 9, 345–359. https://doi.org/10.1038/s41562-024-02077-2
- Lee, H.-P., Sarkar, A., Tankelevitch, L., et al. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025. https://doi.org/10.1145/3706598.3713778
- Macnamara, B. N., & Maitra, M. (2019). The Role of Deliberate Practice in Expert Performance: Revisiting Ericsson, Krampe & Tesch-Römer (1993). Royal Society Open Science, 6, 190327. https://doi.org/10.1098/rsos.190327
- Agarwal, N., Moehring, A., & Wolitzky, A. (2025; revised August 2026). Designing Human-AI Collaboration: A Sufficient-Statistic Approach. NBER Working Paper 33949. https://www.nber.org/papers/w33949
- Roudiere, S., & Lento, B. (2026). From Tacit Knowledge to Structured Documents: A Framework for Knowledge Elicitation. ECKM 2026 Proceedings, 27(2), 1096–1104. https://doi.org/10.34190/eckm.27.2.4875
- Shrestha, Y. R., Ben-Menahem, S. M., & von Krogh, G. (2019). Organizational Decision-Making Structures in the Age of AI. California Management Review, 61(4), 66–83. https://doi.org/10.1177/0008125619862257
- Raisch, S., & Krakowski, S. (2021). Artificial Intelligence and Management: The Automation–Augmentation Paradox. Academy of Management Review, 46(1), 192–210. https://doi.org/10.5465/amr.2018.0072
- Object Management Group. (2026). Decision Model and Notation (DMN) Specification, Version 1.6. https://www.omg.org/spec/DMN/1.6
- SAP. (accessed November 2026). SAP Signavio Process Manager: Decision Model and Notation. https://help.sap.com/docs/SIGNAVIO_PROCESS_MANAGER/user-guide/dmn-intro
- Salesforce. (accessed November 2026). Understand Agentforce Observability KPIs and Tools. https://help.salesforce.com/s/articleView?id=ai.agentforce_understand_observability_kpis_and_tools.htm&language=en_US&type=5
- Microsoft. (accessed November 2026). Evaluation in Microsoft Foundry. https://learn.microsoft.com/en-us/azure/foundry/concepts/evaluation-overview
- Microsoft. (accessed November 2026). Agent Tracing Overview. https://learn.microsoft.com/en-us/azure/foundry/observability/concepts/trace-agent-concept
- Information Commissioner’s Office. (accessed November 2026). Employment Practices and Data Protection: Monitoring Workers. https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/employment/monitoring-workers/
- Information Commissioner’s Office. (2025–26). Data (Use and Access) Act 2025: Data Protection Changes; see also UK legislation section 80. https://ico.org.uk/about-the-ico/what-we-do/legislation-we-cover/data-use-and-access-act-2025/the-data-use-and-access-act-2025-duaa-summary-of-the-changes/data-protection/ ; https://www.legislation.gov.uk/ukpga/2025/18/section/80
- NIST. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0); 2024 GenAI Profile. https://www.nist.gov/itl/ai-risk-management-framework
- OWASP. (August 2026). GenAI LLM Top 10 2026. https://genai.owasp.org/resource/owasp-genai-llm-top-10-2026/
Companion papers
- Layer 3: Workaround Intelligence: the evidence on the hidden work people do around enterprise systems, and the Layer 3 companion to this paper.
- Layer 3: The Workaround Layer: defines Layer 3, the Workaround Record and the legitimacy gate whose Layer 4 stage this paper sets out.
- The Five-Layer Activity Model: the parent framework; this paper is the depth paper for its Layer 4.
Version index and changelog
- v0.1 (14 November 2026): first issued.
From BusinessGPS source R&D.