Protecting Learning In Work When The Work No Longer Requires Learning

May 18, 2026

AI HI Series – Part 3

19 min read

Resources referenced in this article:

The Crossroads: An Interactive Leadership Intelligence Simulation – A 25-30 minute scenario-based simulation that baselines an individual’s capability across seven dimensions of human intelligence.
Download below: The Developmental Weight Index (DWI) – A printable practitioner tool for diagnosing which tasks in any role carry hidden developmental weight.

Developmental_Weight_Index_Tool (1) [Download PDF •16KB]

Disclaimer: Both tools draw from established, peer-reviewed research across cognitive science, moral psychology, organisational theory, and expertise development. Neither has been psychometrically validated. They are practitioner instruments designed to inform professional judgment, not replace it.

Everyone in L&D knows the 70:20:10 model. Seventy per cent of professional development happens through on-the-job experience. Twenty per cent through social learning – mentoring, feedback, observation. Ten per cent through formal training (Lombardo & Eichinger). The profession has spent decades building ever-better interventions for the 10 per cent: workshops, e-learning, leadership programmes, competency frameworks, evaluation models. It has influenced the 20 per cent through coaching schemes and mentoring programmes. And the 70 per cent? The 70 per cent took care of itself.

That wasn’t through negligence. When people did the work, the work developed them. The junior HRBP who sat through sixty disciplinary hearings was building judgment whether anyone designed it that way or not. The finance analyst who reconciled raw numbers month after month was developing an instinct for anomalies that no course could teach. The 70 per cent didn’t need L&D’s intervention because the development pipeline was embedded in the work itself.

AI changed the equation. Part 2 of this series argued that routine work (tame problems) is not just production – it is the training ground where human intelligence gets built. When AI absorbs the routine, the output may improve. But the developmental pathway underneath the task disappears.

And L&D, for the first time, needs to manage a part of development it has never been asked to manage at the task level – a part it has no vocabulary for, no diagnostic for, and no intervention model for.

Why the current toolkit doesn’t reach the 70 per cent

The L&D profession’s instruments – competency frameworks, training needs analyses, Kirkpatrick evaluations – were designed for the 10 per cent. They describe what a person in a role should be able to do, and they measure whether formal interventions helped them do it. They work well for that purpose. But they describe the destination, not the developmental pathway. When AI helps a professional reach the destination without building the intelligence for it – when the work no longer requires learning – it becomes a recipe for a disaster.

That intelligence underneath the tasks – the 70 per cent – is precisely what AI is now displacing. To protect it, we need a vocabulary precise enough to name it, a diagnostic concrete enough to identify where it lives, and an intervention model practical enough for L&D to deploy.

In this (somewhat lengthy) article, I introduce (1) a vocabulary: seven dimensions of intelligence that are categorically human – fundamentally different from AI; (2) a diagnostic tool: the Developmental Weight Index, a 7-question instrument that maps which tasks in any role are quietly building those dimensions; and (3) an intervention model: three moves L&D can make when AI threatens to displace those developmental pathways.

Seven dimensions of distinctively human intelligence

If you asked what makes human intelligence fundamentally different from artificial intelligence, the answer is architectural. AI operates through statistical pattern recognition over large datasets. It is extraordinarily good at finding structure in information. But it does not experience anything – not the situation, not the stakes, not the discomfort of not knowing. Human intelligence is embodied: shaped by sensation, emotion, identity, and lived consequence (Kahneman; Gardner).

That architectural difference produces a practical question: which specific dimensions of intelligence are not just difficult for AI to replicate, but categorically beyond its reach? Not “things AI does less well” – those gaps will close with compute and data. The question is: where does the absence of lived experience, embodiment, and genuine stakes make algorithmic replication impossible in principle?

That question is my selection filter. Drawing on research across cognitive science, moral psychology, organisational theory, and decision science, I have identified seven dimensions that meet this threshold. A note on scope: this is a practitioner framework, not an exhaustive taxonomy of all human cognitive strengths. The seven below are chosen because the strongest distinction is not output quality, but the absence of lived consequence, embodiment, accountability, and identity.

Contextual judgment under ambiguity – the capacity to make sound decisions when information is incomplete, contradictory, or politically charged, where no “correct” answer is derivable from data alone. The expert does not follow a decision tree; she responds to the situation as a whole, reading what it calls for rather than what the procedure says.
Ethical and moral reasoning – the capacity to recognise that a situation has an ethical dimension, to reason through competing values, to hold a line when the easiest path is the most ethically questionable, and to know that one will live with the consequences of their decision. It is about noticing that an ethical question exists and then acting on it under pressure.
Sensemaking and narrative construction – the capacity to impose coherence on messy, contradictory information by constructing a plausible story that enables action. AI can summarise information. It cannot look at five conflicting data points and develop a gut-feel for, “Here is what I think is actually going on.”
Relational intelligence and trust-building – the capacity to build, repair, and leverage trust in high-stakes relationships, requiring vulnerability, reciprocity, and genuine presence. It comprises judgments humans make about other humans, through direct interaction, over time. AI can simulate warmth but it cannot earn trust.
Adaptive risk-taking and courage – the capacity to take risks not probabilistically but existentially – staking reputation, career, or relationships on a judgment call where you might be wrong. The manager who raises an unpopular truth, the HRBP who escalates against a powerful leader – these acts require workplace courage: acting despite personal cost because the situation demands it.
Metacognition and calibrated self-awareness – this dimension carries a dual load. Calibration – knowing what you know and what you don’t – is one process. Bias detection – catching when your own heuristics are distorting your judgment – is a related but distinct one. Both require turning cognition back on itself, and both are built through repeated experience of being wrong in situations that matter. I keep them in a single dimension because in practice they develop together: the task that teaches you to catch an untested assumption is the same task that teaches you how much you don’t know.
Purpose-calibrated judgment – the capacity to connect work to personal values, identity, and legacy in ways that shape not just motivation but decision-making. This dimension operates differently from the six above – it is less a within-task skill and more an orienting force that runs across tasks. But it belongs here because it shapes judgment under ambiguity in ways no algorithm can replicate: the leader who knows what she stands for makes different decisions than the leader who is merely competent. Purpose acts as a compass in situations where data alone will not tell you what matters most.

These seven dimensions are built through repeated exposure to situations that demand them. The 70 per cent.

The Developmental Weight Index: a diagnostic for the 70 per cent

Once we have named the dimensions, the next practical question is: for any given role, which tasks in the 70 per cent are actually building these capabilities – and which are pure production that AI can safely absorb?

You cannot answer this by asking experienced practitioners to introspect on their own tacit knowledge. Polanyi’s insight – we know more than we can tell – means the senior HRBP cannot list the micro-signals she reads in a tense room. She just reads them. And you cannot answer it by auditing every task in a role against every dimension. A role may have a hundred tasks. That does not scale.

The Developmental Weight Index solves both problems. It is a 7-question Yes/No instrument administered to 2–3 experienced practitioners per function. Each question targets a specific behavioural indicator – concrete enough that the respondent can answer without understanding the underlying theory. Each question maps to a specific dimension, but the mapping is in the scoring key, not in the question.

An important boundary to name upfront: the DWI measures developmental demand – the extent to which a task requires a given dimension of intelligence. A task that scores high demands that capability from whoever performs it. Whether it actually develops that capability in a particular person depends on additional conditions: where the person currently sits on the novice-to-expert continuum, whether the task pushes the edge of their competence rather than sitting comfortably within it (Ericsson), and how reflectively they process the experience. Developmental demand is a necessary condition for development, not a guarantee of it. But it is the right thing to measure at the task level, because a task with zero developmental demand will develop nothing in anyone.

Q1. Have you ever handled this task differently depending on who was involved, what else was happening in the organisation, or the political dynamics at the time? → Contextual Judgment

Q2. When doing this task, have you ever had to form your own interpretation of what was going on before a clear or complete picture was available? → Sensemaking

Q3. Does this task put you in situations where the outcome depends on how well you read and respond to another person’s unstated concerns, resistance, or motivations? → Relational Intelligence

Q4. Has this task ever put you in a position where the easiest or most expected course of action wasn’t the one you believed was right? → Ethical Reasoning

Q5. Has this task ever required you to take a position, raise a concern, or make a call that carried personal or professional risk? → Risk-Taking & Courage

Q6. When doing this task, have you ever caught yourself operating on an assumption you hadn’t tested – and had to revise your thinking? → Metacognition

Q7. Does this task connect you to the real human, operational, or societal impact of your function – in ways you can see, not just read about? → Purpose-Calibrated Judgment

How to administer it: List all tasks in the role. Pre-filter: which could AI handle in the next 12–18 months? Only those go through the questionnaire. Ask 2–3 experienced practitioners to answer independently. For each question, majority rules – two of three say Yes, it scores 1. Sum for a Developmental Demand Score out of 7.

Scoring bands: 0–2, safe to automate. 3–4, partial protection – identify which sub-components carry weight. 5–7, protect.

Two worked examples: what the DWI reveals across functions

To make this concrete – and to demonstrate that the DWI is a general organisational diagnostic, not an HR-specific one – here are two worked examples: an L&D practitioner role and a finance business partner role. The tables show both the total score and the per-question breakdown, because the total tells you whether to protect, but the question-level detail tells you which dimensions each task builds.

Example 1: The L&D Practitioner

Task

Training Needs Analysis

Facilitating sessions

360 feedback debriefs

Coaching managers

Stakeholder reporting

Evaluating L3 effectiveness

Writing facilitator guides

Designing assessment centres

Curating external content

Vendor management

Developing content (slides)

Drafting compliance reports

Updating competency libraries

E-learning modules

LMS administration

Scheduling / logistics

Score

Verdict

Protect

Partial

Automate

Example 2: The Finance Business Partner

Task

Advising business leaders on investment decisions

Presenting financials to the board

Challenging business cases from operating units

Building the annual budget with BU heads

Forecasting under uncertainty (e.g. commodity / demand volatility)

Internal audit scoping – deciding what to examine and why

Post-mortem on a missed forecast – diagnosing what the model missed

Interpreting variances – separating signal from noise in monthly results

Preparing management commentary for quarterly results

Vendor contract negotiation (finance terms)

Monthly reconciliations

Consolidating subsidiary financials

Generating standard MIS reports

Processing expense claims

Bank reconciliation (routine matching)

Score

Verdict

Protect

Partial

Automate

A few things to notice – first within each table, then across them.

Within each table: read rows and columns separately. Read across a row to see what a task builds. Training Needs Analysis and advising business leaders on investment decisions both light up every dimension – they are among the most developmentally dense tasks in their respective functions. Drafting compliance reports (L&D) and monthly reconciliations (finance) both score 2, but for entirely different reasons – different dimensions, different developmental profiles hidden behind the same total.

Read down a column to see where a dimension lives. In the L&D table, relational intelligence (Q3) clusters in live, unpredictable human interaction – facilitation, coaching, 360 debriefs, vendor negotiation. If those tasks get automated or reduced, relational intelligence has very few remaining developmental pathways in this role. In the finance table, relational intelligence clusters differently – in advisory and boardroom tasks where the finance business partner must read the room, not just present the numbers.

Across the two tables: the same dimension develops through different work. Sensemaking (Q2) in L&D develops through interpreting what a stakeholder interview actually means versus what was said. Sensemaking in finance develops through looking at a variance report and constructing a narrative about what happened in the business. Same dimension, different task substrate. This matters because when AI displaces the task, the replacement – whether structured immersion or simulation – must replicate the specific sensemaking demand of that function, not a generic SJT exercise.

Now look at the tasks scoring 1–2 – the ones below the protection threshold but not zeros. Drafting compliance reports (L&D) scores 2: it carries a thin thread of ethical reasoning (Q4, because the practitioner must decide what to include, what to flag, and what language to use) and a connection to organisational impact (Q7). Consolidating subsidiary financials scores 2: contextual judgment (Q1, because consolidation rules differ across entities and jurisdictions) and metacognition (Q6, because experienced practitioners learn to catch their own assumptions about inter-company eliminations). These are tasks where something human is happening, but not enough to justify protection. In practice, the right move for tasks in this band is to automate the production but preserve the decision point – let AI draft the compliance report, but keep the human deciding what gets flagged and how. Let AI run the consolidation, but keep the human reviewing the elimination entries.

This column-level view is what makes the DWI more than a task-ranking tool. It becomes a map of where the 70 per cent lives in any given function – and where it is most vulnerable to displacement.

From diagnosis to a personalised development path

The DWI tells you which tasks build which dimensions. But different people in the same role will have different strengths and gaps. A junior L&D practitioner who came from a facilitation background may be strong on relational intelligence and sensemaking but underdeveloped on metacognition and ethical reasoning. A peer who entered L&D through analytics may show the opposite pattern. A junior finance analyst who came through audit may have sharp metacognition but underdeveloped relational intelligence; one who came through commercial finance may show the reverse.

To create a personalised development path, you need a baseline – a way to see where the individual currently stands across the seven dimensions. Situational Judgment Tests and scenario-based simulations are well-suited to this purpose. They place the individual in ambiguous, multi-stakeholder situations with no clean answer and observe how they navigate – which is precisely what the seven dimensions describe.

I have built one such instrument: The Crossroads, an interactive leadership simulation that places a participant into a product launch crisis at a fictional FMCG company, where evidence arrives in fragments, stakeholders have unclear motives, and a late-stage disclosure reframes the entire situation. It produces a profile across all seven dimensions. It is a prototype, not a validated psychometric – but it illustrates how baselining works and what a dimension-level development profile looks like in practice.

Once you have a baseline, the development path writes itself. If the baseline shows you are weak on metacognition, the DWI table tells you which tasks in your role specifically build metacognition (Q6 column). Those are your priority tasks – protect your exposure to them, seek immersion in them, or if they have been automated, find or build a simulation that replicates the metacognitive demand they carried. The baseline tells you what you need. The DWI tells you where to find it. The gap between the two is your development plan – located squarely in the 70 per cent.

The three-move hierarchy: L&D’s intervention toolkit for the 70 per cent

Once you know what to protect and where the individual’s gaps are, you have three moves available, in order of preference.

First preference: protect the task. Do not automate it, or use AI as a scaffold where the human still does the judgment-intensive part. This is the strongest move because the organic developmental pathway stays intact. Bjork’s research on desirable difficulties showed that conditions which make learning slower and more effortful in the short term produce more durable and transferable knowledge in the long term. When AI removes the struggle from a task, it removes the mechanism through which intelligence is built. For tasks scoring 5–7 on the DWI, protecting the task should be the default recommendation unless the efficiency pressure is irresistible. Ericsson’s research on deliberate practice suggests a design principle for scaffolding: keep the human engaged at the edge of competence, with the AI handling components that carry no developmental weight.

Second preference: structured immersion. When a task must be partially automated but experienced practitioners still exist, structured apprenticeship preserves the tacit-to-tacit transfer. Consider what this looks like in practice across functions. If AI absorbs most of the analytical work in a Training Needs Analysis, the junior L&D practitioner loses her exposure to the raw, messy, contradictory data that builds her sensemaking instinct. But if she co-conducts the TNA with a senior practitioner – sitting in on stakeholder interviews, watching how the senior HRBP reads between the lines of what a business leader says, participating in the interpretive conversation about what the data actually means – the developmental pathway is preserved even though the production task has shifted. Similarly, if AI handles the bulk of a monthly reconciliation, the junior finance analyst loses the repetitive exposure that builds her anomaly-detection instinct. But if she reviews the AI-generated reconciliation alongside a senior analyst – learning to spot what the algorithm flagged that doesn’t actually matter and what it missed that does – the tacit knowledge transfer is preserved. The output is still AI-assisted. The learning is still human.

Third preference: tailored simulation. When the task has been fully automated and no organic exposure exists, or when the developmental moment is too rare or too high-stakes to engineer naturally, a designed simulation becomes the alternative developmental pathway. This is where scenario-based learning targeted to specific dimensions sits – not generic leadership case studies, but simulations deliberately constructed around the dimension profile the DWI revealed. If a function’s DWI map shows that ethical reasoning (Q4) and courage (Q5) are concentrated in only two tasks, both of which are being automated, that is a simulation design brief: build a scenario that creates the ethical tension and social pressure those tasks used to provide.

Across all three moves: the structured debrief. Whether someone is doing the real task, shadowing an expert, or navigating a simulation, the debrief is where tacit learning becomes conscious. Gibbs’ reflective cycle provides a practical structure for this: moving from description (what happened?) through feelings, evaluation, analysis, conclusions, and finally an action plan for next time.

The DWI’s seven questions can serve as prompts within this reflective structure. After a significant experience, walk through them not for scoring but for reflection: did I form my own interpretation, or accept someone else’s? Did I catch an assumption? Did I read what was underneath the surface? Did I take a position that carried risk? Used this way, the questions become a lens for debrief’s analysis and evaluation stages – turning a generic reflective cycle into one specifically tuned to the developmental dimensions that matter. This is how you accelerate the novice-to-competent transition even when the organic pathway is compressed.

The manager’s role: where the 70 per cent actually lives

There is one actor conspicuously absent from the model so far: the line manager.

L&D can diagnose developmental weight, build simulations, and design immersion programmes. But the 70 per cent happens in the line manager’s territory. The manager decides which tasks a team member gets exposure to. She decides whether to delegate the messy stakeholder conversation or handle it herself. She decides whether the AI tool gets deployed for speed or held back because the junior team member needs the developmental struggle.

Most managers make these decisions based on output efficiency, not developmental value – and no one has ever given them a vocabulary to do otherwise. The DWI changes that. A DWI map for a function gives the manager a concrete, task-level view of where development sits in her team’s workflow. When she is deciding whether to automate a task, the DWI score tells her what she is potentially giving up. When she is deciding how to allocate work across a team, the column-level view tells her which team members are getting rich developmental exposure and which are being starved of it.

This does not require the manager to become a learning designer. It requires a single conversation – L&D sitting with the manager, walking through the DWI map for her function, and asking: given what we now know about where developmental weight sits, which automation decisions should we rethink, and which task allocations should we adjust? That conversation is where the 70 per cent gets protected or lost. Without it, the DWI is a diagnostic without a decision-maker.

The 70 per cent, claimed

Seventy per cent of professional development happens through on-the-job experience. I cannot claim that L&D has been absent from this territory – action learning, stretch assignments, developmental job design, and the CCL’s own research on developmental experiences have all operated here. But these interventions work at the level of role and assignment. They answer the question: which experiences develop people? What they have not provided is a task-level, dimension-specific diagnostic that can tell you precisely which tasks within a role are building which capabilities – and what is lost when a specific task is automated. That is the gap AI is now forcing open, and it is the gap this framework addresses.

The seven dimensions name the human intelligence that the 70 per cent builds. The Developmental Weight Index diagnoses where it lives in any given role. The baseline – through situational judgment tests or simulations – personalises the development path. The three-move hierarchy – protect, immerse, simulate – gives L&D an intervention toolkit calibrated to the task level, not just the role level. And the manager conversation is where diagnosis turns into decision.

Learning needs deliberate protecting when the work no longer requires learning. And for the first time, L&D has the task-level precision to protect it.

Acknowledgment: This article owes a debt to my husband, Amit Aggarwal, whose observation that the DWI framework applies not just to AI displacement but to the 70 per cent of on-the-job learning more broadly was the insight that expanded its scope from a defensive argument to a constructive one.

End-Notes and References

70:20:10 Model: Lombardo, M.M. & Eichinger, R.W. (1996), The Career Architect Development Planner, Lominger. Originally derived from survey research at the Center for Creative Leadership. The model identifies three sources of professional development: challenging assignments (70%), developmental relationships (20%), and coursework/training (10%). While the precise ratios have been debated, the directional insight – that most development happens through experience, not instruction – remains widely accepted in L&D practice. Directly relevant: AI automation threatens the “challenging assignments” channel for the first time.
Prior L&D interventions in the 70 per cent: Revans, R.W. (1980), Action Learning: New Techniques for Management, Blond & Briggs – action learning as structured on-the-job development. McCauley, C.D., Ruderman, M.N., Ohlott, P.J. & Morrow, J.E. (1994), “Assessing the developmental components of managerial jobs,” Journal of Applied Psychology, 79(4), 544–560 – identified specific job characteristics (unfamiliar responsibilities, high stakes, scope) that drive on-the-job development. These approaches work at the role and assignment level; the DWI extends the diagnostic to the task level within roles.
Dimension 1 – Contextual Judgment: Dreyfus, H.L. & Dreyfus, S.E. (1986), Mind Over Machine, Free Press. Maps expertise development from novice to expert, showing that true expertise operates through holistic pattern recognition rather than rule-following. Relevant: explains why task exposure, not instruction, builds judgment.
Dimension 2 – Ethical Reasoning: Rest, J.R. (1986), Moral Development: Advances in Research and Theory, Praeger. Identifies four components: moral sensitivity, moral judgment, moral motivation, and moral character. The distinction between recognising an ethical dimension and reasoning about it is critical for understanding what repeated exposure to real ethical micro-situations builds. See also Treviño, L.K., Weaver, G.R. & Reynolds, S.J. (2006), “Behavioral ethics in organizations,” Journal of Management, 32(6), 951–990.
Dimension 3 – Sensemaking: Weick, K.E. (1995), Sensemaking in Organizations, Sage. Extended by Maitlis, S. & Christianson, M. (2014), “Sensemaking in organizations: Taking stock and moving forward,” Academy of Management Annals, 8(1), 57–125. Bartunek, J.M. & Moch, M.K. (1987) provide the first-order / second-order / third-order coding framework in Journal of Applied Behavioral Science, 23(4), 483–500.
Dimension 4 – Relational Intelligence: Mayer, R.C., Davis, J.H. & Schoorman, F.D. (1995), “An integrative model of organizational trust,” Academy of Management Review, 20(3), 709–734. The ability–benevolence–integrity framework explains how trust is built and why AI cannot build it.
Dimension 5 – Risk-Taking and Courage: Schilpzand, P., Hekman, D.R. & Mitchell, T.R. (2015), “An inductively generated typology and process model of workplace courage,” Organization Science, 26(1), 52–77. Risk perception versus risk propensity distinguished via Weber, E.U., Blais, A-R. & Betz, N.E. (2002), Journal of Behavioral Decision Making, 15(4), 263–290.
Dimension 6 – Metacognition: Flavell, J.H. (1979), “Metacognition and cognitive monitoring,” American Psychologist, 34(10), 906–911. Kruger, J. & Dunning, D. (1999), “Unskilled and unaware of it,” Journal of Personality and Social Psychology, 77(6), 1121–1134. Koriat, A. (2007), “Metacognition and Consciousness,” in Zelazo, Moscovitch & Thompson (Eds.), Cambridge Handbook of Consciousness – distinguishes monitoring accuracy (calibration) from control processes (strategy selection), supporting the dual-load framing used in this article.
Dimension 7 – Purpose-Calibrated Judgment: Frankl, V.E. (1946), Man’s Search for Meaning, Beacon Press. Professional identity construction in Ibarra, H. (1999), “Provisional selves,” Administrative Science Quarterly, 44(4), 764–791. The job–career–calling distinction from Wrzesniewski, A. et al. (1997), Journal of Research in Personality, 31(1), 21–33.
Taxonomy scope and acknowledged omissions: Guilford, J.P. (1967), The Nature of Human Intelligence, McGraw-Hill — foundational taxonomy of divergent thinking. Senge, P.M. (1990), The Fifth Discipline, Doubleday – systems thinking as a distinct cognitive discipline. Varela, F.J., Thompson, E. & Rosch, E. (1991), The Embodied Mind, MIT Press – embodied cognition. These capabilities are genuinely important and genuinely human, but the “categorically different” versus “currently superior” line is harder to draw for them, which is why they are acknowledged here rather than included in the core framework.
Tacit knowledge and knowledge creation: Polanyi, M. (1966), The Tacit Dimension, Doubleday. Nonaka, I. & Takeuchi, H. (1995), The Knowledge-Creating Company, Oxford University Press. The socialisation mode – tacit-to-tacit transfer through shared experience – is the theoretical basis for the immersion intervention.
Expertise and learning theory: 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. Bjork, R.A. (1994), “Memory and metamemory considerations in the training of human beings,” in Metcalfe & Shimamura (Eds.), Metacognition: Knowing About Knowing, MIT Press. The “desirable difficulties” framework: conditions that slow learning in the short term produce more durable knowledge.
Apprenticeship and situated learning: Lave, J. & Wenger, E. (1991), Situated Learning: Legitimate Peripheral Participation, Cambridge University Press. Collins, A., Brown, J.S. & Newman, S.E. (1989), “Cognitive apprenticeship: Teaching the crafts of reading, writing, and mathematics,” in Resnick (Ed.), Knowing, Learning, and Instruction, Erlbaum.
Reflective practice and debrief design: Gibbs, G. (1988), Learning by Doing: A Guide to Teaching and Learning Methods, Further Education Unit, Oxford Polytechnic. Provides the six-stage reflective cycle (description, feelings, evaluation, analysis, conclusion, action plan) used as the structural basis for the DWI debrief protocol. Boud, D., Keogh, R. & Walker, D. (1985), Reflection: Turning Experience into Learning, Kogan Page – emphasises attending to emotional responses as a necessary component of reflection, not just cognitive replay.
Assessment and simulation design: Lievens, F. & Patterson, F. (2011), “The validity and incremental validity of knowledge tests, low-fidelity simulations, and high-fidelity simulations,” Journal of Applied Psychology, 96(5), 927–940. Thornton, G.C. & Rupp, D.E. (2006), Assessment Centers in Human Resource Management, Lawrence Erlbaum.
AI and intelligence architecture: Kahneman, D. (2011), Thinking, Fast and Slow, Farrar, Straus and Giroux. Gardner, H. (1983), Frames of Mind: The Theory of Multiple Intelligences, Basic Books.
Wicked problems: Rittel, H.W.J. & Webber, M.M. (1973), “Dilemmas in a general theory of planning,” Policy Sciences, 4(2), 155–169.

AI HI Series