One-line proposition

AI changes work most profoundly not by replacing people, but by redistributing cognitive work between humans and machines — creating the opportunity to remove mundane and repetitive tasks while strengthening human judgement, learning and organisational cognition.

What is changing

Much discussion about artificial intelligence and work begins with the question:

Which jobs will AI replace?

That framing is too narrow. Jobs are bundles of activities that involve different combinations of information processing, routine execution, interpretation, interpersonal interaction, contextual knowledge, creativity, ethical judgement and accountability. AI will not necessarily affect all of those activities in the same way.

A more useful question is therefore:

What work should machines do, what work should humans do, and what work can they do better together?

Recent research on Human–AI collaborative decision-making offers a useful way to address that question. Li and Tian (2026), drawing on a systematic review and bibliometric analysis of 627 articles, argue that Human–AI decision-making is not a single mode of interaction. Instead, different configurations distribute reasoning and decision-making between humans and AI.

This shifts the problem from simple human substitution to cognitive work design.

Redistributing cognitive work

Li and Tian (2026) identify four Human–AI collaborative decision-making paradigms.

Programmed Algorithmic Decision describes structured, rule-based environments in which AI performs much of the information processing and decision-making. Human participation may be limited unless an exception or anomaly occurs.

Adaptive Intuitive Decision remains primarily human led. AI provides information processing, pattern recognition and analytical support, while people retain primary responsibility for interpreting the situation and deciding what to do.

Interpretive Analytical Decision involves deeper collaboration. AI expands analytical capacity, searches larger information spaces, and identifies patterns or relationships, while humans contextualise those outputs through experience, domain knowledge, and reflective judgement.

Integrative Hybrid Decision extends collaboration into exploratory and uncertain problem spaces. Generative AI can search broad solution spaces, synthesise information across domains and generate alternatives, while humans provide context, direction, interpretation and normative judgement.

These configurations represent more than different applications of technology. They represent different ways of allocating cognitive labour.

Li and Tian describe this in terms of the locus of bounded rationality: the constraints on reasoning may lie predominantly with the human, predominantly within the algorithmic system, or increasingly within a combined Human–AI configuration.

This is significant because AI does not eliminate bounded rationality. It can redistribute it.

The Conservation of Resources (CoR) perspective adds another important dimension. Human attention, time, knowledge, energy, expertise and supportive relationships are finite and valuable resources. CoR theory proposes that people seek to acquire, retain, protect and build valued resources, and that persistent resource loss can produce reinforcing loss cycles, while resource gains can contribute to reinforcing gain cycles (Hobfoll et al., 2018).

The design question is therefore not simply whether AI can perform a task.

It is:

Where should reasoning, interpretation, judgement and accountability reside for this particular kind of work?

From automation to higher-value human work

A more discriminating approach is needed than simply asking what work AI can automate.

• Automate routine execution.
• Accelerate analysis and synthesis where craft knowledge already exists.
  
• Use AI to support learning while humans develop that craft knowledge.
  
• Co-generate where exploration and uncertainty are involved.
  
• Retain human judgement and accountability where consequences matter.

Some work can be automated with little loss of human capability. Filling a form from information already held elsewhere, retrieving a stored password, reconciling known fields or performing repetitive calculations are examples. The task does not depend upon the person developing significant judgement through its repeated performance.

Other work is different.

Consider developing a Business Capability Model. A competent practitioner can research an organisation, compare industry patterns, identify candidate capabilities, structure them into levels and produce a credible first-cut BCM. Doing that manually may take days or weeks.

AI can analyse and synthesise the same breadth of existing knowledge and generate a credible candidate model in minutes.

That is not simply automation. It is machine-scale acceleration of skilled intellectual work.

The important qualification is that the human must possess sufficient craft knowledge to judge the result.

A person learning how to construct a BCM should not simply ask AI to produce one and accept the output. The act of identifying capabilities, distinguishing capability from process or organisation structure, resolving boundaries, constructing levels and testing definitions is itself how the practitioner develops the mental models needed to recognise whether a later AI-generated BCM is good, incomplete or simply wrong.

AI can support that learning — by questioning, explaining, comparing alternatives and providing feedback — but it should not remove the formative practice through which judgement develops.

The parallel with a skilled trade is useful.

An apprentice carpenter first learns to mark out, cut and fit a mortise-and-tenon joint by hand. Through doing so, the apprentice learns grain, fit, tolerance, alignment, tool control and what a good joint actually feels like.

Once that craft knowledge has been internalised, using a mortising machine and a tenoning jig does not diminish the craft. It amplifies it.

The machine provides speed, repeatability and scale.

The craftsperson provides judgement.

AI can play the same role in knowledge work.

Before craft knowledge: AI should assist learning.
After craft knowledge: AI can accelerate execution.
Throughout: the human must remain capable of judging the result.

This has an important implication for the design of work.

If AI is introduced too early and removes the very activities through which expertise is formed, the organisation may gain short-term productivity while undermining its future capacity for informed judgement. It can create people who know how to obtain an answer without knowing how to recognise a good one.

Conversely, where craft knowledge is deliberately developed, and AI is then used as a cognitive power tool, the organisation can achieve both speed and deeper human capability.

The BCM example makes this particularly visible:

• Human learning builds the capability to model.
• AI generates the first cut at machine scale.
• Human craft knowledge evaluates and reshapes it.
• Gemba tests it against reality.
• The Knowledge Base retains what survives that test.

The principle is:

Do not automate away the work through which necessary human judgement is learned.

And once that judgement exists:

Use AI to remove the drudgery from skilled work, not the skill from the worker.

Capability regeneration or capability erosion

This distinction matters because technological augmentation can produce opposing outcomes.

Li and Tian (2026) identify risks associated with excessive reliance on algorithmic outputs, particularly when work depends on heterogeneous, context-sensitive or tacit knowledge. Blind acceptance of AI recommendations can weaken human judgement, while excessive distrust can prevent organisations from benefiting from useful computational capabilities. They therefore emphasise the ongoing importance of human contextual reasoning, tacit knowledge, interpretive agency, and the ability to interpret, challenge and co-create decisions with AI.

This is why internalised craft knowledge is crucial.

A practitioner can only make an informed judgement about an AI-generated output if they have sufficient knowledge of the craft to recognise what good looks like, detect omissions or implausible assumptions, understand relevant trade-offs, and know when the result must be challenged.

Craft knowledge is not merely information that can be retrieved. It is progressively internalised through learning, practice, feedback, error correction and experience.

For a Business Capability Model, for example, this means learning to:

• distinguish capability from process,
• identify uniqueness,
• understand hierarchy and data inheritance,
• distinguish capability from organisational structure or technology domain,
• establish appropriate levels of abstraction,
• determine and set capability boundaries,
• recognise when a model is internally coherent but does not describe the organisation meaningfully.

Once that knowledge has been internalised, AI can dramatically accelerate the work. It can search, compare, synthesise and generate a credible first-cut model in minutes rather than the days or weeks that may be required manually.

Our Human–AI Knowledge Base Development Guideline externalises and codifies much of this craft knowledge as explicit modelling rules, relationships, and checks, while the practitioner retains the judgement required to interpret and apply them.

But if AI performs the task before the practitioner has developed the craft knowledge needed to judge its output, it may remove the very experience through which that judgement would otherwise have developed.

The result can become a reinforcing capability-loss cycle:

The alternative is a regenerative cycle:

The crucial difference here are the guard rails to AI knowledge creation the Purpose, Ethics and Governance provide. The critical choice is not whether an organisation uses AI, but which of two reinforcing cycles its use of AI creates.

That is where the apprentice carpenter analogy becomes conceptually important rather than decorative. The apprentice learns mortise-and-tenon before relying on the mortising machine because manual practice develops the perceptual and practical knowledge required to judge the machine-produced result and skilfully make any required corrections.

For knowledge work, the equivalent principle is:

AI should not remove the formative work through which necessary human judgement is acquired.

Once that judgement has been developed, AI becomes a cognitive power tool rather than a substitute for capability.

This suggests an important principle for the design of Human–AI work:

• AI should not remove the formative work through which necessary human judgement is acquired.
• Once that judgement has been developed, AI can become a cognitive power tool rather than a substitute for capability.

Work is also a learning system

Changing the allocation of cognitive work has consequences beyond individual productivity or decision quality.

Work is a principal place where organisational knowledge is created, challenged, shared and renewed.

Crossan, Lane and White’s (1999) organisational learning framework describes learning as a multilevel process operating across individuals, groups and organisations through intuiting, interpreting, integrating and institutionalising. Learning feeds forward as new ideas and understandings move towards collective action and institutionalisation, while institutionalised knowledge feeds back to influence subsequent individual and group action.

Kleysen and Dyck (2001) elaborate this framework by adding several processes particularly relevant to Human–AI work.

They add attending, linking organisational learning to changes and signals in the external environment.

They add championing and coalition-building, recognising that knowledge does not become organisational knowledge merely because it is technically correct. Ideas must survive social and organisational processes in Gemba before they are accepted, resourced and incorporated into collective action.

Most importantly for the present argument, they add encoding and enacting as feedback processes.

Encoding occurs when organisational learning is incorporated into structures that guide action — policies, procedures, technical designs, roles, routines, and other forms of organisational knowledge.

Enacting occurs when people encounter those structures in actual work.

The distinction is crucial.

What the organisation says happens and what actually happens are not necessarily the same thing.

Encoding meets reality

Kleysen and Dyck (2001) explicitly recognise that group-level practice can differ from organisational encoding because actual work encounters conditions, novelty and demands that could not have been fully anticipated when the organisational model was created.

That creates a feedback requirement.

If emergent knowledge from actual work does not feed forward again, the organisation may continue encoding a version of reality that no longer describes the work.

Organisational learning is therefore not adequately represented by a simple linear progression:

learn → document → implement.

It is iterative:

interpret → integrate → encode → enact → encounter reality → learn → reinterpret → re-encode.

Kleysen and Dyck describe organisational learning as requiring both feedforward and feedback processes operating iteratively.

This provides an important theoretical foundation for the role of Gemba.

Gemba tests everything

Gemba is the place where work actually happens (Imai, 1997).

• It is where organisational encoding encounters enacted reality.
• A process model can be internally coherent and still be wrong.
• A procedure may describe intended practice but omit the workaround that makes the operation function.
• A data model may correctly represent the information visible to management while missing information actually used by practitioners.
• A performance measure may appear reasonable but generate undesirable behaviour.
• An AI-generated capability model may be logically convincing while failing to recognise a critical activity known immediately to experienced people doing the work.

For Human–AI knowledge development, this distinction becomes especially important because AI can generate plausible structures extremely quickly.

Speed of generation is not evidence of truth.

The Human–AI generated model must therefore remain candidate knowledge until it has encountered reality.

• The model goes to Gemba.
• Gemba challenges the model.
• Humans and AI learn.
• The model changes.

This is an application and extension of the organisational-learning logic rather than a proposition made directly by Kleysen and Dyck.

Their encoding–enacting distinction explains why such a loop is necessary.

Gemba provides the practical location in which that loop can occur.

Golden rule: AI may generate at machine scale. Humans can only validate at human scale.

Learning is social

Knowledge also does not move through organisations on its own.

It moves through relationships.

Abdirahman, Sauvée and Shiri (2014), examining learning and innovation through a network perspective, distinguish three interacting forms of network effect:

• structural effects — who is connected to whom and what resources or knowledge can be accessed;
• interactive effects — the nature and quality of the exchanges occurring through those relationships; and
• cognitive effects — the knowledge, memory, norms, culture and sensemaking that accumulate through interaction over time.

Their argument matters because the relationship between networks and learning is recursive.

Networks enable and constrain learning, while learning also reshapes the network through which future learning occurs.

Relationships determine who encounters new ideas, who shares experience, who challenges assumptions and who can mobilise resources. Interaction subsequently alters relationships, shared understanding, trust and patterns of collaboration.

The result can become a reinforcing cycle:

relationships enable knowledge creation
→ learning changes relationships
→ changed relationships alter future knowledge creation.

Trust, reciprocity, interaction and shared understanding therefore become part of the infrastructure of organisational cognition rather than merely desirable cultural characteristics.

This is also consistent with the social-capital perspective used by Abdirahman et al. (2014), in which access to relationships and the resources embedded within them affects opportunities for experimentation, reflection, communication and exposure to new ideas.

From individual learning to organisational memory

The network perspective also helps explain how learning can accumulate.

In the case examined by Abdirahman et al. (2014), shared resources, interpersonal exchange, cross-auditing and collective learning contributed over time to an organisational memory held partly in shared information resources and partly in relationships, routines, norms and a developing community of practice.

Knowledge therefore existed in more than one place.

It existed:

• in individuals;
• between individuals;
• within shared practices;
• within organisational artefacts;
• within formal systems;
• within memory;
• and within the relationships through which those resources could be interpreted and mobilised.

This matters enormously for AI.

A conversation between one person and an AI may produce insight.

But insight is not yet organisational learning.

Unless that learning can be challenged, shared, retained, connected to existing knowledge, enacted in practice and subsequently revised, it remains largely local and transient.

The role of the Human–AI Knowledge Base

This suggests a new organisational role for the Human–AI Knowledge Base.

The proposition advanced here is that a shared Human–AI Knowledge Base can provide part of the organisational mechanism through which Human–AI interaction becomes cumulative organisational learning.

AI can assist with:

• searching large bodies of information;
• identifying patterns and relationships;
• comparing alternative interpretations;
• generating candidate models;
• maintaining semantic consistency;
• locating gaps and contradictions;
• connecting knowledge across domains; and
• rapidly regenerating models as understanding changes.

Humans contribute what cannot responsibly be delegated simply to computational generation:

• purpose;
• contextual understanding;
• lived experience;
• tacit knowledge;
• interpretation;
• ethical judgement;
• consequence awareness;
• challenge;
• and accountability.

The Knowledge Base then provides a place in which candidate knowledge can be progressively structured, connected, challenged, validated and retained.

This is more than document storage.

Its value lies in supporting cumulative knowledge.

New learning can modify what is already known rather than simply generating another disconnected artefact.

From conversation to cumulative organisational learning

Combining the three bodies of research gives a possible organisational learning cycle:

Human–AI exploration → candidate knowledge is generated

Interpretation and integration → people contextualise, challenge and relate it to existing knowledge

Social validation → relevant practitioners and stakeholders contribute experience and alternative perspectives

Encoding → validated knowledge becomes part of the shared organisational Knowledge Base

Enactment at Gemba → encoded knowledge is encountered in actual work and consequences identified

Consequence and feedback → mismatch, novelty, tacit knowledge and unintended effects become visible

Reinterpretation and learning → humans and AI reconsider the model

Re-encoding → the Knowledge Base changes

Changed shared mental models and relationships
→ future interpretation and action occur from a different starting point.

The key feature is that the cycle does not terminate when a model is generated or even when it is documented.

Knowledge remains provisional and revisable because experience continues.

The Knowledge Base, Human Capital and Social Capital

This also suggests why technology alone cannot create organisational cognition.

A functioning organisational learning system requires at least three interacting forms of capability:

• Human Capital provides knowledge, skill, experience, judgement and the capacity to learn.
• Social Capital provides the relationships, trust, reciprocity and channels through which knowledge can be shared, challenged and combined.
• The Knowledge Base provides cumulative organisational memory: the explicit structures through which validated knowledge can be retained, connected, made accessible and revised.

AI can greatly strengthen the third component and can augment the first two.

It cannot substitute for all three.

Indeed, degrading human judgement or social relationships while improving the technical Knowledge Base could reduce rather than improve overall organisational cognition.

The relevant system is therefore not simply: Human + AI

but: Human Capital ↔ Social Capital ↔ Shared Knowledge Base ↔ AI-enabled learning

operating continuously through actual work.

Governance determines the boundary

Li and Tian (2026) also make clear that different decision contexts require different distributions of authority.

Structured, stable and data-intensive environments may permit substantial algorithmic decision-making.

Ambiguous, ethically sensitive, consequential or strategically uncertain situations require stronger human interpretation and control.

As AI becomes more capable, governance becomes more important rather than less.

Organisations need to decide:

• where AI may act autonomously;
• where AI may recommend but not decide;
• where human judgement must remain central;
• how AI-generated propositions can be challenged;
• how decisions can be contested;
• who is accountable for consequences;
• how bias and error are detected;
• what knowledge is permitted to become organisationally authoritative; and
• when a model must return to Gemba for revalidation.

Li and Tian emphasise transparency, contestability and accountability and warn against inappropriate drift toward automation.

The implication is that Purpose, Ethics and Governance form boundary conditions for Human–AI cognition.

• Capability answers what can be done.
• Governance answers what may be done.
• Ethics asks what should be done.
• Purpose asks why it is being done at all.

Human accountability remains essential where consequences extend beyond the computational problem being solved.

From cognitive erosion to organisational cognition

The literature reviewed here supports the components of a larger argument, although the integrated model proposed below remains a synthesis requiring further empirical testing.

Used poorly:

AI substitutes for judgement
→ people become detached from interpretation
→ tacit capability weakens
→ encoded knowledge drifts away from enacted reality
→ feedback diminishes
→ shared organisational understanding deteriorates
→ cognitive erosion becomes possible.

Used differently:

AI reduces unnecessary information-processing load
→ people retain and develop interpretation and judgement
→ Human–AI collaboration expands analytical and exploratory capacity
→ social interaction challenges and integrates candidate knowledge
→ Gemba introduces tacit, contextual and consequential knowledge
→ validated learning enters the shared Knowledge Base
→ shared mental models change
→ action generates new experience
→ feedback renews organisational knowledge
→ human, social and organisational capability can regenerate.

This is not an argument that AI inevitably creates better organisations.

It argues that AI creates a new design space for organisational cognition.

Whether that space produces erosion or regeneration depends upon the way work, learning, relationships, knowledge and governance are designed around it.

The changing nature of work

The deeper transformation may therefore not be that machines increasingly perform work once undertaken by people.

It may be that the division of cognitive labour within organisations changes.

AI can increasingly perform machine-scale search, comparison, calculation, synthesis and generation.

Humans become increasingly valuable where work requires meaning, context, judgement, relationships, ethics, consequence, learning and accountability.

And Human–AI collaboration becomes most powerful when the knowledge produced by that collaboration does not disappear at the end of an interaction, but enters a living organisational learning system.

The future of work is therefore not simply a contest between humans and machines.

It is an organisational design challenge:

How do we combine machine-scale intelligence with human-scale judgement, social learning and Gemba reality to create better work, better knowledge and better organisations?

References

Abdirahman, Z.-Z., Sauvée, L. & Shiri, G. (2014). Analyzing network effects of Corporate Social Responsibility implementation in food small and medium enterprises. Journal on Chain and Network Science, 14(2), 103–115. https://doi.org/10.3920/JCNS2014.x005

Crossan, M. M., Lane, H. W. & White, R. E. (1999). An organizational learning framework: From intuition to institution. Academy of Management Review, 24(3), 522–537.

Imai, M. (1997). Gemba Kaizen: A Commonsense, Low-Cost Approach to Management. McGraw-Hill.

Kleysen, R. F. & Dyck, B. (2001). Cumulating knowledge: An elaboration and extension of Crossan, Lane & White’s framework for organizational learning. In M. Crossan & F. Olivera (Eds.), Organizational Learning and Knowledge Management: New Directions (pp. 383–394). Richard Ivey School of Business.

Li, H. & Tian, F. (2026). Advancing decision-making through AI-human collaboration: A systematic review and conceptual framework. Group Decision and Negotiation, 35, Article 26. https://doi.org/10.1007/s10726-026-09980-1