The Centaur Organization: Designing Human–AI Collaboration for Decision Quality in Knowledge Work

Jonathan H. Westover, PhD
Jul 28
19 min read

Listen to a review of this article:

Abstract: Organizations are investing aggressively in artificial intelligence, yet many of these initiatives are framed narrowly around automation and headcount reduction. Drawing on the human–AI symbiosis perspective advanced by Jarrahi (2018) and a broader stream of management and information systems research, this article argues that durable value from AI emerges when organizations design for complementarity rather than substitution. The article maps the comparative strengths of humans and AI against three classic challenges in organizational decision making — uncertainty, complexity, and equivocality — and translates that map into a practitioner playbook. It examines five evidence-based organizational responses (task allocation, capability building, explainability, sociotechnical integration, and algorithmic governance) and illustrates them with narratives from healthcare, financial services, consumer goods, and industrial manufacturing. The article closes with three forward-looking pillars for sustaining a "centaur organization": continuous co-learning, distributed decision authority, and stewardship of human judgment. The intended audience is senior leaders, transformation officers, and HCI practitioners shaping AI-augmented work.

For more than a decade, executives have been told some version of the same story: smart machines are coming for knowledge work, and the firms that automate fastest will win. The story is partly true and almost entirely incomplete. Investment is real — surveys of large enterprises consistently rank AI and machine learning among the most disruptive forces in the business landscape (New Vantage Partners, 2017) — but the lived experience of AI deployment inside organizations is messier than the headlines suggest. Pilots stall. Models perform brilliantly in the lab and inconsistently in the field. Workers either ignore the recommendations or, equally problematically, defer to them uncritically. Boards ask why the productivity numbers have not yet moved.

A different framing has been quietly gaining ground in both scholarship and practice. Jarrahi (2018) argued that humans and AI bring distinct cognitive strengths to organizational decision making, and that the most valuable applications are those that exploit their complementarity rather than treating one as a substitute for the other. The vision is not new — Licklider sketched a version of "human–machine symbiosis" in the early 1960s, and Kasparov's "centaur chess" demonstrated empirically that a competent human paired with a capable machine could outperform either alone (Campbell, 2016). What is new, and what makes the topic urgent for managers today, is that the practical scope of this partnership has expanded from chessboards and pathology slides to underwriting, recruiting, R&D portfolio choice, customer service triage, and strategic planning.

The practical stakes are high. Studies of earlier waves of technology-driven change suggest that programs framed primarily as labor substitution often deliver short-term cost savings followed by demoralized workforces, eroded tacit knowledge, and stalled adoption (Mumford, 1994). The same risks apply to AI, with one important addition: contemporary AI systems are themselves dependent on human input — for training data, for edge-case judgment, and for the contextual sense-making that determines whether a model output is actionable or absurd. Designing for that dependency, rather than wishing it away, is the central management challenge of the next decade.

This article synthesizes the human–AI symbiosis literature into a practitioner-oriented playbook. It defines key terms, surveys the state of practice, examines organizational and workforce consequences, and lays out five evidence-based responses with concrete narratives drawn from across industries. It closes with three forward-looking pillars for building what we call the centaur organization — one whose decision systems are deliberately engineered for human–AI collaboration.

The Human–AI Collaboration Landscape

Defining Human–AI Symbiosis and Augmentation

The terminology in this space has multiplied faster than the underlying concepts. A few distinctions are worth pinning down before going further.

Artificial intelligence (AI): a heterogeneous family of computational techniques — including machine learning, natural language processing, computer vision, and expert systems — that enable systems to perform tasks ordinarily associated with human cognition (Russell, Norvig, & Intelligence, 1995).
Automation: the substitution of machine performance for human performance of a defined task.
Augmentation: the use of machine capabilities to extend, rather than replace, human performance, typically by handling computationally intensive subtasks while humans retain interpretive, relational, and judgment-laden ones (Davenport & Kirby, 2016; Jarrahi, 2018).
Human–AI symbiosis: a relationship in which the strengths of human and machine compensate for the limitations of the other, often resulting in joint performance superior to either alone (Jarrahi, 2018).

The empirical case for symbiosis is sturdiest where it has been measured directly. In a now-frequently-cited cancer-detection study, an AI-only model produced a 7.5% error rate on lymph node images and human pathologists produced a 3.5% error rate, while the human–AI combination produced a 0.5% error rate — roughly an 85% reduction relative to the pathologist alone (Wang, Khosla, Gargeya, Irshad, & Beck, 2016). Comparable patterns have been documented in chess (Campbell, 2016) and in laboratory studies of small-group coordination, where the introduction of "noisy" autonomous agents shortened the median time human teams needed to solve a coordination problem by roughly 55% (Shirado & Christakis, 2017).

These results do not show that AI is universally helpful, nor that pairing a human with any model in any workflow will improve outcomes. They show that carefully designed pairings, in domains where each party brings a distinct strength, can outperform either alone. Designing those pairings is an organizational capability, not a technical one.

State of Practice

The current state of organizational adoption is uneven. Executive surveys point to broad enthusiasm: a 2017 Big Data Executive Survey found that the vast majority of large-firm leaders viewed AI and machine learning as among the most disruptive forces facing their businesses (New Vantage Partners, 2017), and consulting research from the same period reported that a large share of executives planned substantial AI investment over the following several years (Accenture, 2017). Yet the dominant business case used to justify those investments has often centered on near-term cost reduction, particularly headcount displacement (Davenport & Faccioli, 2017).

That framing creates two predictable problems. First, it tends to underweight the long lead times required for AI to deliver value in complex sociotechnical environments — environments in which technological breakthroughs prevail only when they are judiciously integrated into the social fabric of an organization (Sawyer & Jarrahi, 2014). Second, it pushes organizations toward applications where AI's analytical strengths are clearest (high-volume, well-defined, data-rich tasks) while neglecting the larger and arguably more valuable territory where AI augments human decision making in ambiguous, equivocal, or strategic contexts.

Beneath the headlines, a more interesting picture is emerging at firms that have lived with these technologies for several years. Procter & Gamble and American Express, for example, have reportedly framed AI not principally as a substitute for human work but as a tool that employees draw on to do their work more effectively (Davenport & Bean, 2017). General Electric, in the course of its widely covered transition toward a "digital industrial" identity, has cultivated cadres of "hybrid" experts — domain specialists trained additionally in machine learning — to serve as bridges between AI capability and operational decision making (CIO Network, 2017). Patterns like these point to a state of practice that is bifurcating: a leading edge that treats AI as a complement to human capability, and a trailing majority still framing it primarily as substitution.

Organizational and Workforce Consequences of AI Integration

Organizational Performance Impacts

Where they are well-designed, human–AI partnerships have delivered measurable performance gains. The pathology results noted above (Wang et al., 2016) translate directly into clinical decision quality. In venture capital, Correlation Ventures became known for compressing investment evaluation cycles to roughly two weeks by combining predictive analytics over large historical datasets with holistic human review of the shortlisted opportunities. In content moderation, large platforms now rely on bots to triage terabytes of user-generated material, with on-demand human workers making final calls on the contested edge cases — what Gray and Suri (2017) memorably called the workforce "behind the AI curtain."

It is tempting to read these numbers as a clean story of AI value creation. They are better read as evidence that complementary architectures outperform either pure-automation or pure-human alternatives. The 85% reduction in pathology error did not come from removing the pathologist; it came from putting the pathologist in a workflow where the AI handled what AI does best, leaving the human in the loop for what humans do best (Jarrahi, 2018).

There is also a less encouraging side to the performance ledger. Lessons from earlier technology-led transformations — most notably business process reengineering — caution that short-term financial gains from labor substitution can be ephemeral, eroded over time by demoralization, loss of tacit organizational knowledge, and weakened informal coordination networks (Mumford, 1994). The contemporary equivalent risk is what some critics describe as a modern-day Taylorism in which algorithmic management aspires to deskill workers or remove them altogether for the sake of efficiency (Frischmann & Selinger, 2017). The resulting performance can look attractive on a quarterly P&L and corrosive over a five-year horizon.

Knowledge Worker and Stakeholder Impacts

For workers, the effects of AI integration are more nuanced than either the utopian or dystopian framings allow. On the positive side, AI tools can offload repetitive analytical drudgery, surface patterns that would otherwise remain invisible, and free knowledge workers to spend more time on the parts of their jobs that draw on intuition, judgment, and relational skill — the parts most workers find meaningful. AI can also support skill development, particularly in analytical reasoning, when its recommendations are accompanied by intelligible explanations of how those recommendations were derived (Davenport & Kirby, 2016).

On the less encouraging side, AI systems deployed without sufficient attention to their human context can erode trust, deskill experienced workers, and obscure accountability when decisions go wrong. There is also a documented concern that organizational members may not "follow" an AI system in the same way they would follow a credible human leader (Parry, Cohen, & Bhattacharya, 2016) — meaning that even technically sound AI recommendations can fail to drive action if they lack the social legitimacy of a human champion.

For external stakeholders — patients, customers, citizens — the impacts depend heavily on whether the organization has built genuine human–AI collaboration or hidden the AI behind a thin layer of human window-dressing. In high-stakes decisions (a credit denial, a clinical diagnosis, a hiring outcome), stakeholders increasingly expect both algorithmic rigor and human accountability, and they are noticing when one is performed without the other.

Evidence-Based Organizational Responses

Table 1: Human-AI Collaboration and Symbiosis Case Studies

Organization or Study	Domain	AI Application/Task	Human Role	Decision Characteristics	Performance Outcome	Key Strategy Applied
Wang et al. (2016)	Healthcare	Identifying metastatic breast cancer in lymph node images	Final diagnostic interpretation and clinical judgment	Complexity	0.5% error rate (85% reduction relative to pathologist alone)	Task allocation (complementary architectures)
Memorial Sloan Kettering	Healthcare	Discerning cancer patterns and suggesting treatment courses (IBM Watson)	Final clinical judgment; weighing patient values and risk tolerance	Complexity and Equivocality	Improved reasoning across vast literature and patient histories	Task allocation (decomposition)
Kasparov (Centaur Chess)	Gaming / Strategy	Chess move calculation and strategy	Guidance and strategic selection of machine moves	Complexity	Superior performance compared to pure machine or pure human	Human-AI symbiosis (complementary strengths)
Shirado & Christakis (2017)	Social Science	Autonomous agents assisting in coordination problems	Global coordination and problem solving	Complexity and Uncertainty	55% reduction in median time to solve coordination problems	Continuous co-learning; human-AI symbiosis
American Express	Financial Services	Fraud detection and customer risk decisions	Analysts overriding models based on contextual cues	Complexity and Equivocality	Enhanced trust and regulatory compliance	Explainability; transparent AI systems
Correlation Ventures	Financial Services	Investment evaluation and predictive analytics	Holistic human review of shortlisted opportunities	Complexity and Uncertainty	Evaluation cycles compressed to roughly two weeks	Task allocation (complementary architectures)
General Electric	Manufacturing	Operational decision making and digital industrial transition	Hybrid experts bridging AI capability and operational context	Complexity and Uncertainty	Durable performance through contextual judgment	Capability building; hybrid talent models
Large Social Media Platforms	Content Moderation	Triaging user-generated material and screening policy violations	Final adjudication on contested edge cases	Equivocality	High-volume screening with human accountability	Algorithmic governance; human-in-the-loop triage
Procter & Gamble	Consumer Goods	Augmenting existing workflows and employee effectiveness	Employees use AI as a tool within established roles	Complexity	Durable integration into the social fabric of the organization	Sociotechnical integration; change management

The literature converges on five categories of organizational response. None is a silver bullet; together they form the operating system of a centaur organization.

Task Allocation by Decision Characteristic

The single most important design choice in any AI-augmented workflow is the allocation of subtasks between human and machine. Jarrahi (2018), building on Choo (1991) and Simon (1982), argues that this allocation should be driven by the underlying character of the decision, which can usefully be decomposed into three challenges:

Uncertainty (missing information): humans tend to outperform when there is no precedent and intuition trained on tacit experience must fill the gap, while AI contributes by surfacing real-time information and anomaly signals (Sadler-Smith & Shefy, 2004).
Complexity (too many variables): AI tends to outperform on retrieval, computation, and pattern detection at scale, with humans choosing among options of roughly equal data support (Marwala, 2015).
Equivocality (divergent interpretations): humans retain strong advantages in negotiation, coalition-building, and persuasion, while AI can contribute through sentiment analysis and stakeholder mapping (Weick & Roberts, 1993).

The practical move is not to dichotomize but to decompose. Most organizational decisions blend all three challenges in different proportions (Koufteros, Vonderembse, & Jayaram, 2005), so the allocation should happen at the subtask level, not the decision level.

Effective approaches include:

Decision audits that score recurring decisions on uncertainty, complexity, and equivocality before assigning AI a role
Workflow choreography that specifies clear handoffs between machine and human at each stage
Reversibility tiers that route low-stakes, reversible decisions toward higher AI autonomy and high-stakes, irreversible ones toward stronger human oversight
Confidence-based routing, in which AI confidence scores trigger human review at defined thresholds

Memorial Sloan Kettering's work with IBM Watson illustrates the principle. Through machine learning techniques and access to medical research articles, electronic medical records, and clinicians' notes, the system was developed to discern cancer patterns and to suggest courses of treatment, while final clinical judgment remained with the oncologist (Captain, 2017). The complexity subtask — reasoning across vast literatures and patient histories — was allocated to AI; the equivocality subtask of weighing the patient's values, family situation, and risk tolerance remained with the human clinician.

Capability Building and Hybrid Talent Models

If task allocation is the architecture, capability is the foundation. Workers need fluency in how AI systems generate recommendations, and AI systems need to be developed and operated by people who understand the domain context they are meant to serve. Both kinds of fluency are scarce, and neither is built quickly.

The most promising model emerging in practice is what some firms call dual experts or hybrid scientists: domain professionals who acquire applied machine-learning skills through structured certification rather than data scientists who try to learn the domain from the outside (CIO Network, 2017). The advantage of this directionality is that domain expertise — particularly the tacit, hard-to-articulate kind described by Sadler-Smith and Shefy (2004) — is harder to teach than applied analytics.

Effective approaches include:

Certification programs that route subject-matter experts (engineers, clinicians, analysts) into structured AI training
Embedded data scientists paired one-to-one with frontline decision makers
Reverse mentoring in which junior data-fluent staff coach senior leaders on model interpretation
Curricular updates in MBA and professional-school programs that maintain analytical rigor alongside intuition and judgment (Martin, 2009)

General Electric has been an instructive case. As part of its broader digital transition, GE invested heavily in upskilling existing subject experts — physicists, aerospace engineers, business analysts — through internal data analytics certification programs, on the premise that these "hybrid scientists" were better positioned to develop workable AI integrations than data scientists imported into unfamiliar industrial domains (CIO Network, 2017). The strategy reflected a recognition that the bottleneck in industrial AI is rarely the algorithm; it is the contextual judgment about which problem to solve, which data to trust, and which recommendation to act on.

Transparent and Explainable AI Systems

A persistent finding across the human–AI literature is that trust mediates use. Workers who do not understand how a model reached its recommendation are likely either to ignore it (reverting to prior practice) or to defer to it uncritically (suspending their own judgment). Both failure modes erode the value of the system. Davenport and Kirby (2016) argued that knowing how smart machines arrive at their analytical recommendations is a key element in helping humans interact effectively with them; transparency is not a nice-to-have but a precondition for symbiosis.

Explainability also has a developmental dimension. When AI exposes its reasoning — the features it weighted most heavily, the historical analogues it drew on — humans not only trust the recommendation more but build their own analytical fluency over time (Hung, 2003). The model becomes a teacher as well as a tool.

Effective approaches include:

Local explanation interfaces that show, for each recommendation, the top contributing features
Counterfactual displays ("if X had been different, the recommendation would have been Y")
Confidence intervals and uncertainty quantification presented alongside point recommendations
Audit logs that allow retrospective review of model inputs and outputs in disputed cases

American Express has been cited as an example of a firm that frames AI as a tool employees draw on rather than a black-box arbiter (Davenport & Bean, 2017). In fraud detection and customer risk decisions, the value of explainable signals is twofold: they allow analysts to override the model when contextual cues warrant, and they support regulatory and customer-facing explanations of why a particular decision was made.

Sociotechnical Integration and Change Management

Decades of research on information systems make it clear that organizations are sociotechnical systems in which technical artifacts and social practices co-evolve, and that interventions that change one without attending to the other tend to fail (Sawyer & Jarrahi, 2014; Mumford, 1994). AI is not exempt from this principle; if anything, it intensifies it, because AI systems learn from organizational data and behavior and therefore embed themselves more deeply into work practices than earlier generations of enterprise software.

The practical implication is that AI integration is a change-management problem at least as much as a technology problem. The questions that determine whether a deployment succeeds are sociotechnical: Whose work changes? Whose authority changes? Whose tacit knowledge is being captured, by whom, and to what end? Who is accountable when the system errs?

Effective approaches include:

Co-design workshops in which frontline workers help shape the system's interfaces and decision thresholds
Phased rollouts with explicit learning checkpoints between phases
Job redesign that reallocates time freed by AI toward higher-judgment work rather than simply increasing throughput
Informal-leader engagement, recognizing that organizational influence often runs through individuals other than formal managers (Cross, Borgatti, & Parker, 2002)

Procter & Gamble offers a useful counterpoint to deployments that treat AI as a substitution play. Rather than framing AI primarily as a means of process automation or job elimination, the firm has reportedly approached AI as a tool that employees use in their work, anchoring deployment in existing roles and workflows (Davenport & Bean, 2017). The result has been less dramatic in the short term than a substitution narrative would predict, and arguably more durable.

Ethical Guardrails and Algorithmic Governance

The fifth response is governance. AI systems make consequential decisions about people — patients, applicants, employees, customers — and they do so at a scale and speed that outstrips traditional oversight mechanisms. Organizations that defer governance until after deployment tend to find that the most damaging failures (biased outcomes, opaque denials, regulatory exposure) emerge in precisely the use cases where governance was weakest.

The literature here is younger and more contested than in the other four areas, but several principles have stabilized. First, governance should be domain-specific: the appropriate oversight regime for a recommendation engine is not the same as for a medical triage system. Second, governance benefits from multi-stakeholder input, including the people most affected by the decisions. Third, governance is continuous, not episodic, because models drift as the data they ingest and the environments they operate in change.

Effective approaches include:

Algorithmic impact assessments conducted before deployment and refreshed at defined intervals
Tiered review boards with authority calibrated to the stakes of the decision
Monitoring for drift and disparate impact on protected or vulnerable populations
Clear escalation paths for workers and stakeholders to flag suspect outputs

A useful illustration comes from large social media platforms, where bots screen enormous volumes of user-generated content for potential policy violations, but final decisions on contested removals typically rest with human reviewers (Gray & Suri, 2017). This architecture — algorithmic detection plus human adjudication — is not perfect, but it represents an explicit governance choice to keep accountability with humans for the decisions that most affect users.

Building Long-Term Human–AI Collaborative Capability

The five responses above describe what to do now. The harder question is how to build the underlying capability for the long term, as AI systems and the work they support continue to evolve. Three pillars are emerging.

Continuous Co-Learning Systems

Human–AI symbiosis is not a steady state; it is a learning loop. Most AI systems improve with exposure to additional data and human feedback, and human decision makers can correspondingly develop richer mental models of how the systems behave (Shirado & Christakis, 2017). The organizational task is to make this co-learning deliberate rather than incidental.

Concretely, that means building feedback infrastructure that captures not only model outputs but also human overrides, the reasoning behind those overrides, and the eventual outcomes. Over time, this corpus becomes a strategic asset: it improves the model, surfaces patterns of where human and machine judgment diverge, and reveals which decisions are most amenable to further automation and which require continued human investment. It also requires investment in workers' AI literacy on a continuing basis, not as a one-time training event but as an ongoing professional development obligation analogous to continuing education in regulated professions.

Distributed Decision Authority

A second pillar is the recognition that decision authority in AI-augmented organizations does not collapse neatly onto formal hierarchies. As Cross, Borgatti, and Parker (2002) showed, much of the work of organizational decision making — particularly the equivocality work of building consensus and aligning interests — flows through informal networks that often do not map onto the org chart. AI does not eliminate these networks; it interacts with them, sometimes in ways that strengthen central nodes and sometimes in ways that bypass them.

The implication for design is that AI should be deployed in ways that respect and reinforce the distributed nature of real organizational decision making. That means giving frontline workers the tools and authority to act on AI recommendations within defined limits, rather than routing every decision up to a managerial chokepoint where the AI's speed advantage is lost. It also means recognizing that intuitive, equivocality-laden decisions occur throughout the organization, not only at the top — product designers, training specialists, market analysts, and frontline service workers all engage in judgment-heavy decision making (Jarrahi, 2018) — and that AI integration should be calibrated to those varied contexts.

Stewardship of Intuition and Judgment

The third and perhaps most counterintuitive pillar is the active stewardship of human intuition and judgment as organizational capital. Intuition, in the sense developed by Sadler-Smith and Shefy (2004) and Dane, Rockmann, and Pratt (2012), is not magic; it is pattern recognition trained on years of experience, much of which is tacit and difficult to articulate. It is also fragile: when organizations route around experienced judgment in favor of algorithmic recommendations, they can erode the very experiential base on which good intuition depends. Mintzberg's (1994) classic argument that strategic thinking is grounded in synthesis, creativity, and intuition rather than in too-precisely articulated planning becomes especially relevant in an AI-saturated environment, where the temptation to substitute computation for judgment is constant.

Stewardship in this context means several things. It means deliberately preserving roles and career pathways in which workers can develop deep domain expertise rather than only learning to operate the machine. It means giving experienced workers the time and authority to override algorithmic recommendations when their judgment so dictates, and treating those overrides as data rather than insubordination. And it means recognizing that intuition tends to be a nontransferable human attribute (Buchanan & O'Connell, 2006) — meaning that once it is lost, it is not easily rebuilt, and certainly not by a model.

The broader point is that the value of human judgment in an AI-augmented organization is not residual — what is left over after the machines have done their work — but constitutive. The decisions that matter most to organizational performance, those involving genuine uncertainty and equivocality, will continue to depend on it (Burke & Miller, 1999; Hayashi, 2001; Gardner & Martinko, 1996). Treating it as a strategic capability worth protecting is one of the harder and more important shifts a leadership team can make.

Conclusion

The most consequential question facing leaders deploying AI is not "what can we automate?" but "how should we redesign the relationship between humans and machines so that decision quality improves over time?" The evidence accumulated from chess to pathology to coordination experiments points consistently in the same direction: well-designed human–AI partnerships outperform either pure-human or pure-AI alternatives, and the gains are largest in domains characterized by the messy combination of uncertainty, complexity, and equivocality that defines most real organizational decisions (Jarrahi, 2018; Wang et al., 2016; Shirado & Christakis, 2017).

Three takeaways are worth carrying forward. First, allocate subtasks rather than whole jobs: the unit of analysis for AI design is the decision component, not the role. Second, invest in capability — particularly hybrid talent that bridges domain and analytical fluency — on the same timescale as the technology investment, not as an afterthought. Third, treat governance, transparency, and stewardship of human judgment as strategic capabilities rather than compliance overhead; they are what determine whether the partnership remains productive as the technology evolves.

Kelly's (2012) framing remains useful: this is a race with the machines, not against them. The organizations that internalize that distinction — and build the operating systems, talent models, and governance structures to live it — will be the ones that turn AI's potential into durable performance. The rest will discover, slowly and at considerable expense, that automation alone is not a strategy.

Research Infographic

The Centaur Organization Slide Deck

References

Accenture. (2017). Technology vision 2017: Technology for people. Accenture.
Brynjolfsson, E., & McAfee, A. (2012). Winning the race with ever-smarter machines. MIT Sloan Management Review, 53(2), 53–60.
Brynjolfsson, E., & McAfee, A. (2014). The second machine age: Work, progress, and prosperity in a time of brilliant technologies. W. W. Norton & Company.
Buchanan, L., & O'Connell, A. (2006). A brief history of decision making. Harvard Business Review, 84(1), 32–41.
Burke, L. A., & Miller, M. K. (1999). Taking the mystery out of intuitive decision making. Academy of Management Executive, 13(4), 91–99.
Campbell, M. (2016, October 24). 20 years later, humans still no match for computers on the chessboard. NPR.
Captain, S. (2017). Can IBM's Watson do it all? Fast Company.
Choo, C. W. (1991). Towards an information model of organizations. Canadian Journal of Information Science, 16(3), 32–62.
CIO Network. (2017, June 7). How AI and machine learning are helping drive the GE digital transformation. Forbes.
Cross, R., Borgatti, S. P., & Parker, A. (2002). Making invisible work visible: Using social network analysis to support strategic collaboration. California Management Review, 44(2), 25–46.
Dane, E., Rockmann, K. W., & Pratt, M. G. (2012). When should I trust my gut? Linking domain expertise to intuitive decision-making effectiveness. Organizational Behavior and Human Decision Processes, 119(2), 187–194.
Davenport, T. H., & Bean, R. (2017). How P&G and American Express are approaching AI. Harvard Business Review.
Davenport, T. H., & Faccioli, G. (2017). Cognitive on the continent. Data Informed.
Davenport, T. H., & Kirby, J. (2016). Just how smart are smart machines? MIT Sloan Management Review, 57(3), 21–25.
Frischmann, B., & Selinger, E. (2017, September 25). Robots have already taken over our work, but they're made of flesh and bone. The Guardian.
Gardner, W. L., & Martinko, M. J. (1996). Using the Myers-Briggs Type Indicator to study managers: A literature review and research agenda. Journal of Management, 22(1), 45–83.
Gray, M., & Suri, S. (2017). The humans working behind the AI curtain. Harvard Business Review.
Hayashi, A. M. (2001). When to trust your gut. Harvard Business Review, 79(2), 59–65.
Hung, S.-Y. (2003). Expert versus novice use of the executive support systems: An empirical study. Information & Management, 40(3), 177–189.
Jarrahi, M. H. (2018). Artificial intelligence and the future of work: Human–AI symbiosis in organizational decision making. Business Horizons, 61(4), 577–586.
Kelly, K. (2012). Better than human: Why robots will—and must—take our jobs. Wired.
Koufteros, X., Vonderembse, M., & Jayaram, J. (2005). Internal and external integration for product development: The contingency effects of uncertainty, equivocality, and platform strategy. Decision Sciences, 36(1), 97–133.
Martin, R. L. (2009). The design of business: Why design thinking is the next competitive advantage. Harvard Business Press.
Marwala, T. (2015). Causality, correlation and artificial intelligence for rational decision making. World Scientific.
Mintzberg, H. (1994). The fall and rise of strategic planning. Harvard Business Review, 72(1), 107–114.
Mumford, E. (1994). New treatments or old remedies: Is business process reengineering really socio-technical design? Journal of Strategic Information Systems, 3(4), 313–326.
New Vantage Partners. (2017). Big data executive survey 2017. New Vantage Partners.
Parry, K., Cohen, M., & Bhattacharya, S. (2016). Rise of the machines: A critical consideration of automated leadership decision making in organizations. Group & Organization Management, 41(5), 571–594.
Russell, S., Norvig, P., & Intelligence, A. (1995). Artificial intelligence: A modern approach. Prentice-Hall.
Sadler-Smith, E., & Shefy, E. (2004). The intuitive executive: Understanding and applying 'gut feel' in decision-making. Academy of Management Executive, 18(4), 76–91.
Sawyer, S., & Jarrahi, M. H. (2014). Sociotechnical approaches to the study of information systems. In A. Tucker & H. Topi (Eds.), Computing handbook, third edition: Information systems and information technology (pp. 5.1–5.27). Chapman and Hall/CRC.
Shirado, H., & Christakis, N. A. (2017). Locally noisy autonomous agents improve global human coordination in network experiments. Nature, 545(7654), 370–374.
Simon, H. A. (1982). Theories of bounded rationality. In H. A. Simon (Ed.), Models of bounded rationality: Behavioral economics and business organization (Vol. 1, pp. 408–423). MIT Press.
Wang, D., Khosla, A., Gargeya, R., Irshad, H., & Beck, A. H. (2016). Deep learning for identifying metastatic breast cancer. arXiv preprint arXiv:1606.05718.
Weick, K. E., & Roberts, K. H. (1993). Collective mind in organizations: Heedful interrelating on flight decks. Administrative Science Quarterly, 38(3), 357–381.
Zuboff, S. (1988). In the age of the smart machine: The future of work and power. Basic Books.

Jonathan H. Westover, PhD is Chief Research Officer (Nexus Institute for Work and AI); Associate Dean and Director of HR Academic Programs (WGU); Professor, Organizational Leadership (UVU); OD/HR/Leadership Consultant (Human Capital Innovations). Read Jonathan Westover's executive profile here.

Suggested Citation: Westover, J. H. (2026). The Centaur Organization: Designing Human–AI Collaboration for Decision Quality in Knowledge Work. Human Capital Leadership Review, 36(4). doi.org/10.70175/hclreview.2020.36.4.4