Designing for Hybrid Intelligence: How Organizations Can Harness the Complementary Strengths of Human and Artificial Intelligence
Download the article here:
Listen to a review of this article:
Abstract: The accelerating deployment of artificial intelligence across industries has intensified debate about the optimal division of labor between humans and machines. This article examines the concept of hybrid intelligence—socio-technical systems that combine human and artificial intelligence to achieve outcomes superior to what either could accomplish alone, while both continue to learn from each other. Drawing on the taxonomy of design knowledge proposed by Dellermann, Calma, Lipusch, Weber, Weigel, and Ebel (2019), and integrating recent evidence from a large meta-analysis, field experiments, and documented organizational examples, the article translates scholarly research into actionable guidance for organizational leaders. Five evidence-based intervention strategies are presented—task architecture design, interactive learning integration, human-centered interpretability, incentive-aligned participation, and adaptive workflow orchestration—each illustrated with examples from retail, healthcare, scientific research, and artificial intelligence research. The article concludes by outlining three forward-looking pillars for building sustained hybrid intelligence capability: continuous co-learning systems, trust calibration infrastructure, and organizational design for human-AI complementarity.
We are living through a peculiar paradox. Artificial intelligence can now draft legal briefs, flag suspicious findings in medical images, and generate production-quality code—yet many organizations struggle to translate experimentation into enterprise-level value. The gap between AI's technical promise and its organizational payoff has become one of the defining strategic puzzles of our time.
Part of the explanation is not that the technology fails, but that organizations too often treat AI as a replacement for human judgment rather than as a complement to it. The most productive frontier does not lie in full automation, nor in keeping humans firmly at the center of every process. It lies in the deliberate, well-designed integration of both—what researchers have termed hybrid intelligence.
Dellermann et al. (2019) offered one of the earliest and most comprehensive conceptualizations of this idea. They define hybrid intelligence systems as systems able to accomplish complex goals by combining human and artificial intelligence so that, together, they achieve better results than either could alone, while continuously improving by learning from each other. Their taxonomy of design knowledge—spanning task characteristics, learning paradigms, human-AI interaction, and AI-human interaction—provides a structured vocabulary that remains relevant as organizations move from AI experimentation to scaled deployment.
This article builds on that foundation. It translates the Dellermann et al. (2019) taxonomy and subsequent empirical research into a practitioner-oriented guide for designing, implementing, and sustaining hybrid intelligence systems. The stakes are considerable: the McKinsey Global Institute estimates that about $2.9 trillion of economic value could be unlocked in the United States by 2030, but only if organizations prepare their people and redesign entire workflows, rather than individual tasks, around people and AI working together (Yee et al., 2025). Capturing that value depends less on algorithmic sophistication than on how thoughtfully organizations design the collaboration between human and artificial minds.
The Human-AI Collaboration Landscape
Defining Hybrid Intelligence in Organizational Context
The term hybrid intelligence has a specific meaning that distinguishes it from related concepts such as automation, augmentation, or decision support. As Dellermann et al. (2019) articulated, hybrid intelligence involves three properties: collective task performance, in which the activities of humans and AI are interdependent; superior results, in which the combined system outperforms either agent working alone; and continuous learning, in which both human and machine components improve through mutual experience over time.
This definition matters because it sets a higher bar than many current AI implementations meet. A chatbot that answers customer queries without learning from human corrections is automation, not hybrid intelligence. A dashboard that presents predictions for a human to accept or reject is decision support. Hybrid intelligence implies a genuine feedback loop—a system that evolves because humans and machines are teaching each other.
Dellermann, Ebel, Söllner, and Leimeister (2019) reinforced this point in a companion article, arguing that hybrid intelligence represents a different paradigm from earlier expert systems and static knowledge repositories. Where traditional decision support systems rely on fixed rules, hybrid intelligence systems adapt, drawing on machine learning techniques that incorporate ongoing human input.
The concept draws on a rich intellectual heritage. Kamar (2016) identified directions for complementing AI systems with human intelligence, arguing that many real-world tasks require capabilities that machines and humans possess in different measure. Amershi et al. (2014) emphasized the role of people in interactive machine learning, showing that end-user involvement can improve models while making machine learning more accessible to non-experts. Dellermann et al. (2019) synthesized these threads into a unified taxonomy, giving system designers a structured way to think about the full range of design decisions involved.
Prevalence, Drivers, and Current State of Practice
The deployment of human-AI collaborative systems has expanded rapidly, driven by several converging forces. First, advances in deep learning have expanded the range of tasks where AI can contribute meaningfully, from pattern recognition and prediction to content generation (LeCun et al., 2015). Second, recognition of AI's limitations—its dependence on large volumes of data, its brittleness outside its training distribution, and its opacity in high-stakes domains—has reinforced the need for human involvement (Doshi-Velez & Kim, 2017). Third, regulation increasingly requires it: the European Union's AI Act requires that high-risk AI systems be designed for effective human oversight (Regulation [EU] 2024/1689, 2024).
A 2024 meta-analysis by Vaccaro et al., synthesizing 106 experimental studies and 370 effect sizes, offers a sobering and nuanced picture. On average, human-AI combinations performed worse than the better of humans or AI working alone. The researchers found important moderators, however: combinations tended to produce gains in content-creation tasks and when humans on their own outperformed the AI, but losses in decision-making tasks and when the AI on its own outperformed humans (Vaccaro et al., 2024). The implication is clear: simply putting a human and an AI together does not guarantee better results. The design of the collaboration, grounded in a clear understanding of each party's strengths, is what matters.
The Dellermann et al. (2019) taxonomy anticipated this insight. It organizes design choices into four meta-dimensions: the characteristics of the task itself, the learning paradigm employed, the ways humans provide input to AI, and the ways AI provides feedback to humans. Each meta-dimension contains sub-dimensions and specific design categories that together form a design space practitioners can navigate when building or improving human-AI systems.
Organizational and Individual Consequences of Hybrid Intelligence
Organizational Performance Impacts
When designed well, human-AI collaboration can produce substantial performance improvements. Brynjolfsson and Mitchell (2017) argued that the workforce implications of machine learning depend on which specific tasks it can perform, and that most jobs combine tasks that are suitable for machine learning with tasks that are not, making complementarity more common than wholesale replacement. Subsequent field evidence has borne this out.
In software development, field experiments across three companies found that developers given access to an AI coding assistant completed about 26% more tasks (Cui et al., 2025). In customer support, access to a generative AI assistant increased issues resolved per hour by 14% on average, with the largest gains for less experienced agents (Brynjolfsson et al., 2025). In a large marketing experiment, people working with AI agents were substantially more productive per worker and produced higher-quality ad copy than people working in human-only pairs, although human pairs produced higher-quality images (Ju & Aral, 2025). And in a field experiment at Procter & Gamble, individuals using AI in product development performed about as well as two-person teams working without it (Dell'Acqua et al., 2025).
Dellermann et al. (2019) identified multiple pathways through which such gains materialize, distinguishing between human augmentation (AI helps humans perform better), machine augmentation (humans help AI perform better), and hybrid augmentation (both improve together). The authors point to AlphaGo as an illustration: the system initially learned from human expert games and later produced strategies that expert players found novel and instructive, a vivid example of the reciprocal learning that defines hybrid intelligence (Silver et al., 2016).
At the organizational level, the McKinsey Global Institute's analysis emphasizes that capturing AI's value depends less on new technological breakthroughs than on redesigning complex, high-value workflows and helping human skills adapt (Yee et al., 2025). This suggests that hybrid intelligence delivers its greatest value when designed around the specific cognitive demands of core business processes rather than applied as a generic overlay.
Individual Wellbeing and Stakeholder Impacts
The effects of hybrid intelligence extend beyond productivity metrics. For knowledge workers, well-designed human-AI systems can reduce cognitive overload by handling routine pattern recognition and data synthesis, freeing attention for judgment, creativity, and interpersonal engagement. Holzinger (2016) argued that in health informatics, interactive machine learning with a human in the loop can help with complex problems where fully automated approaches struggle, while preserving the human judgment essential to patient care.
Poorly designed systems can produce the opposite effect. Research on trust in automation shows that people can both over-rely on automated systems, failing to catch their errors, and under-use them after seeing them fail (Lee & See, 2004). People who observe an algorithm err often lose confidence in it more than they would in a human making the same mistake, a pattern known as algorithm aversion (Dietvorst et al., 2015). These risks are heightened when systems lack the interpretability mechanisms that Dellermann et al. (2019) identified as crucial. Their taxonomy distinguishes three levels of interpretability: algorithm transparency (understanding how the algorithm works), global model interpretability (understanding the general logic of a trained model), and local prediction interpretability (understanding why the model produced a specific output). When none of these is present, people tend either to accept AI outputs uncritically or to reject them reflexively—neither of which produces superior results.
For end users and customers, the stakes are similarly high. Recommender systems that learn from human preferences can substantially improve service; Netflix's product leaders estimated that personalization and recommendations save the company more than $1 billion per year through improved customer retention (Gomez-Uribe & Hunt, 2016). But when such systems lack transparency about how they use personal data, or when they amplify existing biases, they can erode trust and produce discriminatory outcomes. Treating interpretability as a design dimension, not an afterthought, is a corrective organizations ignore at their peril (Lipton, 2018).
Evidence-Based Organizational Responses
Table 1: Hybrid Intelligence Design Strategies, Taxonomy Links, and Illustrative Examples
Strategy | Related Taxonomy Dimension (Dellermann et al., 2019) | Core Design Question | Key Practices | Illustrative Example |
Intentional task architecture | Task characteristics | Which parts of the work should humans, AI, or both perform? | Classify tasks by type; allocate by comparative strength; design hand-offs | Stitch Fix's algorithm-plus-stylist model |
Interactive learning integration | Learning paradigm; AI-human interaction (query strategy) | How and when does the system learn from people, and people from the system? | Active learning; online feedback; capture corrections as training signal | AlphaGo learning from human games, then informing human play |
Human-centered interpretability | AI-human interaction (interpretability) | Can people see why the AI produced a given output? | Global and local explanations; confidence displays; context-appropriate explanation | Prescience hypoxaemia predictions with explanations |
Incentive-aligned participation | Human-AI interaction (incentives, expertise, aggregation) | Why will people keep contributing high-quality input? | Intrinsic, monetary, and customization incentives; role design by expertise | Foldit, the protein-folding game |
Adaptive workflow orchestration | All dimensions, sequenced from task to interaction | Is the whole workflow redesigned around the collaboration? | Redesign before tooling; confidence thresholds; feedback infrastructure | AI-supported mammography screening in the MASAI trial |
Intentional Task Architecture Design
The most fundamental design decision in hybrid intelligence is determining which tasks should be performed by humans, which by AI, and which collaboratively. The Dellermann et al. (2019) taxonomy distinguishes among generic task types relevant to hybrid intelligence—recognition, prediction, reasoning, and action—each of which suggests a different configuration of human and machine involvement.
Recognition tasks: In areas such as object detection or speech recognition, AI can perform at or near human levels, but human oversight remains valuable for edge cases and novel categories.
Prediction tasks: In financial forecasting or demand planning, AI processes large datasets efficiently, while human expertise adds context about market dynamics, regulatory shifts, and qualitative factors.
Reasoning tasks: In strategic planning or diagnosis, humans contribute causal understanding and common sense, while AI provides structured analysis and pattern identification.
Action tasks: In robotic manipulation or automated responses, autonomy levels must be carefully calibrated, with humans providing examples and oversight for novel situations.
The Vaccaro et al. (2024) meta-analysis reinforces these distinctions. Its finding that human-AI combinations tended to underperform in decision tasks, especially where the AI alone was stronger, suggests that organizations should be particularly deliberate about task allocation in decision-heavy domains.
Stitch Fix, the online personal styling company, offers a well-documented example of intentional task architecture. The company built its operating model around algorithms that recommend items for each client and human stylists who make the final selections. Its data science team has described how algorithmic recommendations are presented to stylists through purpose-built tools, and how stylists' choices and client feedback flow back to improve the algorithms—an arrangement the team calls keeping "stylists in the loop" (Stitch Fix, 2022). Algorithms handle the large-scale matching of inventory to client data, while stylists contribute judgment about taste, context, and the client relationship that resists quantification.
Interactive Learning Integration
Static AI models that never learn from their deployment context quickly become stale. The Dellermann et al. (2019) taxonomy emphasizes that the timing of human involvement in the machine learning process and the query strategy used to collect human input are critical design choices. Input can be collected offline, after task completion; online, while the task is being performed; or through active learning, in which the system identifies cases where it is most uncertain and requests targeted human input.
Active learning strategies: Focus scarce human attention on the cases where the model is least confident.
Online query strategies: Enable teaching and learning to proceed nearly simultaneously, creating tight feedback loops.
Implicit teaching interactions: Let the system learn from observing human behavior rather than requiring explicit input, reducing the burden on participants.
Structured machine teaching: Treat the process of teaching a model as an engineering discipline, so domain experts can contribute knowledge efficiently (Simard et al., 2017).
DeepMind's AlphaGo illustrates the potential of well-designed learning integration. The system was first trained with supervised learning on a large dataset of expert human moves and then improved through reinforcement learning from games of self-play (Silver et al., 2016). As Dellermann et al. (2019) note, the system's subsequent play offered strategies that human experts found novel and instructive. This reciprocal learning—AI learning from humans, then humans learning from AI—embodies the co-evolution that defines mature hybrid intelligence.
Human-Centered Interpretability and Transparency
Trust is the oxygen of hybrid intelligence. Without appropriately calibrated trust, people either ignore AI recommendations or accept them uncritically (Lee & See, 2004). Dellermann et al. (2019) identified interpretability as a core dimension of AI-human interaction, distinguishing between algorithm transparency, global model interpretability, and local prediction interpretability.
Algorithm transparency: Allows stakeholders to understand the fundamental logic of a model, building baseline trust.
Global model interpretability: Helps users grasp what features and patterns the model relies on in general, supporting calibrated reliance.
Local prediction interpretability: Explains why a specific output was produced for a specific input, enabling informed human override.
Context-appropriate explanation: Recognizes that what counts as an adequate explanation varies by audience and purpose (Lipton, 2018).
Doshi-Velez and Kim (2017) argued that rigorous evaluation of interpretability is especially important in high-stakes domains such as medicine, criminal justice, and finance, where erroneous AI outputs can cause serious harm.
The Prescience system developed by Lundberg et al. (2018) shows interpretability in practice. Using operating-room records from University of Washington hospitals, the researchers built a model that predicts hypoxaemia (low blood oxygen) during surgery and explains each prediction by showing which patient and procedure factors drove the risk. In the study's evaluation, anesthesiologists anticipated hypoxaemia events more accurately with the system's predictions and explanations than with their own judgment alone. Clinicians retained full authority over decisions, while local explanations gave them a basis for judging when the model's reasoning was sound.
Incentive-Aligned Participation Structures
Human input is not free. One of the most practical dimensions of the Dellermann et al. (2019) taxonomy addresses why humans participate in hybrid intelligence systems. The authors distinguish monetary incentives, common on crowdsourcing platforms; intrinsic incentives such as enjoyment, learning, and purpose, as in citizen science and games with a purpose; and customization, in which users receive better service the more input they provide. They also distinguish among participants with different expertise, such as machine learning experts, domain experts, and end users, and among ways of aggregating input from many people.
Monetary incentives: Effective for discrete labeling and annotation tasks, but may not sustain deep, ongoing engagement.
Intrinsic motivators: Enjoyment, challenge, and purpose can sustain participation; games designed around useful computation show how play can produce valuable human input (von Ahn & Dabbish, 2008).
Customization-based incentives: Create a virtuous cycle in which more input yields better service, which motivates further input.
Expertise-calibrated participation: Designing different roles for machine learning experts, domain experts, and end users matches each contributor to suitable tasks (Simard et al., 2017).
Foldit offers a striking example from scientific research. Researchers turned the problem of protein structure prediction into an online game in which players manipulate three-dimensional protein models, guided by scores computed from an energy function and assisted by automated optimization tools. Players were motivated by competition, collaboration, and the intellectual challenge of the puzzles rather than payment, and top players developed strategies that, in some cases, outperformed the leading automated method then available (Cooper et al., 2010). Foldit shows how intrinsic incentives and well-designed tools can combine human spatial reasoning with computational search to produce results neither could achieve as well alone.
Adaptive Workflow Orchestration
Hybrid intelligence is not only a technology decision but a workflow design decision. Organizations that deploy AI tools without redesigning the processes those tools operate within tend to capture less value than those that rethink workflows from first principles (Yee et al., 2025). The hierarchical structure of the Dellermann et al. (2019) taxonomy, which moves from task characteristics through learning paradigms to interaction mechanisms, carries the same message: effective hybrid intelligence begins with a clear understanding of the task, not with the technology.
Process redesign before tool selection: Identify where in the workflow human judgment adds value and where AI speed and consistency matter most.
Confidence thresholds: Govern hand-offs between AI and human review according to the stakes of each decision.
Feedback infrastructure: Ensure that human corrections and overrides become learning signals that improve future performance.
Coordination mechanisms: Reconcile the continuous, batch-oriented rhythms of machines with the intermittent, attention-constrained rhythms of people.
Breast cancer screening in Sweden provides a well-documented example. In the randomized Mammography Screening with Artificial Intelligence (MASAI) trial, an AI system assigned each screening exam a risk score; lower-risk exams were read by a single radiologist, higher-risk exams by two radiologists, and the AI highlighted suspicious findings to support readers. Compared with standard double reading by two radiologists, AI-supported screening detected cancer at a similar rate while reducing radiologists' screen-reading workload by 44% (Lång et al., 2023). The gains came not from inserting AI into an unchanged process but from redesigning who reads which exams and how, with radiologists retaining responsibility for every decision.
Building Long-Term Hybrid Intelligence Capability
Continuous Co-Learning Systems
The third property in Dellermann et al.'s (2019) definition of hybrid intelligence—continuous learning—may be the most important for long-term advantage. Systems that improve through use create compounding returns, while static systems depreciate as their environment changes.
Building co-learning infrastructure requires several commitments. First, organizations must instrument human-AI interactions to capture learning signals. Every human correction, override, or approval is potential training data; organizations that discard it, or fail to collect it, leave a valuable asset on the table. Second, learning must run in both directions. It is not enough for AI to learn from human input; the system should also surface insights that help people refine their own understanding. Carter and Nielsen (2017) described this vision as using artificial intelligence to augment human intelligence—AI as a cognitive partner that expands human understanding, not merely a tool that executes instructions.
Third, organizations need feedback loops at multiple timescales. Real-time feedback, such as a stylist adjusting a recommendation, coexists with periodic model retraining on accumulated human judgments and strategic reassessment of how tasks are allocated between humans and AI. The Dellermann et al. (2019) distinction among offline, online, and active learning query strategies offers a useful framework for designing these loops.
Trust Calibration Infrastructure
Trust is not binary. Both overreliance and underuse degrade the performance of human-automation systems (Lee & See, 2004), and visible algorithm errors can trigger lasting aversion (Dietvorst et al., 2015). The goal is calibrated trust: people rely on AI when it is likely to be correct and exercise independent judgment when it is likely to be wrong.
Calibrated trust requires organizational investment in three areas. First, transparency mechanisms that make AI confidence visible to users, so they can learn when to rely on it. Second, performance tracking that gives people ongoing evidence about where the AI is more or less reliable. If, for example, clinicians could see that a tool performed very well on common conditions but much less well on rare ones, they would know where to concentrate their own vigilance. Third, training and literacy programs that help workers understand AI's strengths and limits in their specific domain. The distinction Dellermann et al. (2019) draw among machine learning experts, domain experts, and end users underscores that trust calibration must be tailored to each participant's role and expertise.
Organizational Design for Human-AI Complementarity
Sustained hybrid intelligence capability ultimately requires changes to structure, roles, and culture—not just technology. The Dellermann et al. (2019) taxonomy is explicitly socio-technical, reflecting the principle that design knowledge for hybrid systems must address both technical architecture and the human context in which it operates.
Several organizational design principles support this vision. First, roles should be defined around contribution types, not legacy job descriptions. In a hybrid intelligence organization, an analyst is less someone who builds spreadsheets than someone who contributes domain judgment to a human-AI analytical system. Second, teams should be composed with complementarity in mind. Field experiments by Ju and Aral (2025) found that AI agents' randomly assigned personality traits interacted with human collaborators' traits, with some pairings improving output quality and others reducing it—suggesting that even the "personality" of AI teammates may need to be designed to complement the people they work with. Third, governance must evolve to accommodate shared agency. When outcomes result from human-AI collaboration, accountability frameworks must clarify who is responsible for which aspects of a decision, consistent with emerging regulatory expectations for human oversight (Regulation [EU] 2024/1689, 2024).
Dellermann et al. (2019) acknowledged that their taxonomy was primarily descriptive and called for future research providing prescriptive guidance on which design choices suit which situations. Progress has been made, as the Vaccaro et al. (2024) meta-analysis and recent field experiments show, but the fundamental challenge remains: organizations must coordinate design choices across task, learning, and interaction dimensions in a strategically aligned way. Those that treat human-AI collaboration as an organizational design challenge, not just a technology purchase, will be best positioned to capture its value.
Conclusion
The central insight of hybrid intelligence research—articulated by Dellermann et al. (2019) and reinforced by subsequent evidence—is that the future of AI in organizations is not about humans versus machines. It is about the quality of the collaboration between them. The Dellermann et al. (2019) taxonomy provides a structured vocabulary for designing that collaboration, from task characteristics and learning paradigms to the mechanisms by which humans teach AI and AI informs humans.
The evidence also offers a caution. Human-AI combinations do not automatically outperform their parts; on average, they have not (Vaccaro et al., 2024). Gains depend on design: where the task plays to complementary strengths, where people can see and calibrate the AI's reasoning, and where workflows are rebuilt around the collaboration.
For leaders, several practical takeaways follow:
Design the task architecture deliberately: Allocate recognition, prediction, reasoning, and action tasks according to where human and machine capabilities are genuinely complementary.
Build interactive learning into operations: Create feedback loops so human input improves the AI and AI insight improves human judgment.
Treat interpretability as a first-class requirement: Provide explanations and confidence information suited to each user's role.
Sustain participation with the right incentives: Match monetary, intrinsic, and customization incentives to the contribution needed.
Redesign workflows, not just tools: Rebuild processes around the collaboration and measure team-level outcomes.
Most importantly, organizations should treat hybrid intelligence as a capability built over time—a co-evolutionary process in which humans and machines learn to work together more effectively with each iteration. The organizations that master it will not necessarily be those with the most advanced algorithms, but those with the most thoughtfully designed collaborations between human and artificial minds.
Research Infographic
References
Amershi, S., Cakmak, M., Knox, W. B., & Kulesza, T. (2014). Power to the people: The role of humans in interactive machine learning. AI Magazine, 35(4), 105–120.
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942.
Brynjolfsson, E., & Mitchell, T. (2017). What can machine learning do? Workforce implications. Science, 358(6370), 1530–1534.
Carter, S., & Nielsen, M. (2017). Using artificial intelligence to augment human intelligence. Distill, 2(12).
Cooper, S., Khatib, F., Treuille, A., Barbero, J., Lee, J., Beenen, M., Leaver-Fay, A., Baker, D., Popović, Z., & Foldit players. (2010). Predicting protein structures with a multiplayer online game. Nature, 466(7307), 756–760.
Cui, Z. K., Demirer, M., Jaffe, S., Musolff, L., Peng, S., & Salz, T. (2025). The effects of generative AI on high-skilled work: Evidence from three field experiments with software developers [Working paper]. SSRN.
Dell'Acqua, F., Ayoubi, C., Lifshitz-Assaf, H., Sadun, R., Mollick, E., Mollick, L., Han, Y., Goldman, J., Nair, H., Taub, S., & Lakhani, K. R. (2025). The cybernetic teammate: A field experiment on generative AI reshaping teamwork and expertise (Working Paper No. 25-043). Harvard Business School.
Dellermann, D., Calma, A., Lipusch, N., Weber, T., Weigel, S., & Ebel, P. (2019). The future of human-AI collaboration: A taxonomy of design knowledge for hybrid intelligence systems. In Proceedings of the 52nd Hawaii International Conference on System Sciences (pp. 274–283).
Dellermann, D., Ebel, P., Söllner, M., & Leimeister, J. M. (2019). Hybrid intelligence. Business & Information Systems Engineering, 61(5), 637–643.
Dietvorst, B. J., Simmons, J. P., & Massey, C. (2015). Algorithm aversion: People erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General, 144(1), 114–126.
Doshi-Velez, F., & Kim, B. (2017). Towards a rigorous science of interpretable machine learning (arXiv:1702.08608). arXiv.
Gomez-Uribe, C. A., & Hunt, N. (2016). The Netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems, 6(4), Article 13.
Holzinger, A. (2016). Interactive machine learning for health informatics: When do we need the human-in-the-loop? Brain Informatics, 3(2), 119–131.
Ju, H., & Aral, S. (2025). Collaborating with AI agents: Field experiments on teamwork, productivity, and performance (arXiv:2503.18238). arXiv.
Kamar, E. (2016). Directions in hybrid intelligence: Complementing AI systems with human intelligence. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence (pp. 4070–4073).
Lång, K., Josefsson, V., Larsson, A.-M., Larsson, S., Högberg, C., Sartor, H., Hofvind, S., Andersson, I., & Rosso, A. (2023). Artificial intelligence-supported screen reading versus standard double reading in the Mammography Screening with Artificial Intelligence trial (MASAI): A clinical safety analysis of a randomised, controlled, non-inferiority, single-blinded, screening accuracy study. The Lancet Oncology, 24(8), 936–944.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521(7553), 436–444.
Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80.
Lipton, Z. C. (2018). The mythos of model interpretability. Queue, 16(3), 31–57.
Lundberg, S. M., Nair, B., Vavilala, M. S., Horibe, M., Eisses, M. J., Adams, T., Liston, D. E., Low, D. K.-W., Newman, S.-F., Kim, J., & Lee, S.-I. (2018). Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nature Biomedical Engineering, 2(10), 749–760.
Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (2024). Official Journal of the European Union.
Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489.
Simard, P. Y., Amershi, S., Chickering, D. M., Pelton, A. E., Ghorashi, S., Meek, C., Ramos, G., Suh, J., Verwey, J., Wang, M., & Wernsing, J. (2017). Machine teaching: A new paradigm for building machine learning systems (arXiv:1707.06742). arXiv.
Stitch Fix. (2022, September 2). Stylists in the loop. Multithreaded [Stitch Fix technology blog].
Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303.
von Ahn, L., & Dabbish, L. (2008). Designing games with a purpose. Communications of the ACM, 51(8), 58–67.
Yee, L., Madgavkar, A., Smit, S., Krivkovich, A., Chui, M., Ramírez, M. J., & Castresana, D. (2025). Agents, robots, and us: Skill partnerships in the age of AI. McKinsey Global Institute.
Jonathan H. Westover, PhD, Chief Research Officer (Nexus Institute for Work and AI); Co-Founder & Chief Workforce and Learning Officer (Future State University); Founder & CEO (Human Capital Innovations); Professor of Organizational Leadership & Change (UVU). Read Jonathan Westover's executive profile here.
Suggested Citation: Westover, J. H. (2026). Designing for Hybrid Intelligence: How Organizations Can Harness the Complementary Strengths of Human and Artificial Intelligence Human Capital Leadership Review, 39(4). doi.org/10.70175/hclreview.2020.39.4.3
