Other

Osypov Pavlo

Founder & CEO, OPV Systems LLC, North Carolina, USA

Senior Software Developer, INSPYR Solutions Inc.

ORCID: 0009-0006-5289-3690

https://www.doi.org/10.25313/2524-2695-2026-4-05-38

HUMAN-AI COLLABORATION AS A MODEL FOR SCALING INDIVIDUAL EXPERTISE: FROM AUGMENTATION TO IDENTITY-LEVEL PARTNERSHIP

​​Summary. Artificial intelligence has grown capable enough to reopen a basic question: how should work be divided between humans and machines? Fully autonomous AI performs well on structured, data-rich tasks, yet collaboration between humans and AI consistently produces stronger results in complex, context-dependent, and creative settings. This article reviews how human–AI interaction models have evolved and proposes a three-level classification of collaboration: automation, augmentation, and identity-level partnership. Drawing on established complementarity frameworks and hands-on experience in AI-driven software development, the author introduces the Scalable Human Model (SHM) – a framework in which AI operates as an extension of one individual’s cognitive style, communication patterns, and decision-making logic. A case study of OPV Systems validates the model: one developer, supported by AI agents, reached productivity levels normally associated with a team of five to eight. The findings suggest that this mode of partnership differs qualitatively from augmentation – it lets professionals scale their expertise without sacrificing authenticity or strategic control. Implications for software engineering, digital legacy systems, and human-centric AI design are discussed.

Key words: human-AI collaboration, augmentation, human-centric AI, scalable human model, identity-level partnership, AI agents, complementarity.

Introduction. The rapid advancement of artificial intelligence has reshaped the discourse on the future of work, with much of the public and industry attention focused on the capacity of AI to replace human labor. The automation-first perspective, reinforced by breakthroughs in large language models, computer vision, and autonomous agents, frames AI primarily as a substitute for human effort [1; 4]. A growing body of evidence complicates this picture. Agrawal, Gans, and Goldfarb [12] show that AI’s performance ceiling is set by the quality and scope of its training data; where tasks demand tacit knowledge, contextual judgment, or ethical reasoning, human cognition still holds a clear advantage.

Empirical studies increasingly confirm that the highest-quality outcomes emerge from structured collaboration between humans and AI. Kellogg, Lifshitz-Assaf, and Shrestha [3] report that in image classification tasks, human-AI teams achieved 90% accuracy, compared to 81% for humans alone and 73% for AI alone. Hemmer et al. [4] formalize this observation through a task allocation framework in which AI automates routine tasks, augments human judgment on moderately complex ones, and defers entirely to humans on novel or ambiguous problems. Steyvers et al. [5] provide further theoretical grounding by identifying information asymmetry and capability asymmetry as the two primary sources of complementarity: humans and AI systems tend to make qualitatively different errors, creating conditions under which their combination outperforms either in isolation.

Accumulated evidence has catalyzed a shift toward human-centric AI design. The World Economic Forum [1] and the EIT Deep Tech Talent Initiative [2] identify the future of work as a redesign of professional activity itself, where AI amplifies human creativity, strategic thinking, and emotional intelligence. McKinsey’s 2025 report [9] estimates that AI could automate up to three hours of daily routine activity per worker by 2030, with the primary economic value lying in augmentation. Deloitte [8] similarly emphasizes that generative AI achieves its full potential when integrated with uniquely human skills – critical thinking, empathy, and complex problem-solving. Gonzalez et al. [13] at Carnegie Mellon University reinforce this perspective through the COHUMAIN framework, which positions AI as a collaborative partner that strengthens existing team dynamics under human direction.

For all this progress, current collaboration models stop at the task level: they decide which discrete tasks to hand to humans and which to machines, or they feed AI-generated recommendations into human decisions [4; 6]. A deeper form of partnership – one in which AI internalizes a specific individual’s cognitive style, communication habits, and decision-making logic, then preserves and scales that person’s professional identity across parallel workflows – has received little formal attention. The distinction from augmentation is qualitative, not incremental. An augmentation tool offers generic capabilities any user can apply; an identity-level system picks up how a specific person reasons, writes, and decides, then handles enough parallel work that one professional covers ground usually requiring a full team.

What follows is an account of how human-AI collaboration models have developed, leading to a new framework: the Scalable Human Model (SHM). SHM is organized in three layers. In the Identity Layer, AI builds and maintains a working model of how the person thinks and communicates. In the Execution Layer, AI agents perform tasks according to that model; and the Control Layer, where the human retains strategic oversight, ethical governance, and creative direction. Empirical grounding comes from the author’s experience developing OPV Systems, an AI platform for constructing persistent digital personas, where a single developer working with AI agents achieved output comparable to a conventional multi-person team. Through comparative analysis with established models such as the EPOCH methodology [10], the COHUMAIN architecture [13], and the automation-augmentation framework [4], the article positions identity-level partnership as a qualitatively new stage in the evolution of human-AI collaboration, with practical implications for software engineering, digital legacy, and the broader design of human-centric AI systems.

Literature Review

Evolution of Human-AI Interaction: From Tool to Partner. The conceptual understanding of AI’s role in professional workflows has undergone three distinct shifts since the mid-20th century. First-generation AI systems, rooted in rule-based expert systems and statistical classifiers, functioned as deterministic tools that executed pre-programmed instructions without adaptation to user context [12]. A second generation, driven by advances in machine learning and neural networks, introduced AI as an intelligent assistant capable of pattern recognition, recommendation, and limited autonomous judgment – exemplified by diagnostic support systems in radiology and fraud detection in financial services [6]. A third and current generation, catalyzed by large language models and multimodal architectures, positions AI as a collaborative partner that can generate novel content, engage in multi-turn reasoning, and co-create with human operators in real time [7; 11].

Brynjolfsson [11] frames this transition through what he terms the “Turing Trap” – the tendency to evaluate AI systems by their ability to replicate human performance instead of complementing it. He argues that the economic and social value of AI is maximized when machines enable humans to accomplish what was previously impossible. Dellermann et al. [14] build on this reasoning in their Hybrid Intelligence paradigm: the best outcomes, they argue, emerge from networked systems that pair human creativity and ethical reasoning with AI’s computational speed and scalability. Lin et al. [7] offer concrete evidence from science itself – their SciSciGPT platform shows that pairing a researcher’s domain intuition with AI-driven bibliometric analysis measurably accelerates the pace of discovery.

Mechanisms of Human-AI Complementarity. The theoretical foundation for understanding why collaboration outperforms solo performance lies in the concept of complementarity. Steyvers et al. [5] identify two distinct sources. Information asymmetry, the first source, occurs when humans and AI have access to qualitatively different information about the same problem: for instance, a physician draws on patient history, non-verbal cues, and clinical intuition, while an AI model analyzes imaging data across thousands of cases. Capability asymmetry, the second source, arises because humans and AI excel at fundamentally different cognitive operations – humans at causal inference, abstract reasoning, and ethical judgment; AI at statistical pattern detection, exhaustive search, and consistent execution under high data volumes.

Hemmer et al. [4] put this theoretical distinction to work in a task allocation framework. In their experiments, concentrating 80% of human effort on the hardest 20% of tasks – those marked by novelty, ambiguity, or thin training data – while letting AI handle routine work autonomously produced the strongest results. A key finding is that between-task complementarity (assigning whole tasks to whichever agent is better suited) and within-task complementarity (AI supporting human judgment on the same task) reinforce each other; the highest aggregate performance appeared only when both operated together.

Gomez et al. [6] map the full range of interaction patterns in AI-assisted decision-making, from complete AI autonomy through turn-taking collaboration to human-led processes with AI support. Surveying deployments in healthcare, finance, and content moderation, they find that most production systems use only a handful of these patterns – a gap that points to substantial untapped design space for richer collaborative configurations.

Fig. 1. Comparative accuracy across three domains: human only, AI only, and human-AI collaboration. Illustrative representation of reported performance trends across three domains, based on empirical findings in [3; 4; 5]. Exact values are approximate and compiled for comparative visualization

Existing Frameworks for Human-Centric AI Collaboration. Several frameworks have been proposed to operationalize the principles of human-AI collaboration, each addressing a different level of analysis. At the task level, Hemmer et al. [4] provide a decision-theoretic model for optimal work distribution between humans and AI, demonstrating measurable accuracy improvements. At the team level, the COHUMAIN framework developed by Gonzalez et al. [13] at Carnegie Mellon University draws on organizational psychology to examine how AI agents can strengthen group dynamics, improve communication, and offset individual decision-making biases within existing teams. COHUMAIN treats AI as organizational infrastructure that enhances collective intelligence.

At the occupation level, the EPOCH methodology introduced by Loaiza and Rigobon [10] at MIT provides a quantitative scoring system that evaluates jobs across five dimensions of human-intensiveness: Empathy, Presence, Opinion, Creativity, and Hope. Tasks scoring low on the EPOCH index are candidates for automation, while high-scoring tasks represent domains where human agency remains irreplaceable. The framework enables organizations to diagnose which roles are most susceptible to AI substitution and which require augmentation strategies, but it functions as a workforce planning tool rather than a model for real-time interaction.

Dellermann et al. [14] take a system-level view. In their Hybrid Intelligence framework, productive collaboration is not a property of any single human–AI pair but emerges from networks in which multiple humans and AI agents jointly tackle complex problems. Drawing on collective intelligence research, they stress that how the interface between humans and AI is designed matters as much as what either side can do on its own.

Table 1

Comparative analysis of human-AI collaboration frameworks

Source: compiled by the author based on [4; 5; 10; 13; 14]

Identified Gap: The Absence of Identity-Level Models. Table 1 lines up six comparison criteria against each reviewed framework. Between them, the models cover tasks, teams, occupations, and networked systems. The consistent blind spot: none describes a long-term arrangement where an AI system learns one professional’s way of thinking, communicating, and deciding, then reproduces it over time.

On the ground, this kind of partnership is already taking shape. In software engineering, AI coding agents pick up repository-specific conventions, architectural preferences, and naming patterns from the developer who directs them [15]. In content creation, AI systems are being fine-tuned to match an individual author’s tone and rhetorical strategies. Yet no published framework gives the phenomenon a formal name or structure.

Without such a model, the field has no shared vocabulary for evaluating, designing, or governing AI systems that persistently extend one person’s professional identity. SHM fills that space. Its three-layer architecture captures what makes identity-level partnership distinct and draws a clear boundary against conventional augmentation or automation.

Materials and Methods. A mixed-method research design was employed, combining conceptual modeling, structured comparative analysis, and a single embedded case study. Conceptual modeling serves as the primary method for constructing the Scalable Human Model (SHM). The model was developed through an iterative process: five established human-AI collaboration frameworks (automation-augmentation [4], COHUMAIN [13], EPOCH [10], complementarity [5], and Hybrid Intelligence [14]) were systematically analyzed across four parameters – unit of analysis, AI role typology, degree of personalization to individual users, and temporal scope of collaboration. Dimensional mapping revealed that none of the existing frameworks explicitly models a scenario in which AI internalizes and reproduces the cognitive and communicative identity of a specific individual over time, providing the structural basis for the three-layer SHM architecture.

Structured comparative analysis was applied to position SHM within the existing body of knowledge. Six uniform criteria – collaboration level, AI role, human role, personalization degree, temporal scope, and primary limitation – were used to evaluate each framework, ensuring that the claimed novelty of SHM is substantiated by systematic demonstration of how it extends dimensions left unaddressed by prior work [6]. Where a framework does not explicitly address a comparison dimension (e.g., identity consistency), it was coded as “absent” instead of being inferred, to avoid overinterpreting the original authors’ intent.

The empirical component follows an embedded single-case design, with OPV Systems – an AI platform for constructing persistent digital personas, founded and developed by the author – serving as the primary unit of analysis. Its core functionality of capturing and reproducing an individual’s communication style and decision-making patterns makes it a naturally occurring instance of identity-level partnership. The case study examines three aspects: (a) the technical architecture and its role in preserving identity consistency; (b) the AI-augmented development workflow, where a single developer used AI agents configured to maintain the author’s architectural conventions and quality standards for code generation, documentation, and project management; and (c) a quantitative comparison of output against industry benchmarks for conventional teams. Data were collected from the project’s version control system (Git), structured development logs, and the publicly deployed MVP.

A single case cannot support statistical generalization, but statistical generalization is not the objective here. The goal is analytical generalization – testing whether a theoretical framework holds when confronted with a concrete instance, not extrapolating frequencies across populations. Potential bias from the author’s dual role as researcher and practitioner is mitigated by grounding the analysis in objectively verifiable data (version control records, publicly available product) and by subjecting the proposed model to structured comparison with independently published frameworks.

Results

Three-Level Classification of Human-AI Collaboration. Comparing the five frameworks produced a three-level classification. At Level 1, Automation, AI executes predefined tasks under human supervision while the person monitors and handles exceptions. At Level 2, Augmentation, AI generates recommendations, drafts, or surfaced information to support the person’s decisions; authorship stays with the human. Level 3, Identity-Level Partnership, is the configuration where AI persistently extends one individual’s professional identity, reproducing that person’s reasoning, communication style, and quality standards across parallel workflows.

What distinguishes Level 3 from Level 2 is the nature of AI involvement, rather than its volume. An augmentation tool (Level 2) delivers generic capabilities that any user can apply: code completion, grammar correction, data summarization. An identity-level partner (Level 3) delivers outputs indistinguishable from those the specific human would produce independently, because the AI has internalized that person’s particular conventions, preferences, and decision heuristics. At Level 2, switching the human operator changes the character of the output; at Level 3, the AI keeps outputs consistent even when scaling across tasks the human could not personally handle at once.

The Scalable Human Model: Architecture and Layer Functions. SHM translates identity-level partnership into three layers. The Control Layer is exclusively human territory. It houses creative direction (vision, goals, novel ideas), ethical governance (checking AI outputs against professional and moral standards), and strategic decision-making (setting priorities, managing trade-offs, pivoting). None of these functions can be delegated; Because they stay with the person, authorship and accountability hold no matter how much AI handles downstream.

AI manages the Identity Layer, but the human’s actual work shapes it. The layer acts as a persistent persona memory with three components. The cognitive profile records how the person reasons, decomposes problems, and applies heuristics. Communication patterns cover vocabulary, tone, rhetorical structure, and preferred formats. Decision logic captures risk tolerance, priority weightings, and evaluation criteria. The layer updates incrementally with each new piece of work the person produces, keeping the AI’s representation current without manual reconfiguration. Output quality rests on three factors working together: the human’s creative input, the number of parallel AI agents, and how accurately those agents mirror the human’s profile. Weaken any one of them and output quality drops.

Specialized AI agents occupy the Execution Layer. They work autonomously, but the Identity Layer configures each one to produce outputs matching the human’s professional profile. At OPV Systems, agents covered four domains: code generation in the author’s architectural style, documentation in the author’s voice, correspondence following the author’s rhetorical patterns, and sprint planning with risk flags aligned to the author’s priorities. A feedback loop carries corrections from the Control Layer back into the Identity Layer, refining its parameters with each review cycle.

Fig. 2. Three-layer architecture of the Scalable Human Model (SHM). The Control Layer (human) directs the Identity Layer (AI persona memory), which configures the Execution Layer (AI agents). A feedback loop enables continuous refinement

Source: developed by the author

Case Study Results: OPV Systems Productivity Analysis. The OPV Systems case study provides quantitative evidence of the productivity impact achievable through identity-level human-AI partnership. Over the MVP development cycle, a single developer operating within the SHM configuration delivered a fully functional web platform comprising a Next.js frontend, Express.js backend, OpenAI API integration, Three.js-based visualization layer, authentication system, and investor-facing presentation – in approximately eight person-weeks of effort. Industry benchmarks for comparable full-stack products, based on published estimation data and the author’s nine years of professional experience, place the equivalent effort for a conventional team at 28 to 42 person-weeks distributed across five to eight specialists.

Table 2 shows the comparison by task category. Documentation and investor materials had the widest gap (×6.7) – agents delivered structurally complete drafts in the author’s writing style, and the human contributed only strategic edits. Core architecture followed at ×5.3: agents took care of boilerplate code, standard patterns, and routine debugging; the developer focused on architectural choices and novel logic. Testing and deployment had the narrowest ratio (×3.5), reflecting the context-sensitivity of QA work and correspondingly more frequent human corrections.

Table 2

Productivity comparison: conventional team versus SHM configuration (OPV Systems MVP)

Source: compiled by the author based on project version control data and industry estimation benchmarks

Two qualitative patterns stood out. First, identity consistency held from start to finish. Code kept the same naming conventions, architectural patterns, and commenting style throughout the project; documentation maintained a single authorial voice with no post-hoc harmonization needed. Second, agents got better with use. Substantive revision, meaning any change to logic, architecture, or naming conventions rather than cosmetic formatting, was needed for roughly 35–40% of AI-generated code in week one but only 12–15% by the final week. Each correction cycle fed back into the Identity Layer and sharpened its parameters.

Discussion

Why Identity-Level Partnership Is Qualitatively Distinct. What the results point to is a categorical, not incremental, shift. An augmentation tool speeds up whoever happens to use it; between sessions it retains nothing about that person. An identity-level partner, by contrast, shapes every output around one individual’s internalized profile. The organizational parallel is the difference between hiring a generic contractor who can do the work and training a long-term associate who anticipates your decisions. SHM makes the boundary explicit: the Execution Layer handles the what, the Identity Layer encodes the who, and the same execution infrastructure produces fundamentally different outputs depending on whose identity it carries.

Separating execution from identity also reframes the “Turing Trap” identified by Brynjolfsson [11]. Brynjolfsson warns against designing AI that merely imitates average human performance. Identity-level partnership avoids this trap by making AI more specific: the goal is to extend a particular human’s unique capabilities. In economic terms, the value equation flips. Instead of replacing workers with cheaper automation, the goal becomes amplifying what each professional can do: more clients served, more products built, more projects contributed to, all without diluting personal quality standards.

Relationship to Existing Frameworks. As a conceptual model, SHM is complementary to the frameworks reviewed in Section 2. The task allocation model of Hemmer et al. [4] remains valid within the Execution Layer, where the distribution of specific tasks between AI agents and the human operator follows the same efficiency logic. The complementarity framework of Steyvers et al. [5] operates at each layer: information asymmetry exists between the Control Layer (where the human holds strategic context) and the Execution Layer (where AI agents hold operational data), while capability asymmetry governs the Identity Layer (where AI excels at pattern storage and retrieval, but the human defines what patterns matter). COHUMAIN [13] addresses scenarios where multiple humans collaborate with AI; SHM addresses the scenario where a single human collaborates with multiple AI agents. The two frameworks apply to different organizational configurations.

The EPOCH methodology [10] provides a useful diagnostic for determining which professional roles are most suitable for SHM deployment. Roles scoring high on the Creativity and Opinion dimensions but moderate on Presence and Empathy – such as software architects, technical writers, and strategic consultants – represent the strongest candidates, because their work depends heavily on individual cognitive style (which the Identity Layer can capture) but less on real-time physical or emotional co-presence (which AI cannot replicate). Roles scoring high on Empathy and Presence, such as therapists or emergency responders, may benefit from SHM in auxiliary functions (documentation, administrative communication) but are less likely to delegate core tasks to the Execution Layer.

Ethical Dimensions of Identity-Level AI. Working at the identity level raises ethical questions that simpler tools do not. The most immediate is authorship. If an AI agent writes a document following the author’s style and reasoning, and the author reviews and approves it, who authored it? Under SHM the answer is built into the design: the human holds final review authority and accepts responsibility for every output, which places the AI in the role of a production tool. A law firm offers the closest parallel: a senior partner signs briefs that associate drafted, and institutional authorship sits with whoever reviews and takes responsibility.

A more complex concern arises in the context of digital legacy, where the Identity Layer preserves a person’s communicative profile beyond their active participation – or, in the case of OPV Systems’ “Last Call” product, beyond their lifetime. The digital legacy application raises questions of consent (did the individual authorize the creation and use of their digital persona?), boundaries (what topics should the persona be permitted to address?), and psychological impact (does interacting with a digital persona help or hinder the grieving process?). While these questions exceed the scope of the present study, they represent a critical research frontier. SHM’s Control Layer provides a structural hook for governance policies, but filling in the substance of those policies will require technologists, ethicists, psychologists, and legal scholars working together.

Practical Implications and Application Domains. Five application domains were identified for SHM deployment, mapped in Table 3, specifying for each domain the relevant Identity Layer components, the Execution Layer tasks, the expected impact, and the current level of technological readiness. Software engineering is the most immediately viable domain — the case study reported here documents a fivefold productivity increase with direct empirical backing. Digital legacy and grief therapy sit at medium readiness: OPV Systems has reached TRL4 with the “Last Call” prototype, built for veterans and their families. Education, professional services, and creative industries offer longer-term prospects but will need domain-specific tuning of the Identity Layer and resolution of regulatory questions.

Table 3

Application domains for the Scalable Human Model: Identity Layer configuration, tasks, impact, and readiness

Source: developed by the author based on the SHM framework and OPV Systems project data

Limitations and Future Research Directions

The most obvious limitation is the single-case design. OPV Systems offers an internally consistent picture of identity-level partnership, but productivity ratios and calibration dynamics may look different in other domains, with other team sizes, or on other AI platforms. Measurement is a second concern: the author assessed identity consistency qualitatively – and was simultaneously the person whose identity the system replicated, which invites subjectivity. A natural next step is developing quantitative metrics. Blind protocols where independent reviewers try to distinguish AI-generated outputs from human-generated ones would be one approach.

Three lines of follow-up research stand out. First, multi-case studies deploying SHM in software engineering, content creation, and consulting would reveal whether the three-layer architecture transfers cleanly or needs adaptation. Second, tracking the Identity Layer’s calibration curve over months or years would clarify whether the correction-rate improvement seen here, from 35–40% down to 12–15%, is replicable. Third, digital legacy, where getting identity wrong carries the highest stakes, calls for interdisciplinary research on ethical governance.

Conclusions. The goal was to formalize a mode of human–AI collaboration that current frameworks overlook: identity-level partnership, where AI acts as a lasting extension of one person’s professional identity. The three-level classification and SHM’s three-layer architecture supply both the vocabulary and the structural scaffold needed for designing, analyzing, and governing such systems.

In the OPV Systems case, one developer operating under SHM achieved roughly fivefold productivity gains relative to conventional team benchmarks, peaking in documentation (×6.7) and bottoming in context-sensitive testing (×3.5). Correction rates declined steadily over the development cycle, offering early evidence that the Identity Layer learns through feedback and progressively improves output fidelity.

Signs point beyond software engineering. Early work on digital legacy, specifically the “Last Call” prototype, and identified openings in education, professional services, and creative industries suggest the approach is not locked to one domain. The guiding principle is to keep creative and ethical control with the human while delegating scalable execution to AI. That principle applies wherever individual expertise needs to reach further without losing its character.

In this case, the most productive collaboration happened when AI worked the way the specific human would. Fidelity to one person’s cognitive profile, communication style, and quality standards let a single professional match the output and coherence of a full team. Whether the pattern holds in other fields is an open empirical question, but the architectural principle of separating what the AI does from whose identity it carries gives future research a concrete place to start.

References

1. World Economic Forum (2025). How Human-Centric AI Can Shape the Future of Work. URL: https://www.weforum.org/stories/2025/09/human-centric-ai-shape-the-future-of-work/
2. EIT Deep Tech Talent (2025). The Future of Human and AI Collaboration. URL: https://www.eitdeeptechtalent.eu/news-and-events/news-archive/the-future-of-human-ai-collaboration/
3. Kellogg, K., Lifshitz-Assaf, H., Shrestha, Y. R. (2025). When Humans and AI Work Best Together – and When Each Is Better Alone. MIT Sloan Management Review. URL: https://mitsloan.mit.edu/ideas-made-to-matter/when-humans-and-ai-work-best-together-and-when-each-better-alone
4. Hemmer, P., Schemmer, M., Vetter, S., Kühl, N. (2025). Roles of Artificial Intelligence in Collaboration with Humans: Automation, Augmentation, and the Future of Work. Management Science. DOI: 10.1287/mnsc.2024.05684
5. Steyvers, M., et al. (2025). Complementarity in Human-AI Collaboration: Concept, Sources, and Evidence. European Journal of Information Systems, 34(6), pp. 979–1002. DOI: 10.1080/0960085X.2025.2475962
6. Gomez, C., Cho, S. M., Ke, C., Huang, C.-M., Unberath, M. (2024). Human-AI Collaboration Is Not Very Collaborative Yet: A Taxonomy of Interaction Patterns in AI-Assisted Decision Making. Frontiers in Computer Science, 6. DOI: 10.3389/fcomp.2024.1521066
7. Lin, Z., Ding, Y., Wu, L., Wang, D. (2025). SciSciGPT: Advancing Human–AI Collaboration in the Science of Science. Nature Computational Science. DOI: 10.1038/s43588-025-00906-6
8. Deloitte (2024). Generative AI and the Future of Work. URL: https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/articles/generative-ai-and-the-future-of-work.html
9. McKinsey & Company (2025). Superagency in the Workplace: Empowering People to Unlock AI’s Full Potential at Work. URL: https://www.mckinsey.com/capabilities/tech-and-ai/our-insights/superagency-in-the-workplace
10. Loaiza, I., Rigobon, R. (2024). The Epoch of AI: Human-Machine Complementarities at Work. Available at SSRN 5028371. Massachusetts Institute of Technology.
11. Brynjolfsson, E. (2022). The Turing Trap: The Promise and Peril of Human-Like Artificial Intelligence. Daedalus, 151(2), pp. 272–287. DOI: 10.1162/daed_a_01915
12. Agrawal, A., Gans, J., Goldfarb, A. (2018). Prediction Machines: The Simple Economics of Artificial Intelligence. Harvard Business Review Press, Boston.
13. Gonzalez, C., Admoni, H., Brown, S., Woolley, A. W. (2025). COHUMAIN: Building the Socio-Cognitive Architecture of Collective Human–Machine Intelligence. Topics in Cognitive Science, 17(2), pp. 180–188. DOI: 10.1111/tops.12673
14. Dellermann, D., Ebel, P., Söllner, M., Leimeister, J. M. (2019). Hybrid Intelligence. Business & Information Systems Engineering, 61, pp. 637–643. DOI: 10.1007/s12599-019-00595-2
15. Osypov, P. (2026). Why Human + AI is Stronger than Just AI. OPV Systems Blog. URL: https://opvsystems.com/blog/why-human-ai-is-stronger-than-just-ai