When Humans and AI Agents Form a True Team: A New Framework for Collective Intelligence at Work

Artificial intelligence systems are no longer occasional tools summoned for a single question. They are becoming recurrent members of organizational life, participating day after day in interdependent work alongside human colleagues. Yet, according to a new theoretical framework published in Frontiers in Organizational Psychology, research on this phenomenon remains fragmented: AI studies focus on system capabilities, adoption research on individual human responses, experiments on one-off interactions, and team science on constructs built exclusively for human groups. A team of Italian researchers from the University of Studies G. d’Annunzio Chieti and Pescara, led by Alessandro Rovetta with Alessandro Bortolotti and Claudio Di Berardino, now proposes a disciplined way out of this fragmentation. Their framework, called Organismic Human-Artificial Intelligence, or OHAI, treats the persistent human-AI team itself as the central unit of analysis.

The core idea is deceptively simple but consequential. A persistent organizational human-AI team is defined as a bounded collective in which distinguishable human and artificial contributors engage in interdependent activity across successive cycles, and in which prior collective activity can affect later artificial participation. That last condition, termed functional continuity, is what separates genuine hybrid teams from the far more common pattern of repeated, episodic use. A technically stateless base model can still belong to a persistent team if the deployed system can access and use externally maintained state, such as interaction history stored in an orchestration layer or shared knowledge infrastructure. Conversely, simply archiving records is not enough if that history never enters subsequent interaction. One-shot exchanges, sessions that functionally reset, and the static aggregation of separately produced human and AI outputs all fall outside the domain.

OHAI organizes inquiry around three analytically distinct objects: system configuration, collective functioning, and collective intelligence. System configuration is a structured, potentially time-indexed description of the pre-specifiable human, artificial, task, organizational, and temporal features of the hybrid system, including expertise, model capabilities, memory architecture, decision rights, roles, and information access. Collective functioning concerns what actually happens when configured contributors work together: how expertise is accessed, distributed, evaluated, combined, and updated, and how errors are detected and repaired. Collective intelligence, the framework’s primary theoretical target, is defined in line with landmark work by Anita Woolley and colleagues as a general group-level capacity manifested across diverse tasks, whose scope of generalization must be specified and empirically demonstrated rather than assumed.

The framework’s central inferential rule is its most provocative feature: a property implemented in an artificial component does not by itself establish a corresponding team-level phenomenon. An AI system that maintains an expertise map, for instance, does not automatically create a transactive memory system in the team. The team-level phenomenon must be demonstrated in the enacted organization and use of distributed contributions, that is, in evidence showing that questions are actually directed to appropriate contributors, that information is routed to where it can be used, and that source selection changes with experience. The rule cuts both ways: a team-level function need not depend on any particular AI architecture, since similar collective functions may be realized through different combinations of human practices, artificial features, task structures, and organizational arrangements.

To make collective functioning measurable, the authors propose three selective lenses drawn from human team research. Partner modeling, related to theory of mind, concerns the representations contributors hold of one another, whether psychological beliefs on the human side or computational user models, state tracking, and behavioral prediction on the artificial side. Transactive-memory-system-related coordination concerns how differentiated expertise is located and used across the collective, extending Wegner’s classic account of who knows what. Shared-mental-model-related representational alignment concerns whether human psychological representations and artificial computational task states are sufficiently compatible to support anticipation and coordination. In every case, the framework demands evidence at the team level and carefully distinguishes human psychological states from artificial computational operations, without assuming psychological symmetry between contributors.

Persistence is where the framework becomes genuinely dynamic. Once prior collective activity can shape later participation, the team cannot be represented as a sequence of independent encounters. Recursive updating names the temporal relation through which earlier outputs, expectations, routines, and interaction histories enter the conditions of subsequent activity, whether deliberately or not, symmetrically or not. Change can unfold at overlapping timescales, from within-episode adjustments to slow accretion of roles and procedures. This history sensitivity makes collective intelligence trajectories, the temporal patterns in repeated estimates of the same CI construct, legitimate empirical objects. But the authors insist that improvement, learning, or adaptation must never be presumed: accumulated history may just as easily stabilize errors, obsolete representations, or ineffective routines. Path dependence is an open empirical question, not a developmental promise.

Estimating a CI trajectory is methodologically demanding. Because performance in a single episode can reflect task difficulty, temporary coordination failures, or stochastic variation in AI outputs, claims about change in the general collective capacity require repeated estimation through comparable or defensibly linked multi-task task batteries, with sufficient task diversity to separate general capacity from episodic performance. Studies must also keep three questions distinct: whether the same collective continues, whether its claim-relevant configuration remains stable, and whether the measure remains comparable across periods. A team that swaps its AI model or restructures roles may still be a continuing collective, but attributing observed change to accumulated experience, to learning, or to the reconfiguration itself requires a comparison design appropriate to the specific claim.

The framework also launches a secondary comparative program on cross-substrate correspondence, asking whether selected human-team phenomena recur or perform comparable functions in hybrid teams. It distinguishes observable-pattern correspondence, functional correspondence, and a third form, relational-pattern correspondence, reserved as a limited future methodological possibility. Four test families organize the empirical work: interaction-pattern correspondence, informational synergy, failure-process correspondence, and correspondence in coordination structures. Notably, similar final errors do not establish the same underlying process, a caution drawn from the classic hidden-profile tradition, and informational synergy alone does not identify any particular coordination mechanism. Divergence from human-team references is treated as theoretically valuable in its own right, potentially revealing genuinely hybrid forms of organization.

To make the agenda cumulative, the authors describe four nonmandatory stages of research maturity. Stage one qualifies the configuration and measurement, documenting model versions, memory architectures, retrieval arrangements, roles, and authority structures. Stage two runs circumscribed local tests between selected configuration features, collective functioning, and performance or CI. Stage three addresses longitudinal team-level phenomena such as recursive updating and CI trajectories, with material configuration changes carefully documented. Stage four supports integrated multilevel studies linking configuration, functioning, intelligence, and downstream consequences. The authors stress that these stages organize research maturity, not team development, and that a well-qualified negative result should constrain or narrow a local claim rather than be expanded away post hoc.

Perhaps the most human dimension of the paper lies in its treatment of employee wellbeing. Persistent hybrid teams organize not only collective problem solving but also the relationships through which people experience their work: whom members consult, whose judgments carry weight, and how responsibilities and recognition are distributed. Citing evidence that transactive memory systems and shared mental models relate to job satisfaction, engagement, and identification in human teams, the authors call for research on how artificial participation reshapes these processes. They propose practical tools, including role-and-expertise maps, participative reviews, and longitudinal assessment profiles, to detect whether improved collective capability comes with concentrated verification burdens or narrowed recognition. The broader message is that organizations configure the conditions under which AI functions become part of collective life, and that governing that configuration, including the history it carries forward, is now an organizational responsibility, not merely a technical one.

Subject of Research: A theoretical framework for collective intelligence in persistent organizational human-AI teams

Article Title: When humans and agents form a collective: a framework for collective intelligence in persistent organizational human-AI teams

Article References: Rovetta, A., Bortolotti, A., & Di Berardino, C. (2026). When humans and agents form a collective: a framework for collective intelligence in persistent organizational human-AI teams. Frontiers in Organizational Psychology, 4, Article 1941058. https://doi.org/10.3389/forgp.2026.1941058

Image Credits: AI Generated

DOI: 10.3389/forgp.2026.1941058

Keywords: collective intelligence, human-AI teams, organizational psychology, transactive memory, team cognition, OHAI framework, employee wellbeing, recursive updating, cross-substrate correspondence, AI agents, team dynamics, theory of mind