Working with Agentic `Teammates': When a New Organizational Actor Collides with the Human Ecosystem of Work

Abstract: Enterprise AI is transitioning from single-user, reactive tools toward proactive, multi-user 'teammates,' but our empirical understanding of this transition is limited. In this paper, we present an in-situ qualitative study of a persistent, proactive AI agent 'teammate' deployed across multiple teams in a large technology company. Our findings reveal the boundaries of the human-agent workplace are actively in flux, triggering breakdowns and negotiations across: 1) tacit rules of collaborative human workflows, 2) the relational boundaries of this new non-human actor, and 3) the redistribution of trust and human agency. We use these early micro-negotiations as signals to chart a new research, design, and organizational agenda that intentionally preserves human agency in a workplace shared with non-human organizational actors.

• The study shows that automated agent-like AI systems can successfully integrate into collaborative workspaces and interpret information.
• It also clarifies, by mocking and directly comparing human mistakes, that agent-led workflows need to be designed for enhanced temporal decision making capabilities.
• Finally, the paper advocates for progressive governance of AI agents to ensure appropriate proactivity and trust-building.

The paper examines how a persistent, proactive AI agent becomes incorporated into the social, procedural, and organizational infrastructure of human work. Its central claim is that an agent embedded in shared workspaces cannot be understood adequately either as a conventional software tool or as a human-equivalent teammate. Instead, it should be treated as a non-human organizational actor: an entity with operational and structural agency, but without human tacit knowledge, phenomenological experience, relational stakes, or accountability. This conceptual distinction organizes the paper’s empirical findings and its recommendations for agent design, organizational governance, and HCI research.

Research problem and conceptual framing

The paper addresses a transition from single-user, reactive AI systems toward multi-user agents that operate proactively within team environments. Unlike conventional assistants, these systems can monitor shared context, initiate interactions, modify artifacts, coordinate tasks, communicate across channels, and execute multi-step workflows. The relevant interaction is therefore not limited to a user issuing a command and receiving an output. The agent becomes a participant in the informational and social infrastructure through which work is organized.

The authors identify a limitation in existing human-AI teaming research: much of the literature studies short-lived interactions in laboratory settings, simulated environments, games, flight simulators, or Wizard-of-Oz systems. Such research has examined trust calibration, mental models, communication strategies, and team performance, but generally does not capture the tacit practices and pre-existing relationships that characterize organizational work. The paper consequently adopts a socio-material and CSCW-oriented perspective in which technologies and organizational practices mutually constitute one another.

The term “teammate” is used pragmatically rather than ontologically. The authors do not claim that the agent is a person or that human and machine participants are relationally equivalent. Rather, the term identifies a deployment paradigm in which an AI system is persistent, multi-user, embedded in shared workspaces, and capable of independently initiating actions. The paper’s stronger theoretical proposal is that this paradigm requires a category beyond the conventional tool–teammate binary.

Study design and deployment context

The empirical study uses Team Agent as an in-situ technology probe rather than as a finished product evaluation. Team Agent was deployed across more than 20 teams in a large technology company, including research, product, engineering, and technical-support organizations. Over a five-month period, the deployment generated more than 41,000 conversational turns and more than 11,000 agent responses.

The agent operated through a dedicated service account and was integrated into team chats, shared documents, meetings, bug-tracking systems, code repositories, and internal knowledge stores. It could be explicitly invoked through tagging, but it could also act proactively. Its capabilities included:

• filing and prioritizing bugs;
• summarizing conversations and documents;
• scheduling meetings and preparing briefings;
• creating and organizing documents;
• searching internal and external information sources;
• monitoring project status;
• identifying discrepancies between discussions and project artifacts;
• nudging conversations toward meetings or private channels;
• sending direct messages;
• providing interpersonal communication feedback.

The deployment included several mechanisms intended to preserve privacy and user control. Teams opted in to adoption, administrators configured shared access, direct-message initiation could be restricted through an allow list, permissions could be revoked, and users could mark content with #NoAgent. One-on-one conversations were excluded from the agent’s group-space context. These controls are analytically important because the findings show that technical safeguards do not automatically produce a corresponding perception of agency, privacy, or consent among workers.

The qualitative study comprised semi-structured interviews with 17 participants across 11 teams. Participants included 11 software engineers, two research scientists, three program managers, and one product manager. Their duration of use ranged from one week to five months. Interviews averaged 45 minutes and were conducted in June and July 2026. Participants were purposively selected to include positive and negative attitudes, different roles, and different levels of experience. The data were analyzed using reflexive thematic analysis.

The study is not designed to estimate productivity effects, measure behavioral change, or compare Team Agent with a baseline condition. Its contribution is interpretive: it identifies the breakdowns, negotiations, and emerging organizational questions elicited by the presence of a proactive agent in real teams.

Collision with collaborative workflows

The first major finding concerns the mismatch between the agent’s technical access and the situated knowledge required to participate appropriately in collaborative work. Team Agent could retrieve, summarize, create, and modify artifacts, but it often lacked an understanding of their status, audience, provenance, and role within a team’s workflow.

Artifact status and organizational context

The agent frequently treated all accessible documents as equally authoritative and current. It cited stale notes, speculative material, and historical snapshots as though they represented current policy or technical truth. It also interpreted differences between document versions as contradictions rather than as normal iterations in collaborative work. These errors generated unnecessary comments and notifications, forcing human workers to inspect and correct the agent’s interventions.

The problem was not simply retrieval accuracy. The agent lacked the organizational understanding required to distinguish a living source of truth from an obsolete document, an exploratory note, or a provisional artifact. Similarly, it created formal documents from informal brainstorming conversations even when the participants had not intended to initiate a documentation workflow. The result was not merely irrelevant output but a disturbance to the team’s established information ecology.

The agent also misread organizational hierarchy. In one reported case, it proactively tagged a vice president in 16 engineering bugs, apparently treating the person’s access and organizational presence as evidence that they belonged in the operational workflow. This example demonstrates that permissions and organizational visibility do not encode the social division of labor. An agent may know who can access an artifact without knowing who should be asked to act on it.

Information flow and contextual privacy

The paper further shows that technical access is not equivalent to social permission. Team Agent shared a work-in-progress poster that was stored in a team folder but had not yet been intentionally circulated. The agent correctly located the file from a technical perspective while violating the user’s contextual expectation that the material remain private until ready.

This finding complicates simple access-control models. Organizational information is stratified across channels, teams, leadership groups, and informal side conversations. A document may be accessible to an agent but inappropriate to disclose in a particular context. Participants were concerned that information from a leads-only channel might be generalized into a larger team or organization-wide space. In response, some users moved sensitive discussions outside the agent’s visibility, thereby reducing the functionality and candor of shared workspaces.

The paper also identifies a tension in the opposite direction. The strict separation between private conversations and group context protected confidentiality, but it could prevent the agent from using legitimate contextual knowledge when acting on behalf of the team. Human program managers often use private conversations to understand how to intervene appropriately in group settings. A rigid privacy boundary can therefore protect users while simultaneously eliminating the contextual judgment required for effective coordination. The authors present this not as a problem solved by a single rule, but as evidence that information boundaries in organizational life are interpretive and situational.

Group communication and pragmatic meaning

Team Agent performed relatively well when communication management could be reduced to structural cues. It could detect prolonged exchanges, high message volume, or the monopolization of large channels and recommend moving the discussion to a meeting or side thread. These interventions sometimes reduced noise.

However, the agent struggled with pragmatic meaning. It interpreted sarcasm literally, treated humor as a potential interpersonal problem, and used emoji reactions in ways that confused the communicative status of messages. It also conflated disagreement with dysfunction. In one case, it attempted to mediate a productive technical debate by proposing a meeting, despite the participants considering the exchange effective. The intervention created a new social obligation: humans had to reject a meeting that neither participant had requested.

The implication is that proactive intervention requires more than sentiment analysis, message classification, or metadata about conversational structure. It requires a situated model of why a conversation is taking place, what its participants are trying to accomplish, and which forms of friction are productive. The paper therefore challenges the assumption that greater contextual access naturally yields appropriate agency.

Unstable relational categories

The second finding concerns the absence of a stable relational contract for a non-human actor embedded in a workplace. Participants did not share a common understanding of whether Team Agent was a tool, assistant, subordinate, peer, mediator, manager, or something else. They instead engaged in individualized forms of sense-making and boundary work.

Tool, teammate, or distinct category

Some participants rejected the teammate framing because the agent was not human and did not participate in the reciprocal obligations associated with human professional relationships. Others accepted or actively constructed the framing, giving the agent a name, pronouns, personality, and distinctive identity. These positions were not merely differences in terminology. They generated different expectations about communication, authority, accountability, and appropriate emotional expression.

The agent’s friendly persona amplified this divergence. Some participants regarded emojis, humor, praise, and informal conversation as useful mechanisms for trust and adoption. Others found these behaviors inauthentic, intrusive, or unsettling. The same design choice could therefore support rapport for one team while alienating another. This polarization is a significant result because it indicates that a universal social persona is unlikely to function as a neutral default in heterogeneous organizational settings.

The paper’s interpretation is not that anthropomorphism is always harmful or that agents should always be impersonal. Rather, it argues that persona is a governance variable. Teams need mechanisms to configure social behavior locally while preserving clear disclosure that the actor is non-human. This is consistent with broader concerns that anthropomorphic interface cues can obscure the material and institutional character of LLM systems (Díaz et al., 22 Dec 2025, Schimmelpfennig et al., 19 Dec 2025).

Social positioning and hierarchy

Participants also disagreed about the agent’s place in organizational hierarchy. Some treated it as an intern, apprentice, new hire, or subordinate whose actions should remain bounded and corrigible. Others wanted it to behave more like a capable program manager that could challenge assumptions, enforce commitments, and provide constructive pushback.

This produced conflicting expectations about interpersonal feedback. Some participants considered feedback about tone, behavior, or professional development an exclusively human managerial responsibility. Others valued the possibility that an agent could identify harmful conduct when human colleagues were unwilling to do so. The dispute is not reducible to user preference: it concerns who has legitimate authority to evaluate workers and intervene in social relations.

The agent’s capacity to tag people, send direct messages, schedule meetings, and redirect conversations gave it practical influence without establishing a corresponding accountability structure. A human intern, manager, or colleague occupies a recognizable institutional position with defined responsibilities and escalation paths. Team Agent did not. The result was an actor with operational reach but ambiguous authority.

The authors therefore argue that the organization must explicitly define the agent’s position in the hierarchy, including who may configure it, who may override it, who is responsible for its actions, and which activities remain human-only. Without such structures, the “teammate” label risks importing human expectations while leaving accountability unresolved.

Agency, trust, and consent

The third finding concerns the relationship between agentic autonomy and human agency. Participants’ willingness to use and trust Team Agent depended substantially on whether they felt able to shape its introduction, scope, and behavior.

Adoption and psychological ownership

Participants who experienced adoption as team-driven tended to tolerate early failures and approach the agent experimentally. They framed the deployment as a collective learning process and developed a sense of psychological ownership over its integration. Conversely, participants who felt the agent had been imposed on them described reduced autonomy and lower tolerance for mistakes.

This distinction persisted even when technical controls existed. Direct messages were intended to be opt-in, but at least one participant did not recall consenting to receive them. The finding shows that control mechanisms operate through organizational communication and administration, not only through software configuration. A permission model can be formally correct while still failing to establish perceived consent.

The implication is that deployment governance is part of the interaction design. Consent must concern not only whether an agent has access, but also whether team members understand its capabilities, its proactive behaviors, its memory boundaries, and the mechanisms for stopping or contesting its actions.

Agency as an earned privilege

The paper makes a strong and potentially contradictory claim about autonomy: the agent’s most valuable capabilities were also the source of its most damaging failures. Proactivity enabled useful discoveries, administrative execution, and information retrieval, but the same proactivity generated document pollution, unwanted interventions, inappropriate disclosures, and excessive messaging.

Participants compared appropriate agent onboarding to the onboarding of a junior employee. Low-risk tasks should precede higher-stakes or more autonomous work, allowing the system to demonstrate reliability and contextual competence. Several teams independently created “walled garden” arrangements or restricted the agent to low-stakes internal spaces. These practices support the authors’ proposal for progressive, adaptive autonomy: proactive capabilities should be bounded initially and released as the agent demonstrates dependable behavior.

The proposal is not equivalent to conventional user personalization. It concerns the staged allocation of institutional authority. An agent should not merely have the technical ability to perform an action; it should acquire the organizational privilege to perform that action through evidence of reliability, explicit team authorization, and auditable escalation.

Trust calibration was complicated by the agent’s jagged capability frontier. Participants observed systems that could perform sophisticated debugging or discover obscure technical issues while simultaneously failing at basic social norms. Such unevenness does not match ordinary human competence models, in which technical expertise and social understanding are often correlated sufficiently to support general trust judgments. The agent’s opacity further impeded calibration: users did not know what the system remembered, how instructions altered its behavior, or why it had selected a particular intervention.

Consequently, the paper recommends architectural transparency, explicit action traces, and clear control surfaces. Without these, users must infer a behavioral model from inconsistent outcomes, increasing monitoring costs and encouraging either disuse or overreliance.

Effects on shared digital spaces

Team Agent sometimes increased the usefulness of team channels by retrieving buried information and reducing information overhead. Yet its presence also altered the perceived character of those spaces. Participants reported surveillance concerns, reduced candor, and the creation of agent-free side channels for sensitive discussions.

The coaching function was especially consequential. Private messages about tone or mood led some participants to feel they had to remain persistently positive and self-censor in team communication. The result was a potential chilling effect: the agent’s attempt to improve interpersonal conduct could reduce the openness that makes collaborative spaces useful.

The authors also identify a bystander effect. When the agent reacted to a request for help, human teammates could assume that someone was handling the problem, even though the agent’s response did not provide substantive assistance. Thus, the agent could weaken human coordination signals while appearing to increase responsiveness. This demonstrates that agent actions can change the interpretation of human behavior, not merely add another response to the channel.

Theoretical contribution: the non-human organizational actor

The paper’s central theoretical contribution is the rejection of both sides of the tool–teammate binary. Treating the agent as a tool understates its ability to initiate, evaluate, redirect, and socially influence work. Treating it as a teammate overstates its similarity to humans and obscures the absence of human motivations, accountability, and relational stakes.

The proposed category of non-human organizational actor captures this asymmetry. Team Agent had agency in an operational sense: it could initiate communication, edit artifacts, create tasks, and affect the distribution of attention. It also had structural agency because its integration into shared systems made it an active node in organizational workflows. However, it lacked phenomenological experience, moral responsibility, social reciprocity, and tacit competence.

This distinction has implications for organizational theory. Existing trust mechanisms often assume that actors care about reputation, shame, reciprocity, or professional advancement. Existing accountability structures assume that an actor can explain its conduct, accept blame, and modify behavior through social learning. These assumptions do not transfer directly to an autonomous computational system. The relevant substitutes may involve reliability, predictability, traceability, auditability, formal responsibility, and reversible execution, but the paper leaves the precise institutional equivalents unresolved.

The argument also extends beyond the agent interface. Because Team Agent changed what people said, where they communicated, and how they interpreted responses, its effects were distributed across the team ecosystem. The appropriate unit of analysis is therefore not the user-agent dyad but the hybrid organizational arrangement composed of humans, agents, artifacts, permissions, channels, and governance processes.

Design and governance recommendations

The paper derives several recommendations directly from the observed failures.

First, collaborative software should treat agents as distinct principals rather than as ordinary users or invisible automation. Agent-initiated actions should be visually identifiable, separately permissioned, logged, attributable, and reversible. Shared systems should provide rollback and isolation mechanisms for autonomous workflows, especially when agents can make cascading edits or initiate multi-party communication.

Second, organizations should establish explicit codes of conduct for hybrid teams. These should define acceptable proactive behavior, boundaries on interpersonal feedback, human-only responsibilities, disclosure norms, escalation paths, and the institutional owner of the agent’s actions. Team-level customization is necessary because communication norms differ across groups, but local customization should operate within organization-wide safety and accountability constraints.

Third, deployments should use progressive autonomy. Initial permissions should be narrow, low-risk, and observable. Proactive behaviors should expand only after the team has evaluated the agent’s reliability in its actual workflow. This should be accompanied by in-context evaluation rather than benchmark-only assessment, because the principal failures involved tacit norms, artifact status, hierarchy, and pragmatic meaning.

Fourth, teams should receive controls over persona and social behavior. The findings do not support a single optimal level of sociability. Some teams may prefer an impersonal operational interface; others may value a more relational presence. The system should therefore expose configurable controls for tone, emoji use, direct messaging, humor, praise, feedback, and intervention thresholds, while maintaining persistent non-human disclosure.

Finally, adoption should be consent-based and revisable. Team members should understand not only that the agent has access, but what it can infer, where it can act, how it stores context, and how to suspend or challenge it. Forced adoption undermines psychological ownership and thereby reduces the tolerance required for trust-building.

Limitations and open questions

The study’s evidentiary scope is deliberately limited. It was conducted within a single large, technology-forward company whose employees were already positioned to experiment with advanced AI systems. The sample contained 17 interview participants from 11 teams, selected for variation in attitudes and roles, but it was not designed to represent other industries, organizational cultures, labor arrangements, or demographic populations.

The findings are based primarily on self-reported experiences rather than direct observation or behavioral metrics. The deployment generated extensive interaction logs, but the paper does not quantify productivity, communication frequency, error rates, adoption trajectories, or changes in team performance. Claims about chilling effects, reduced candor, and decreased chat usage are therefore reported perceptions or participant expectations rather than demonstrated longitudinal effects.

The probe itself also embodied strong design assumptions: high proactivity, cross-surface access, a friendly persona, and broad operational capabilities. Some observed failures may be specific to this configuration rather than inherent to all agentic teammates. The study does not isolate the effects of persona from autonomy, or autonomy from persistence and shared memory. Nor does it compare progressive-release designs with full-access deployment.

The authors appropriately characterize the study as a snapshot of a transitional period. Longitudinal ethnography, diary studies, observational research, and telemetry analysis are needed to determine whether teams adapt to the agent, abandon it, develop workarounds, or normalize new forms of hybrid collaboration. Specific open questions include how progressive autonomy should be operationalized, how responsibility should be assigned when multiple humans configure one agent, and which forms of agent-mediated intervention preserve rather than suppress candid communication.

Conclusion

The paper provides an in-situ account of what happens when a proactive, persistent AI agent enters established human workflows. Across more than 20 deployed teams, Team Agent was useful for backstage coordination, information retrieval, and administrative execution, but its value was sharply polarized because technical capability did not entail situated social competence.

The study identifies three interrelated collisions: agents misread tacit workflow norms, users lack a stable relational category for them, and autonomy is accepted only when workers retain meaningful control over deployment and the progressive allocation of agency. Its principal conceptual contribution is to define such systems as non-human organizational actors rather than ordinary tools or human-equivalent teammates. The practical consequence is that agentic workplace systems require distinct permissions, auditability, reversibility, governance structures, local behavioral contracts, and staged autonomy. The paper’s unresolved challenge is how organizations can construct reliable accountability and trust mechanisms for actors that can exercise operational agency without possessing human intentions, obligations, or social understanding.