Source-linked AI summary
Beyond the Personal Assistant: How Expectations for Enterprise AI in Teamwork Diverged as Generative AI Took Shape, 2023-2025
Qing Xiao, Xinlan Emily Hu, Mark E. Whiting, Arvind Karunakaran, Hong Shen, Hancheng Cao
TL;DR
As technologies evolve, users’ articulated expectations may change, making it important to understand what becomes of earlier design possibilities. This study compares practitioners’ team-AI expectations across two interview phases; by 2025, expectations had diverged, while personal assistants were routine and team-focused systems had not arrived.
Problem
Users’ articulated expectations are shaped by what technologies can do and what people have experienced, so they may change as technologies evolve.
Source: §1 · p. 1Method
The study compares interviews with practitioners in one early-adopting software organization, returning to the same participants in 2025 to ask about AI for teamwork.
Source: §3.1 · p. 4, §8 · p. 13Results
By 2025, no team-focused system had entered participants’ work, personal assistants were routine, and earlier expectations had been deferred, rejected, or no longer stated.
Source: §8 · p. 13Takeaways & Limitations
Following expectations over time reveals design possibilities that remain open, become objectionable, or disappear from later accounts.
Source: §1 · p. 2The two-phase study in one organization establishes that these expectation trajectories occur, not how they are distributed or how quickly expectations evolve elsewhere.
Source: §7 · p. 13
Abstract
from arXiv · showhide
HCI often draws on users' articulated needs and expectations to explore design opportunities for emerging technologies. Yet, these accounts are shaped by how technologies take form over time. We examine this dynamic through a two-phase interview study of enterprise AI in a project-based software development organization. We interviewed 15 practitioners in 2023 and re-interviewed 10 in 2025. In 2023, participants expected enterprise AI built around teams: systems integrating signals across workplace tools, detecting coordination problems, and intervening in team processes. By 2025, no such team-focused systems had entered their work, while personal AI assistants were prevalent. These expectations diverged from one another: some expectations were deferred, some rejected as undesirable or infeasible, and some no longer articulated as expectations they had held. These trajectories show how design possibilities can remain open-ended, narrow, or disappear as technologies take shape, with implications for designing enterprise AI and interpreting present-day user accounts.
1 Introduction
The study follows how expectations for enterprise AI in teamwork changed as AI took shape at work. Participants first imagined team-oriented systems, but by 2025 their expectations had diverged while personal assistants had become routine.
1 Introduction
- The study compares the same practitioners’ accounts of team AI expectations in 2023 and 2025, when generative AI use at work had changed substantially.The researchers interviewed 15 practitioners in 2023 and re-interviewed 10 in 2025.Source: §1 · p. 2
- The study asks what workers expected AI to do for their teams and how those expectations changed as workplace AI evolved.The focus is timely because enterprise AI includes both personal assistants and systems proposed to support coordination across shared workplace tools.Source: §1 · pp. 1–2
- In 2023, participants expected AI to read across workplace tools, detect coordination problems, and intervene in team processes, subject to limits on data collection and human judgment.They wanted assessments to reach the person concerned before a manager, visible data collection, and AI that informed rather than replaced human judgment.Source: §1 · p. 2
- By 2025, no team-focused system had entered participants’ work; four deferred, three rejected, and three dropped their 2023 expectations as personal assistants became routine.The authors describe changes in articulated expectations, not directly observed underlying beliefs.Source: §1 · p. 2
- Following expectations over time distinguishes possibilities that remain open, become objectionable, or disappear from later accounts.The authors argue that present-day accounts alone cannot reveal these different histories.Source: §1 · p. 2
2 Background and Related Work
Prior work examines individual workplace AI use, AI systems supporting teamwork, and how expectations change with experience. This study connects these areas by following the same practitioners’ team-AI expectations over time.
2.1 AI at Work: Individual Use in Interdependent Settings
- Research on individual AI use also considers organizational relationships and collective effects, while this study follows team expectations alongside workers’ growing experience with AI.Prior studies cited here report shifts toward individual coding effort, changes in coordination time [82], and AI use disconnected from collaborative workflows.Source: §2.1 · pp. 2–3
2.2 AI for Teamwork and Coordination
- CSCW research and newer AI systems address coordination, awareness, information sharing, interaction, and direct contributions to team tasks.The study complements system- and use-centered work by examining practitioners’ expectations for team AI and how they changed.Source: §2.2 · p. 3
2.3 User Expectations and Design Opportunities Over Time
- Although HCI elicits expectations to explore design possibilities, less work follows expectations voiced during an emerging technology’s development as users’ experiences change.The authors build on longitudinal studies by asking the same practitioners in 2023 and 2025 what they expected AI to do for their teams.Source: §2.3 · p. 3
3 Method
The study uses two interview phases in one early-adopting, project-based software organization to compare participants’ expectations for AI in teamwork. Researchers analyze each phase separately before pairing returning participants’ accounts to trace change.
3 Method
- Fifteen practitioners were interviewed in 2023 and 10 were re-interviewed in 2025 using the same question about what AI should do for their teams.The study compares accounts from the same people, while acknowledging that it did not continuously observe them between phases.Source: §3 · p. 3
3.1 Study Site and Participants
- The organization’s distributed, project-based teams coordinate with changing collaborators, and early internal AI deployment made it a useful setting for studying expectations as AI entered work.The organization deployed GitHub Copilot and GPT-based coding assistants internally in early 2023.Source: §3.1 · p. 4
3.2 Data Collection: Two Interview Phases
- The 2023 protocol elicited teamwork frictions, participants’ understandings of AI, and expectations about automation, human judgment, and acceptable intervention; the 2025 protocol revisited practice and expectations.Before comparing accounts, researchers reminded participants of what they had said in 2023 so they could affirm, revise, or disown those accounts.Source: §3.2 · pp. 4–5
3.3 Analysis
- Researchers coded each phase independently, then paired returning participants’ accounts to identify what persisted, changed, appeared, or was absent.The 2025 codebook was developed without reference to the 2023 one; findings describe articulated expectations, not directly observed beliefs, and an absent expectation’s reason cannot be determined from interviews alone.Source: §3.3 · pp. 5–6
4 Findings: Expectations of Enterprise AI Built Around Teams in 2023 (Phase 1)
Participants expected enterprise AI to integrate workplace records and support team coordination, but placed limits on how such systems should assess and intervene. Their expectations addressed late detection, fragmented understanding, and coordination work that fell to individuals.
4 Findings: Expectations of Enterprise AI Built Around Teams in 2023 (Phase 1)
- Participants expected an enterprise layer across Jira, Slack, GitHub, and project plans to convert existing work traces into early information about team functioning.Source: §4 · p. 6
4.1 The Problems These Expectations Were Anchored To
- In distributed project work, disengagement surfaced late and structured tools left shared understanding incomplete, while informal repair fell to experienced contributors.Source: §4.1 · p. 6
4.2 What Participants Expected AI to Do in Future Teamwork
- The authors organize expectations as sensing, diagnosis, and mediation; most participants described more than one function.Source: §4.2 · p. 6
4.2.1 Sensing: Expecting AI to Detect When a Team Departs from Its Own Baseline.
- Sensing meant comparing team records with a learned historical baseline, especially silence, stalled artifact movement, or schedule slippage, early enough to correct course.Source: §4.2.1 · p. 6
4.2.2 Diagnosis: Expecting AI to Combine Signals That No Single Tool Held.
- Diagnosis would combine signals across systems to interpret context, assess team or project health, and inform staffing before a team formed.Source: §4.2.2 · pp. 6–7
4.2.3 Mediation: Expecting AI to Act on Relationships Between Team Members.
- Mediation ranged from milestone-triggered huddles and time-zone matching to interaction feedback, sentiment tracking, and team-composition visibility from isolated conversations.Source: §4.2.3 · p. 7
4.3 The Limits Participants Placed on Their Team-Level Expectations
- Participants wanted disclosed assessments shown first to individuals as human-interpreted prompts, not verdicts, and doubted conversational assistants could deliver the cross-tool team system.Source: §4.3 · p. 7
5 Findings: What Became of Those Expectations in 2025 (Phase 2)
By 2025, participants described personal AI assistants for individual tasks, but no team-focused systems of the kind they had expected in 2023. Their earlier team-level expectations were deferred, rejected, or no longer articulated, while some related concerns appeared as expectations of colleagues.
5 Findings: What Became of Those Expectations in 2025 (Phase 2)
- No team-level system arrived; participants instead described personal AI for documentation, coding, research, and task-starting, while meeting notes ran alongside rather than within collaboration.Source: §5 · pp. 7–8
5.1 Three Trajectories: How the 2023 Expectations of AI for Teamwork Diverged by 2025
- Four participants deferred their expectations, three rejected them, and three no longer stated them in 2025.Source: §5.1 · p. 8
5.1.1 Deferred: Keeping the Expectation and Making It More Spe-
- Deferred expectations became more specific—from team-level intent alignment and adaptive work allocation to persistent feedback and automated integration—amid organizational and architectural obstacles.Source: §5.1.1 · p. 8
5.1.2 Rejected: Remembering the Expectation and Turning Against It.
- Rejected expectations reflected doubts that teamwork could be represented reliably and concern that supportive feedback might become managerial surveillance.Source: §5.1.2 · pp. 8–10
5.1.3 Dropped: No Longer Stating the Expectation.
- For participants who dropped expectations, teamwork no longer seemed a relevant AI domain or the earlier expectation simply went unmentioned; interview data cannot establish why it disappeared.Source: §5.1.3 · p. 10
5.2 From Team-Level AI Expectations to Colleague Norms
- Participants did not describe the missing team systems as a loss, even as personal AI exceeded expectations; concerns about pace and disclosure appeared as colleague norms, not literal transfers of requirements.Source: §5.2 · p. 10
6 Discussion
Expectations for AI in teamwork acquired different histories: some persisted, some were rejected, and some disappeared from later accounts. Comparing those histories gives designers evidence that a present-day account alone cannot provide.
6.1 Expectations Have Histories, Not Just Current States
- Similar present-day positions can conceal whether an expectation was rejected, dropped from later accounts, or never considered.Longitudinal records preserve what persisted, became objectionable, or fell out of view.Source: §6.1 · p. 11, §6 · p. 11
6.2 The Design Space Users Articulate Is Historically Contingent
- Accounts from 2023 and 2025 came from different technological worlds; neither is ground truth, and the data do not show that personal assistants caused the divergent trajectories.Each account made some possibilities salient while leaving others less visible.Source: §6.2 · p. 11
6.3 Different Expectation Trajectories Offer Different Design Evidence
- Deferred expectations identify opportunities to revisit, rejected expectations reveal constraints, and dropped expectations preserve alternatives for possible reintroduction.Longitudinal research preserves differences across moments rather than converging on an increasingly accurate requirements document.Source: §6.3 · pp. 11–12
- Longitudinal elicitation complements rapid prototyping by preserving earlier alternatives that testing current concepts cannot recover without a record.The design concepts illustrate how expectation trajectories can shape responses; they are not an agenda for the studied organization.Source: §6.3 · p. 12, §6.4 · p. 12
- Two phases in one organization do not show how quickly expectations evolve elsewhere, how often trajectories occur, or which expectations designers should act on.Repeated qualitative work carries costs, so the authors do not advocate longitudinal panels for every design project.Source: §6.3 · p. 12
6.4 Designing "AI for Teamwork" From Understanding Expectation Trajectories
- Project Memory refines deferred expectations by retaining context across project activity, reflecting later emphasis on persistence, shared context, and coordination practices.The concept builds on earlier ideas for AI that understands team activity across tools.Source: §6.4 · p. 12
- Evidence Lens redesigns rejected expectations by integrating workplace information while avoiding automated team-health verdicts and personnel judgments.It responds to concerns about validity, interpretation, and oversight while preserving information synthesis.Source: §6.4 · p. 12
- Reconnect Prompt reintroduces a dropped possibility as a probe, testing whether AI has no place in teamwork or should simply not judge teams.The 2025 accounts did not settle that distinction.Source: §6.4 · p. 12
7 Limitations and Future Work
The study establishes that expectation trajectories can occur and shows how they cohere, but its sample, single-organization setting, interview evidence, and two time points limit broader claims. Further work could combine interviews with behavioral evidence and follow participants over a longer period.
7 Limitations and Future Work
- The 15- and 10-person phases, with trajectory groups of four, three, and three, show that the patterns occur but not how prevalent they are.Estimating prevalence would require surveys or archival analyses of changing worker expectations.Source: §7 · p. 13
- Because this is one early-adopting, project-based software organization, the findings support claims about what can happen and how, not how often across settings.Expectations may follow different paths in co-located teams, constrained sectors, or organizations that adopted AI later.Source: §7 · p. 13
- Interview accounts show what participants said they expected, not private belief or team behavior, and revisiting earlier statements may prompt reconciliation.For dropped expectations, interviews cannot distinguish forgetting from quiet disavowal.Source: §7 · p. 13
- Communication logs or observation combined with repeat interviews could test whether the described relocation of pace and disclosure into peer norms appears in team interactions.Source: §7 · p. 13
- Two time points bracket change but cannot establish when or why trajectories occurred or what happens next.A third phase could examine whether dropped expectations remain dropped and whether reopened expectations resemble those from 2023.Source: §7 · p. 13
8 Conclusion
In this organization, team-focused enterprise AI had not reached participants by 2025, while personal assistants had become routine. Their earlier team-AI expectations diverged, and each trajectory offers distinct design material.
8 Conclusion
- By 2025, no team-focused system had reached participants, and their 2023 expectations had been deferred, rejected, or no longer stated.Personal AI assistants had become routine.Source: §8 · p. 13
- Deferred expectations mark opportunities to refine, rejected expectations mark constraints to redesign around, and dropped expectations preserve alternatives that earlier records may make possible to reintroduce.Source: §8 · p. 13
- A present-day account alone cannot distinguish whether an unvoiced expectation is still wanted, has been rejected, or is no longer thought of.In rapidly evolving domains, earlier accounts document how the design space users articulate has changed.Source: §8 · p. 13