TL;DR for operators
More direct human control does not necessarily produce a preferred human-AI workflow. In a laboratory assembly study, 77% of participants ultimately preferred robot-initiated tool delivery, even though human-initiated delivery was associated with higher perceived control among those same participants.
Most also preferred fewer confirmations and shorter explanations. Yet these preferences do not establish that less interaction is universally better, or that automation improves productivity.
The paper’s principal contribution is a design framework connecting worker needs, AI capabilities, task requirements, organisational conditions, and societal obligations. For operational teams, the practical implication is to make appropriate interactions configurable, preserve meaningful intervention rights, and evaluate trust, workload, and control alongside efficiency. The framework is a structured design approach, not a validated formula for optimal collaboration.
The workflow workers preferred gave them less direct control
Consider a worker assembling a mechanical component while a robot delivers tools. The robot can wait until the worker requests each tool, or identify the assembly stage and deliver the next tool automatically. It can ask permission before acting, explain its intended action, or proceed with minimal communication.
Each configuration offers a different balance between worker intervention and uninterrupted task execution.
A University of Auckland study described in Working with AI: A Design Framework for Human-AI Collaboration1 illustrates why this balance is difficult.
Thirty participants completed six assembly runs each, producing 180 observations. At the end, 23 participants—77% of the sample—preferred robot-led tool delivery.
Yet among those 23 participants, mean perceived control during earlier researcher-assigned runs was 5.58 under human-led initiation versus 4.37 under robot-led initiation, measured on a seven-point scale (Figure 9).
The distinction matters. Participants could prefer automatic assistance while feeling more directly in control when they initiated each handover themselves.
For an organisation designing AI-assisted workflows, perceived control and preferred interaction are therefore different evaluation targets. Increasing one does not necessarily improve the other.
Fewer interactions were preferred, but not universally validated
The laboratory system allowed three principal configuration choices: who initiated tool delivery, whether the robot requested confirmation, and how much explanation accompanied its actions.
After repeated use, participants generally favoured less interaction overhead:
- 77% preferred robot-led tool initiation.
- 87% preferred confirmation disabled.
- 29 of 30 preferred brief or no explanation.
- 77% changed their selected configuration at least once across the three self-selected runs.
These are descriptive preference results, reported in Section 4.1.5 and Figure 12. They are not independent tests establishing the optimal setting for each interaction feature.
The preference pattern is consistent with a plausible mechanism: when a task is predictable, repeated requests and explanations can impose cognitive and communication costs without providing equally valuable information.
The study also reports mean mental-demand scores of 2.90 under human-led initiation versus 2.22 under robot-led initiation. The authors suggest that remembering when to request the next tool adds mental effort.
However, the value of confirmation changes when an AI action is difficult to reverse, when mistakes carry substantial consequences, or when the system is uncertain. A worker’s preference for fewer routine interruptions should not be interpreted as permission to remove necessary safeguards.
Crucially, the laboratory system preserved human override in every configuration. The worker could reject, change, or stop robot actions.
This separates control over every routine initiation from authority to intervene when necessary. A system can reduce repetitive interaction without transferring final operational authority to the AI.
Collaboration design begins outside the interface
The paper’s larger contribution is a socio-technical framework: collaboration quality depends on more than whether the AI performs its technical task correctly.
It identifies five interacting contextual domains:
| Domain | What the designer must understand |
|---|---|
| Human | Skills, experience, motivation, confidence, cognitive capacity, and autonomy needs |
| AI | Accuracy, reliability, transparency, adaptability, safety, and privacy protections |
| Task | Complexity, workload, time pressure, observability, uncertainty, and risk |
| Organisation | Training, management support, incentives, resources, and maintenance |
| Society | Regulation, professional standards, rights, public expectations, and accountability |
These domains are contextual preconditions, not a readiness score or a checklist that produces a universal design.
The framework then distinguishes three levels of desired outcomes: human experience and well-being; collaboration quality and task performance; and longer-term organisational and societal consequences.
That distinction changes how an AI implementation should be assessed. Faster task completion may coexist with excessive reliance, diminished autonomy, or unresolved accountability.
To translate these concerns into requirements, the authors combine human needs, human-AI teamwork, and responsible deployment. Their conceptual foundations include self-determination theory, teamwork research, and OECD responsible AI principles.
Six design principles follow: task and role design, transparency, communication, human-centred design, responsible design, and sustainability design.
The implementation process is iterative. Designers first understand the use case, identify desired outcomes and requirements, apply the design principles, implement concrete decisions, and evaluate observed results. Outcomes inform the next design iteration rather than closing the process permanently (Section 3.2, Figure 6).
The assembly system demonstrates this translation. Task allocation became a fixed division between human assembly and robot tool delivery. Transparency became configurable explanations. Communication became voice commands, speech output, and on-screen feedback. Human-centred design became onboarding and adjustable settings. Responsible design included override and supervision.
Sustainability design, however, was not directly tested.
The framework’s strength lies in connecting these choices to explicit requirements. Its causal effectiveness over alternative design frameworks remains unestablished.
A wrong tool changed trust and completion time
The laboratory also examined what happened when assistance failed.
During Run 5, the robot deliberately delivered an incorrect tool. Mean self-reported trust fell from 5.87 in Run 4 to 4.60 in Run 5, then returned to 5.83 in Run 6 (Figure 8).
Completion time followed a related pattern: 123 seconds in Run 4, 157 seconds in the error run, and 118 seconds in Run 6.
This was an illustrative failure intervention, not a robustness experiment with independently randomised error timing. It shows that participants’ reported trust and task performance were responsive to a visible failure in this particular setting. It does not establish how trust would recover after repeated errors or during long-term deployment.
The study also reports a decline in average completion time from 174 seconds in Run 1 to 118 seconds in Run 6, approximately 32%.
That improvement cannot be attributed to configurability. Researcher-assigned settings were used in Runs 1–3, while participants selected settings in Runs 4–6. Increasing task familiarity, changing configurations, and run order therefore overlap.
Even the reported difference between assigned and self-selected runs—150 versus 132 seconds on average—cannot isolate the effect of user choice. The authors explicitly recognise this limitation.
Reported reliance requires similar care. Participants depended on the robot for tool delivery in every configuration, so reliance was partly imposed by the task rather than freely chosen.
These distinctions prevent a descriptive laboratory demonstration from becoming an unsupported productivity or resilience claim.
What an operational deployment should test
Cognaptus inference: For a product owner deploying an AI assistant into a recurring business workflow, the appropriate interaction policy should depend on task risk, user experience, and the consequences of an incorrect action.
For low-risk, reversible actions, experienced users may benefit from fewer routine confirmations and shorter explanations. For consequential or difficult-to-reverse actions, explicit confirmation and accessible explanations may remain necessary.
The framework suggests a practical design sequence.
First, specify who initiates actions, who verifies them, and who can intervene. Then decide which interactions can be adjusted without undermining safety, privacy, or accountability.
Next, evaluate more than successful task completion. Track processing time, error rates, recovery behaviour, perceived control, cognitive demand, and whether users rely on the system appropriately. Trust should be calibrated to demonstrated reliability, rather than simply maximised.
Finally, reassess settings as users gain experience or tasks change. The observation that 77% of participants modified configurations provides a reason to investigate adaptable interfaces, but not proof that adaptation independently improves performance.
For procurement, this implies asking vendors about configurable interaction policies, intervention mechanisms, failure recovery, and maintenance responsibilities—not merely model accuracy or automation coverage.
These are proposed operational applications of the framework, not enterprise outcomes demonstrated by the assembly experiment.
Where the laboratory evidence stops
The empirical study involved only 30 participants, predominantly engineering students, performing one short-duration, structured assembly task under laboratory supervision. Most lacked prior collaborative-robot experience.
The evidence consists mainly of descriptive averages, participant preferences, brief self-report measures, and system logs. The study reports no formal causal identification strategy or statistical uncertainty analysis. The deliberately introduced error occurred at a fixed run position, and self-selection was introduced only after assigned configurations.
More broadly, the study did not independently evaluate employee retention, financial return, workforce displacement, long-term safety, organisational adoption, or societal outcomes. Responsible deployment requirements were treated as baseline conditions rather than tested outcomes.
A stronger enterprise evaluation would compare interaction configurations at comparable levels of worker experience, vary configuration order where feasible, observe failure recovery, and measure outcomes over a longer period.
The design objective is appropriate control
The paper offers a useful correction to a common deployment assumption: more human initiation, more confirmation, and more explanation do not automatically produce better collaboration.
The laboratory participants often preferred AI assistance that reduced routine interaction, while still retaining the ability to intervene. That preference is context-dependent, and the evidence does not identify an optimal configuration beyond the studied task.
For organisations, the enduring contribution is the framework’s insistence that collaboration be designed and evaluated as a complete working arrangement. The relevant question is how much automation, information, and human authority a particular user needs for a particular task under particular risks.
That question requires explicit design choices, repeated measurement, and evidence from the environment where the system will actually operate.
Cognaptus: Automate the Present, Incubate the Future.
-
Yuqian Lu and Regina Lee and Rui Zhou and Lixin Jiang and Andrew McDaid and Amy Lawrence (2026). Working with AI: A Design Framework for Human-AI Collaboration. arXiv:2609.31793. https://arxiv.org/abs/2609.31793 ↩︎