Fiona, the facilitator
Process-focused AI support
- Suggested collaboration strategies
- Gave time reminders
- Generated periodic summaries of team discussion
Introduction
From reactive tools to active teammates
Most AI systems still behave like tools. They wait for prompts, answer questions, and stay at the edge of collaboration.
But real teamwork is dynamic. Teams coordinate, split work, get stuck, recover, and rethink. I wanted to understand what happens when AI becomes part of that process.
In this project, I designed and studied two proactive AI roles for co-located collaborative problem-solving: a facilitator that supported coordination and reflection, and a peer that contributed ideas, hints, and follow-up support.
I built the system, designed the study environment, and ran a mixed-methods evaluation with 24 participants. The work was published at ACM CHI 2026.
Format
Technology probe study
Two proactive AI teammates were designed to test contrasting roles in the same collaborative task.
Setting
Co-located teamwork
Teams solved time-sensitive collaborative problems in a controlled yet socially dynamic environment.
Cross functional Collaboration with Honda Research Institute as a part of internship
This project was conducted in collaboration with Honda Research Institute US during my internship. The work brought together research questions about human-AI teamwork, product-facing questions about proactive behavior, and practical implementation decisions for the AI probes.
That collaboration shaped the study into something both rigorous and future-facing: not just whether proactive AI can work, but how different participation styles might translate into real collaborative systems.
My Role as Lead Researcher
End-to-end ownership of the research, prototype, and evaluation
Research framing
I defined the core UX question around proactive AI roles in teamwork and translated it into two comparable AI probes.
System and study design
I built the system, designed the collaborative study environment, and created the task flow needed for a within-subjects comparison.
Mixed-methods research
I ran the evaluation, captured behavioral and self-report evidence, and connected qualitative interpretation with team outcomes.
Design synthesis
I synthesized the findings into design principles for how collaborative AI should time, frame, and adapt its participation.
Project Objective
Understand what role proactive AI should play in human teamwork
The objective was not to build a final product. It was to create two clearly different AI participation styles and evaluate how each changed collaboration.
From a UX research perspective, the project focused on role design, interaction cost, team dynamics, and legibility. The core question was whether proactive AI should guide the team as a facilitator or join the work as a peer.
The Question
If AI becomes part of a team, what role should it play?
Most collaborative AI today is reactive. Someone in the group has to pause, prompt, interpret the result, and bring it back into the conversation. That interaction cost matters in fast, shared work.
So I asked a more specific UX question: should AI behave like a facilitator, helping the group coordinate and reflect, or like a peer, offering ideas and problem-solving support as an equal participant?
What I Designed
Two proactive AI probes for the same collaborative task
Ava, the peer
Task-focused AI support
- Offered proactive thoughts every 3 minutes
- Supported lightweight follow-up prompts
- Stayed available as an embedded chat partner
The goal was not to build a final product, but to instantiate two clearly different ways AI could participate in teamwork.
Study Design
Digital escape rooms as a testbed for collaborative problem-solving
I used digital escape rooms because they naturally create time pressure, distributed information, shared attention demands, and fast coordination cycles. That made them a strong setting for observing how proactive AI affected real team interaction.
24
participants in the mixed-methods evaluation
6
teams of four in a within-subjects design
3
conditions: no AI, peer AI, facilitator AI
The study combined behavioral observation, real-time transcripts, surveys, and focus group interviews. This made it possible to compare not only task outcomes, but also how teams felt, coordinated, and adapted over time.
What I Found
AI changed team dynamics, not just performance
Qualitative results snapshot
The study interface and probe behaviors helped surface how teams interpreted, used, ignored, or pushed back on proactive AI participation over time.
The peer agent felt helpful, but came at a cost
Ava sometimes gave timely hints, supported memory offloading, and opened space for exploration. But it also increased workload, disrupted flow, and sometimes pulled people into side conversations with the AI.
The facilitator agent was quieter, but more effective
Fiona often faded into the background, especially when teams felt the summaries were repetitive. Yet teams performed best overall in the facilitator condition. That revealed an important design tension between felt usefulness and actual team support.
Role design shaped trajectories across teams
Some teams relied on AI and later became more reflective. Others became dependent and then frustrated. This suggested that collaborative AI should adapt over time rather than hold a fixed role.
Design Takeaways
Four principles for collaborative AI design
Use proactivity carefully
Too much intervention increases cognitive load, even when the AI is technically helpful.
Support the team, not just the task
AI can disrupt communication even when its ideas are good, so social fit matters as much as output quality.
Make reasoning legible
People need enough rationale to evaluate confident AI suggestions and stay appropriately critical.
Adapt the role over time
Early-stage structure and later-stage idea support may require different AI behavior within the same workflow.
Impact
Published research with product-level implications
This work was published at ACM CHI 2026 and conducted in collaboration with Honda Research Institute US.
More importantly, it helped me articulate a product-level insight I now carry into my broader UX research work:
The best collaborative AI is not the AI that speaks the most. It is the AI that knows when, how, and why to participate.
Research takeaway