Microsoft Research and Aarhus Prototype an AI Workspace That Tracks Collaboration
Stund links conversation, edits and media to a shared canvas so AI can query and reshape meeting history, but the prototype has no user study and only minimal sandboxing.
Microsoft Research listed Stund on August 17 as an accepted UIST 2026 paper developed with Aarhus University. The work describes a web-based research prototype that records voice, video, chat, canvas interactions and content edits as linked activity traces, then lets AI query or transform those traces inside a shared workspace.
The project matters to people designing collaborative software because it moves beyond the usual meeting-assistant pattern. Instead of treating a transcript as a separate record, Stund connects what was said to the workspace artifacts and interactions around it. The sources describe a prototype and an accepted paper, not a Microsoft product, public preview or validated workplace system.
How the workspace uses activity traces
Stund divides its workspace into two connected layers. A shared infinite canvas holds sticky notes, shapes, video feeds and user-made widgets. A temporal layer stores interactions with that canvas alongside chat, audio and video traces. The sidebar chat can issue AI-mediated commands that query or manipulate both layers.
The prototype can turn recurring actions into processors and widgets. Script processors handle user-supplied logic, prompt processors send selected text traces to a language model, and vision processors analyze video frames. Widgets can surface derived information in the canvas, such as a count of elements or a summary of the previous five minutes of transcript activity.
The paper demonstrates the design through meeting scenarios rather than a field deployment. Its examples include replaying the conversation surrounding a spreadsheet edit, detecting that someone was absent and summarizing what they missed, and creating a processor that notices when participants shake a digital sticky note. These examples show technical range, not measured improvements in collaboration.
What is different from a transcript assistant
The research team argues that many meeting assistants work mainly from speech transcripts or chat logs while remaining detached from the changing workspace. Stund instead binds conversation, content and interaction history to the same canvas. That lets a later query refer to both an artifact and the activity that produced it, and it lets AI create a situational tool inside the workspace rather than only return a chat response.
That design is useful as a research direction because collaboration context often sits across several surfaces. It is also demanding: the system has to retain richer traces, connect them correctly and decide which information a processor may use. The paper does not provide a benchmark for trace accuracy, summary quality, processor reliability or the quality of generated widgets.
Surveillance and control remain central risks
The authors explicitly warn that the infrastructure supporting awareness can also support surveillance. Richer traces create more value for recall and coordination, but they also expose more about participants. In the prototype, people can disable their own recording, and processors currently operate only from the point when they are added. The authors note that retroactive processing is technically possible, which could let a future processor reinterpret earlier activity in a way participants did not anticipate.
Those boundaries make consent, retention, access and processor governance part of the core design rather than optional deployment details. Stund is especially sensitive because it brings together media, interaction events and editable workspace content instead of handling only a meeting transcript.
What the evidence does not establish
This is a functional research prototype, not evidence that the approach improves real work. The paper says the team has not demonstrated whether users understand the system or receive value in actual collaboration, and calls for empirical studies and participatory design.
The implementation also supports only one session per workspace, flushes session history nightly because traces consume disk space, and gives widgets and processors only minimal sandboxing. The authors say stronger security measures would be needed for production. Those limits are the clearest checkpoint for future work: user studies, durable multi-session history, clearer controls and production-grade isolation would provide stronger evidence than additional scenarios alone.
Status
Learning. Internal confidence is medium because Microsoft Research and the author copy establish the design and limitations, while no independent deployment, user study or product-readiness evaluation is included.
Sources
- Microsoft Research — English archive
- Microsoft Research — Stund publication page
- Aarhus University — Stund author copy
Update note: Last reviewed 2026-08-17. We will revise this post if the researchers publish user-study results, production safeguards or a material system update.
Sources
- Microsoft Research — English archive — official
- Microsoft Research — Stund publication page — official
- Aarhus University — Stund author copy — research
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.