Stund: Activity-Enriched Collaborative Workspaces Mediated by Artificial Intelligence

Authors: Marcel Borowski, Janus Bager Kristensen, Payod Panda, Sean Rintel, Niklas Elmqvist, Jens Emil Sloth Grønbæk, Clemens Nylandsted Klokmose
Venue: UIST 2026
Author's version (Aarhus University)

AThe paper

A systems paper presenting Stund, a working shared online workspace that records how a team works and lets AI build on that history. It is demonstrated through a scenario and has not yet been tested with users.

  1. A conceptThe authors define activity-enriched collaborative workspaces, which capture talk and actions as part of the workspace itself.
  2. A working systemStund was built as a shared online canvas with live video, sticky notes, documents and an AI chat.
  3. A scenarioThree remote colleagues write a grant proposal over a day, using the system's capabilities.
  4. Comparison and testingA table contrasts Stund with a basic chatbot given transcripts and logs. The authors also tested AI models and costs informally.

BThe system

  1. Record activityAs people talk and work, Stund stores audio, video, canvas actions and chats as time-stamped traces, and transcribes speech.
  2. Ask about the pastPeople select items and ask the AI chat about past activity, such as what a note meant, or have it arrange content.
  3. Derive new tracesOn request, the AI writes processors that detect events in the traces, such as a shaken note or someone leaving the camera.
  4. Build widgetsThe AI also writes widgets, small interactive tools on the canvas such as a note counter or a catch-up summary.
  5. Share and recombineWidgets can be shared and recombined through a repository.

The chat passes requests to specialised AI experts, which can use different models. There was no user study.

CWhat the paper proposes

In the scenariowhat the system did for three remote colleagues

  • In the scenario, the AI explained an unclear note by finding the talk that produced it.The authors know of no current system that links a note back to the conversation in which it was made.
  • In the scenario, a colleague had the AI build a detector for a team habit.During a coffee break, the AI made a detector for shaken sticky notes and a widget listing what it caught. The team defined this kind of awareness itself.
  • In the scenario, the AI built a catch-up tool for a colleague who had stepped away.The widget summarises the last few minutes of talk and activity, assembled from live recordings rather than after the meeting.
  • In the scenario, the team linked spreadsheet edits to the talk behind each edit.Combining shared widgets let the team play back the voice recording behind an edit, such as why a budget figure was lower.

Informal testing and open questionsthe authors' own use and the paper's discussion

  • In the authors' informal use, only a frontier model wrote widget code smoothly.A frontier model is among the most capable AI models. Smaller models handled most organising and summarising. Simple retrieval queries usually cost less than 0.05 USD each.
  • The paper argues the same activity traces could also enable surveillance.What matters is who controls what is captured, and why. Past actions could be reinterpreted by processors that do not yet exist.

DKey takeaway

Activity-enriched collaborative workspaces could let teams use AI to build their own tools from how work happened. The same records raise questions about who controls what is captured.

Made for this site with AI assistance from the verified research brief. It is not the paper's own figure.

Abstract

Time is a second-class citizen in most collaborative software: activity traces are fragmented across tools or entirely ephemeral. We introduce Stund, an activity-enriched collaborative workspace where all activities—voice, video, interactions, and content edits—leave traces that become part of the medium of the ongoing collaboration. Unlike traditional chatbots or meeting assistants, artificial intelligence (AI) is embedded in the medium as well, allowing users to directly leverage traces through it. We demonstrate this through scenarios such as: spontaneously creating an instrument during the meeting to play back the recorded conversation around a particular spreadsheet edit; automatically detecting when someone is absent to summarize what was missed when they return; or detecting and resurfacing when sticky notes have been shaken in excitement when ideas are presented. The entire workspace is queryable, transformable, and reified through AI. Stund introduces a new approach to malleable collaborative software, where activity traces and AI are an integral part of the workspace medium.

Research brief

By Marcel Borowski, Janus Bager Kristensen, Payod Panda, Sean Rintel, Niklas Elmqvist, Jens Emil Sloth Grønbæk, and Clemens Nylandsted Klokmose.

Takeaway: Collaborative software can become malleable by remembering how work happened. When every voice, video and edit leaves a trace that AI can query, history becomes working material. Teams could build tools mid-meeting, like replaying talk behind a spreadsheet edit or catching up a returning colleague.

Why this matters

What was known
Most software treats activity history as an afterthought, and AI meeting tools work on single modalities such as transcripts, detached from what is happening in the shared workspace. Existing AI-assisted meeting systems-commercial or research-operate on single modalities or treat AI as a conversational layer detached fromworkspace state. (§2.2, p. 3)
What this paper adds
Rather than adding another AI assistant beside the workspace, we make the workspace's whole recorded history part of the medium, building on the theory of interaction substrates. We demonstrate this with a working system and a five-part scenario. This accumulation of content, context, and bespoke capability growing together in AI-mediated substrates is, we argue, what distinguishes AECWs from existing collaborative tools (§5.6, p. 8)
Who it matters to
Researchers and designers of collaborative, malleable and AI-enabled software, and teams who meet and work together remotely. The paper also flags the privacy and surveillance stakes for participants. We see Stund as an experimental platform that enables research to explore these tensions empirically (§7.3, p. 11)

Research questions

Method and data

Design
Systems contribution: a conceptual model, a functional implementation, and a five-part use scenario demonstrating its capabilities (no user study).
Prototype
Stund, a web-based shared infinite canvas (built on Webstrates, Codestrates and the tldraw canvas) holding live video feeds, sticky notes, documents and shapes, with an AI chat sidebar. As people talk and work, Stund records their audio, video, canvas actions and chats as time-stamped traces on a Java server and transcribes speech automatically. Users can select items and ask the chat about past activity ('What did this mean?') or have it arrange content. They can also have it write 'processors' that derive new traces (for example, detecting a shaken note or someone leaving the camera) and 'widgets' (small interactive tools on the canvas such as a note counter or a catch-up summary). Widgets can be shared and recombined through a repository. The chat passes requests to specialised AI 'experts', which can use different models, such as frontier models for writing code. There was no comparison in a study. A table contrasts Stund's capabilities with those of a basic chatbot given transcripts and interaction logs.
Annotated screenshot of Stund: a canvas with a video feed, spreadsheet widget, document, sticky notes and search widget, a chat sidebar with tabs and ability toggles, and a right-hand bar of repositories and media controls.
Figure 2: Stund Interface. The interface is structured around a shared canvas and a chat sidebar. Custom elements, including documents, widgets, and video feeds, can be placed along regular elements (here notes, images, arrows, shared cursors) in the canvas (paper cover image by [31]). The chat supports multiple tabs-local or shared-where shared tabs show presence indicators and can be collaboratively authored. Chat messages can include elements or images in their context. Here, the chat summarizes the contents of the spreadsheet widget, with its reply displayed in the context of that element. The chat has toggles to activate the write-access abilities: canvas, widget, and processor. A sidebar on the right gives access to control panels for model selection, repository dialogues, and buttons to start or stop media streams. From the paper, Fig. 2, p. 5.
Setting
Remote collaboration on a shared online canvas with audio and video. The scenario follows a team writing a grant proposal over a day.
Participants / data
Scenario of three remote collaborators (Anna, Bo and Carl) writing a grant proposal over a day. The authors also used Stund in their own meetings and other domains.
Analysis
Demonstration of capabilities through the scenario, a capabilities overview comparing Stund's components with a basic LLM chatbot baseline (Table 1), and informal testing of models and costs.

Findings

Questions this paper answers

Can AI explain why something is on a shared whiteboard?
It can, if the workspace records what was said as well as what was edited. AI can then find when a note was made, retrieve the talk around that moment and explain it. A chatbot given only transcripts might answer too, but without the link back to the note. (See §5.1, p. 6.)
Can people build their own meeting tools during a meeting with AI?
In principle. In a demonstration scenario, collaborators ask the workspace's AI chat to write small tools on the shared canvas. Examples are a note counter, a detector for a team habit and a catch-up summary, which they then reuse and combine. Whether real teams can do this is untested. (See §5.2, p. 6; §5.3, p. 7; §5.4, p. 7.)
Which AI models are needed to build tools inside a collaborative workspace?
In informal testing, smaller models handled most organising and summarising, but writing widget code ran smoothly only with a frontier model. An open-source model could get similar results with more iteration. (See §7.2, p. 10.)
Is recording everything in a shared workspace a privacy risk?
Yes. The activity history that makes past work useful can also enable surveillance, and past actions could be reinterpreted by analyses that do not yet exist. Monitoring that workers control may constrain them less than imposed monitoring. Letting individuals switch off their own recording may partly help. (See §7.3, pp. 10–11.)

Quantitative results

MeasureResultWhere
Cost of simple retrieval queries with Claude Sonnet 4.6usually less than 0.05 USD each§7.2, p. 10

Contributions and implications

Limitations

›

Concepts

Activity-enriched collaborative workspace (AECW)
Our concept, inspired by Hollan's vision of activity-enriched computing: computer-mediated collaboration spaces where heterogeneous activity traces from both communication and interaction are captured as part of the workspace (§1, p. 2)
Material and temporal substrates
The two substrates of an AECW, after Mackay and Beaudouin-Lafon. The difference between them: State in the material substrate is mutable, i.e., commands can manipulate it, while activity traces are immutable, making the temporal substrate append-only. (§3, p. 4)
Reification
Turning commands and effects into reusable instruments. For example, the command to create four sticky notes in different colors can be reified into a "create-four-sticky-notes" instrument (§3.1, p. 4)
AI-as-medium
AI doing clerical, custodial and creational work inside the workspace, rather than acting as a separate assistant, tool or teammate: This approach represents an "AI-as-medium" view of integrating generative AI into software, where AI is part of the computational fabric of the workspace (§1, p. 2)

Figures

Left: timeline of activity trace tracks over time (video, person present, audio, transcripts, topics, AI chats, element moved, shaken and edited, spreadsheet edited). Right: the shared canvas with notes, video feeds, widgets and chat.
Figure 1: Activity-Enriched Collaborative Workspaces. A workspace consists of a material substrate where collaborative activity happens, and a temporal substrate where this activity leaves traces over time. In Stund, the material substrate is a collaborative canvas that includes notes, video feeds, and user-made interactive widgets such as a custom spreadsheet widget for a grant budget or a situational widget showing a list of notes that have recently been shaken to indicate importance. Audio/video feeds, canvas interactions, widgets, and the chat all leave traces in the temporal substrate. Activity traces can be derived from others. Transcripts are derived from audio feeds, and a user-made processor may derive shake traces from a note being moved back and forth in rapid succession. In a sidebar chat, users can issue commands to query activity traces ("Why is this here?"), manipulate the canvas ("Put to-dos in red notes beside our notes."), or create processors ("Detect when a note is shaken.") and widgets ("Show a live-updating list of recently shaken notes."). From the paper, Fig. 1, p. 1.
Annotated screenshot of Stund: a canvas with a video feed, spreadsheet widget, document, sticky notes and search widget, a chat sidebar with tabs and ability toggles, and a right-hand bar of repositories and media controls.
Figure 2: Stund Interface. The interface is structured around a shared canvas and a chat sidebar. Custom elements, including documents, widgets, and video feeds, can be placed along regular elements (here notes, images, arrows, shared cursors) in the canvas (paper cover image by [31]). The chat supports multiple tabs-local or shared-where shared tabs show presence indicators and can be collaboratively authored. Chat messages can include elements or images in their context. Here, the chat summarizes the contents of the spreadsheet widget, with its reply displayed in the context of that element. The chat has toggles to activate the write-access abilities: canvas, widget, and processor. A sidebar on the right gives access to control panels for model selection, repository dialogues, and buttons to start or stop media streams. From the paper, Fig. 2, p. 5.
Architecture diagram: client recorders send traces via WebRTC to the server's ingest and processing modules, and a filtered client model streams back via WebSocket to the activity session manager, used by the AI chat experts and widgets.
Figure 4: Stund's Client-Server Architecture. Active code components are colored orange and passive data components are colored green. The client collects data with recorders and sends them via WebRTC to the ingest module, which unpacks and stores the data. The processing module creates a raw activity model, which is filtered by the model module. The client model is streamed to the activity session manager of the client via WebSocket. From there it is used by the AI chat and widgets. The experts of the AI chat have read/write access to elements and widgets on the canvas and to custom processor schemas. The latter can be used to add new processors to a session via WebSocket calls. From the paper, Fig. 4, p. 9.

How to cite

Marcel Borowski, Janus Bager Kristensen, Payod Panda, Sean Rintel, Niklas Elmqvist, Jens Emil Sloth Grønbæk, and Clemens Nylandsted Klokmose. 2026. Stund: Activity-Enriched Collaborative Workspaces Mediated by Artificial Intelligence. UIST 2026.

This brief was prepared with AI assistance and published by Sean Rintel. It is not the authors' published text. Page numbers refer to this PDF.

BibTeX

@inproceedings{borowski2026stund,
author = {Borowski, Marcel and Kristensen, Janus Bager and Panda, Payod and Rintel, Sean and Elmqvist, Niklas and Grønbæk, Jens Emil Sloth and Klokmose, Clemens Nylandsted},
title = {Stund: Activity-Enriched Collaborative Workspaces Mediated by Artificial Intelligence},
booktitle = {UIST 2026},
year = {2026},
publisher = {ACM},
url = {https://seanrintel.com/publications/stund-activity-enriched-collaborative-workspaces-mediated-by-artificial-intelligence.html},
}