Research note · Product design
Designing movement tools people actually use: lessons from a chat-based product
A movement tool only works if it is used, and most are not. MoveBreak, our assistant for short breaks inside Microsoft Teams and Slack, has supported 2,000+ employees and 30,000+ wellbeing moments since early 2026 (internal data). It contains no pose model and no generative AI. This note sets out the six design decisions behind it, and what they changed in how the lab works.
Last updated: By the Movelytics team
01Starting point
The problem is not the exercise
When a team sets out to fight sedentary work, it usually starts with content: which stretches, which videos, which programme. Content is the easy part. Physiotherapists and osteopaths know what a body needs after two hours on a chair.
The hard part is the moment. A person at work is busy, in the middle of something, and not thinking about their neck. A tool that waits to be opened will wait. We came to see adoption, not content, as the engineering problem.
Seated neck and shoulder stretch. Illustration. Authored pose sequence, not live inference.
02Lessons 1 to 3
Reduce the distance to the movement
1. Go where people already are
The usual answer to sedentary work is one more mobile app for every employee to install. We designed the opposite: a small chatbot inside the messaging tool everyone already has open. A message from Noa arrives in the same place as a message from a colleague. There is no new habit to build, only a prompt in an existing one.
2. Ask for a minute
A break lasts one to three minutes. This is a deliberate ceiling. A twenty-minute session is better for the body and is skipped; a one-minute movement is modest and is done. We chose the version that happens. The library holds about 300 breaks, covering movement, posture, eyes, focus and team cohesion, so that the minute is not always the same minute.
3. Remove the installation
HR and IT deploy the assistant together from the admin centre. No employee installs anything or creates an account. Every step removed between the decision of the company and the first break of the employee is a group of people who would otherwise have dropped out.
03Lessons 4 to 6
Be predictable where it matters
4. Write the scenarios in advance
Noa uses predefined scenarios and no generative AI. In a tool that tells people how to move their body at work, an improvised sentence is a risk with no upside. A written scenario can be checked by a professional before anyone receives it. This costs flexibility, and we accept the cost.
5. Give the employer less than it could have
Only aggregated anonymous usage goes to the employer. Noa does not read other conversations. A wellbeing tool that reports on individuals stops being used honestly on the day people understand it. Protecting the employee is not only an ethical position; it is a condition of use. Data is hosted in the European Union.
6. Choose the level of technology on purpose
Movelytics trains pose models. We could have put a camera and a skeleton in this product. We did not, because a camera at the office adds friction and doubt, and the problem did not need it. Deeptech when the problem needs it, low-tech when that works better: the decision is part of the design, and it should be made on purpose rather than by habit.
04What changed
What this changed for a lab that trains models
These lessons did not stay in one product. They changed three things in how we work on the others.
- We ask where people are before we ask what to build. For a pose feature, that means the phone the user already owns, not the phone we would like them to own. It is why we measure our model on low-end Android devices first.
- We treat every extra step as a loss. An upload, a wait, an account, a permission screen: each one loses users. Running the model on the device removes the upload and the wait.
- We tell clients when a model is not the answer. Some movement problems are solved by a message at the right time. A lab that only sells models cannot say so.
The same reasoning runs through our model work. The PoseTracker Model exists because real-time tracking that only works on recent phones is, in practice, a tool many people cannot use. Teams who want to test it on their own devices can join the private beta.
05Limits
What we cannot claim
This note describes design decisions and the reasoning behind them. The usage figures are internal data, given as totals. We do not publish here a study of health outcomes, a comparison with other tools, or a measure of how each decision contributed. The lessons are those of one product, in one kind of workplace: office teams who work in a messaging tool all day.
The product itself, its offer and its security documentation are on move-break.com. Our page on why we built MoveBreak tells the product story, and the technology page covers the other side of the lab.
Private beta
Measure it on the devices your users hold.
The PoseTracker Model is in private beta.