SquadBear and the EU AI Act
Where AI appears in SquadBear, what it never does, and how that maps to the EU AI Act's rules for AI at work. Our reading, for your DPO or counsel to check.
No emotion recognition
Check-ins are ratings and comments people choose to give, never inferred from a face or voice.
No biometrics
The office kiosk identifies people by NFC card or QR badge. Its camera only reads badges.
No scoring of people
No performance score, ranking, attrition prediction or candidate screening.
A person signs off
Sensitive changes an assistant prepares wait for the person to approve them in the app.
Where AI appears in SquadBear
1. Optional summaries inside the product
When an authorized user asks for it, SquadBear sends selected team-health comments or retrospective cards, without author labels, to Cloudflare Workers AI for a summary, suggested actions or grouping of similar cards. The output is a suggestion a person reads and may discard. It is about what a team wrote, not about any one person.
2. The assistant you connect over MCP
SquadBear's MCP server is a set of tools, not a model. You choose the assistant, such as Claude or ChatGPT, and its provider. Each tool call runs with the connecting person's permissions and approved scopes. When the assistant asks to change leave policies or entitlements, workspace settings or a person's role or status, to invite or add an admin, to erase a person or to start a process, SquadBear stores a proposal for 30 minutes and runs it only after the same person approves it, signed in to the app.
3. Everything else is rules
Leave balances, entitlements, coverage checks, timesheets and attendance totals are computed by the product's rules. A language model never calculates a balance or decides a request.
4. Performance reviews run no model here
Self, peer, manager and upward reviews, goals and anytime feedback are stored and shown as people wrote them. SquadBear does not send a review to Workers AI or to any model of its own. A person may connect their own assistant over MCP. That assistant can read the review its author was asked to write and save a draft, including a suggested rating. The rating that counts is the one the author picks, and the author submits in SquadBear. The assistant cannot submit, share or rank anyone.
Peer answers reach that assistant as Peer 1, Peer 2, in the order they were submitted, with no name and no id. Upward answers are not in that packet. When upward collection closes, unanswered rows are deleted. Submitted rows are stored again under new ids, in random order, with no author and no timestamp, and the originals are deleted. A question is released only when at least three people answered, and the cycle cannot be set below three. No assistant can read an upward response. A writing style can still identify someone. That is not a claim of formal anonymity of the comment text.
Finished reviews are deleted a set number of months after the round closes. The control is Review settings, "Delete reviews after", from 12 to 120 months, 24 by default. The same nightly job deletes anytime feedback that many months after it was sent, and goals that many months after they were closed. Open goals, and a round that never closed, are kept. Cycles, rounds and forms are kept. Deletion cannot be undone. The product interface is English or Polish.
How that maps to the AI Act
Prohibited practices (Article 5)
Since 2 February 2025 the AI Act prohibits emotion recognition in the workplace, except for medical or safety reasons, and biometric categorisation that infers sensitive traits. SquadBear does neither. The kiosk once had a face-recognition option; it has been removed, and the kiosk now reads only cards and QR badges.
High-risk uses in employment (Annex III, point 4)
Annex III lists AI systems used to recruit or select people, to make decisions on promotion or termination, to allocate tasks based on behaviour or personal traits, and to monitor and evaluate performance and behaviour. SquadBear's built-in model path is the team-health summary above. It does not score a person, rank a team or decide a promotion. Performance reviews are a record people write. SquadBear does not run a model over them. If you connect an assistant and use it while drafting a review, you deploy that assistant, and whether Annex III, point 4 applies to that deployment is for your counsel. This page does not make that classification.
Transparency
This page, the Security page and the Privacy Policy describe when an AI request happens and what data it carries, so you can tell your people.
What stays with you
- Deciding whether to turn on AI summaries for team health and retrospectives.
- Choosing, configuring and governing any assistant you connect over MCP.
- Telling your people how AI is used in your workplace, and training the staff who use it.
Frequently asked questions
Does SquadBear use emotion recognition?
No. Emotion recognition at work is prohibited under Article 5(1)(f) of the AI Act. Team health check-ins are ratings and comments people choose to give. Nothing is inferred from faces, voices or other biometric data.
Does SquadBear score or rank employees?
No composite score and no ranking. A review can ask for a rating. The person writing picks it, and it is shown as picked, never averaged. There is no attrition prediction and no hiring screen. Reports such as the Bradford factor are arithmetic on recorded absences, and a person decides what to do with them.
Who is responsible for the AI assistant we connect over MCP?
You choose and deploy the assistant, and its provider is responsible for the model. SquadBear supplies tools that apply the connecting person's permissions, and holds sensitive changes for a person to approve. How you use the assistant is yours to assess.
Is this legal advice?
No. It is our reading of how the product works against the Regulation, published so your counsel can check it. Write to legal@squadbear.com with questions.