Meet Tsuno at Sirha Lyon 2027 · Stand S.57Learn more
EN FR
Try the demo
Menu

Team and management

AI, voice and privacy at work: the guide

A voice AI at work is lawful in the EU with a written purpose, staff told first, minimal data, confirmed actions and no emotion reading. The full checklist.

Published on

In a restaurant office, a manager and an employee sit side by side reading a printed privacy notice together.

A voice AI at work is possible in France and across the EU under five conditions: a purpose written in one sentence, every employee informed before go-live, data kept to the minimum and for a short time, no action without human confirmation, and no emotion reading or judgment of people. In French companies with at least 50 employees, the employee representative body is consulted before the decision. This guide puts in order what the law requires, what a tool must guarantee, what an employee can ask for and what to demand from the vendor. It gives general pointers, not legal advice: for your own case, talk to your data protection officer or a lawyer.

Monday, 2:50 p.m., a privacy notice on the office desk

Ophélie runs a brasserie with 22 staff. She is about to offer a voice assistant in the kitchen and stockroom. First, she has printed a one-page notice. Karl, a section chef, sits down next to her. He reads in silence, then puts his finger on one line: "Retention period: to be completed."

"You don't know how long it keeps what I say?"

Ophélie does not. She knows what the tool does, not what it keeps. Karl is not refusing anything. He just wants the notice finished before he talks to a machine. The people and the scene are fictional; Karl's question is one every employee has a right to ask.

Why voice changes the privacy question

We say more than we write

In writing, we choose our words. Out loud, we say what comes: a colleague's name, a tired remark, a joke. A voice assistant therefore receives more than it is asked, and it has to be designed to keep only what serves the purpose.

The worry is not technical

Teams do not ask how speech recognition works. They ask who is listening, who is reading, and whether it will be used against them. Until those three questions have a written answer, no feature matters.

The employer stays accountable

In most cases the employer is the controller and the vendor its processor. So the employer informs staff, keeps the records and answers employees' requests. It cannot just point to the vendor's privacy policy.

What the law requires of the employer

The GDPR, in six principles

The General Data Protection Regulation (Regulation (EU) 2016/679) sets the frame. Article 5 requires a specified purpose, data minimisation, limited retention and security. Article 6 requires a legal basis for each purpose: at work, consent is almost never free, so employers usually rely on legitimate interest or performance of the contract. Article 9 prohibits in principle the processing of special category data (health, religion, opinions, trade union membership and so on). Articles 13 and 30 require a notice to the people concerned and a record of processing. Article 35 requires an impact assessment when the risk is high, which is common for a new technology used by staff.

The French Labour Code, in two sentences

In France, article L1121-1 bans restrictions on rights and freedoms that are not justified by the nature of the task or proportionate to the aim. Article L1222-4 states that no information about an employee personally may be collected by a device that has not been brought to their attention beforehand. Notice comes before go-live, not after.

Employee representatives, from 50 staff

In French companies with at least 50 employees, the CSE is informed and consulted on "the introduction of new technologies" (L2312-8) and, before the decision to implement them, on any means of monitoring employees' activity (L2312-38). The file, the deadlines and the pitfalls are covered in should you consult employee representatives before using a voice assistant?. The CNIL side (continuous listening, impact assessment, consent) is in a microphone at work: what the law says in France.

The AI Act and its timeline

The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) applies in stages. Since 2 February 2025, its Article 5 has banned practices deemed unacceptable, including inferring a person's emotions in the workplace, except for medical or safety reasons. The obligations for so-called high-risk systems, which include certain employment uses (assessing, monitoring, allocating tasks based on behaviour), were pushed back to 2 December 2027 by Regulation (EU) 2026/1744, the "digital omnibus", published in the EU Official Journal on 24 July 2026. What the emotion ban covers in practice is explained in AI and emotions at work: what has been banned since 2025. What is coming for employers is in AI at work: what changes for employers in 2027.

What a tool must guarantee, point by point

Here are twelve guarantees an employer should get in writing. They turn the principles above into behaviour you can check.

  1. One written purpose: helping the employee find a procedure, a task, a checklist. Not "keeping a closer eye on the team".
  2. Minimisation: the tool keeps only what serves that purpose, and keeps nothing of what was not a request.
  3. Personal sessions: each person opens their own, the tool answers when spoken to, never listening in the background.
  4. The person's rights, no more: the assistant only sees and does what the person can already see and do. Details in can your manager read your questions to an assistant?
  5. Confirmation: the tool prepares, sums up the effect, and waits for a "yes" before acting.
  6. A log: who confirmed what, and when, with the state before and after.
  7. Written retention periods for each type of data: audio, text, action. See voice data at work: what to keep and for how long.
  8. Named subprocessors, with their country and the safeguards for transfers outside the EU.
  9. Reversibility: getting your data back in a standard format when the contract ends, then seeing it deleted.
  10. Refusing special category data: the tool does not record an employee's health, pregnancy, religion or opinions. That is the subject of MCP and sensitive data: what AI must never record about an employee.
  11. No emotions: no analysis of tone, stress or mood.
  12. No judgment: the tool lines up facts, it does not score an employee or predict who will leave.

The last three cannot be adjusted in a menu: a tool either holds to them or it does not. We explain why in our red lines: the ethics of AI at work, and the difference between helping and monitoring in a companion, not a supervisor.

Who sees what, and who answers for what

The manager, head office, the franchisor

A good tool follows your organisation: an employee sees their role, a manager their sites, an owner their scope. In a franchise network the question gets subtler, because franchisor and franchisees are separate businesses, each the employer of its own staff. We cover it in franchises: what head office sees and what it does not.

When the answer is wrong

An assistant reading an outdated procedure gives a wrong instruction in a confident voice. Who is responsible: the author of the sheet, the vendor, the manager? The answer depends on how the tool shows its sources and how the site approves its sheets. That is the subject of who is responsible if a procedure is wrong?

What an employee can ask for

The employee does not have to wait to be told. They can ask, and the employer must answer.

  1. To be informed beforehand: purpose, data processed, legal basis, recipients, retention period, rights and the data protection officer's contact if there is one (GDPR, Article 13; in France, Labour Code L1222-4).
  2. To access their data: what the tool has kept about them, including the text of their requests if it is stored (GDPR, Article 15). The French data protection authority, the CNIL, points out in its guidance on access to one's employee file that the employer answers within one month, extendable by two months for complex requests, provided the employee is told within the first month. The right of access does not open a colleague's data.
  3. To have inaccurate data corrected (Article 16), and to ask for erasure of what no longer needs to be kept (Article 17).
  4. To object to processing based on legitimate interest, on grounds relating to their situation (Article 21). The employer may continue if it shows compelling legitimate grounds.
  5. Not to be subject to a decision based solely on automated processing that significantly affects them (Article 22).
  6. To complain to the data protection authority if they believe their rights are not respected (Article 77).

In practice, a screen-based alternative must exist for anyone who would rather not talk to a machine. And an access request is made in writing, to the address given in the notice.

The questions to ask the vendor

TopicThe questionA good answer looks like
AudioIs the voice recording kept? For how long?A written period, or "not kept beyond processing the request"
TextAre transcribed questions stored? Who can read them?A period, and a precise list of roles
ActionsWhat does the log keep?Who, when, what, with the state before and after
HostingWhere is the data?A country, not "in the cloud"
SubprocessorsWho processes voice and text? In which country?Named categories and transfer safeguards
AccessCan the assistant see more than the person?No, enforced on the server
RefusalsWhat does the tool refuse, even when asked?A written, testable list
End of contractHow do we retrieve and delete our data?A period, a format, a deletion date

Be wary of answers that sound too good. "Nothing is ever kept", "no mistakes possible", "100% compliant": a serious vendor qualifies its claims. We list the ones we refuse to make in the promises we will never make.

Common mistakes

  • Rolling out first and informing later. French labour law requires employees to know about a device before it collects anything.
  • Getting a consent form signed and stopping there. It replaces neither the legal basis nor proportionality.
  • Leaving the retention period "to be completed". It is the first line an employee like Karl will read.
  • Describing the tool in monitoring language: "see who does their checks", "measure productivity".
  • Thinking a banned feature is just badly configured. The AI Act prohibits emotion analysis at work in the EU, whatever the settings.
  • Forgetting third parties. A microphone in the dining room picks up guests who asked for nothing.

What to measure

  • The share of employees informed in writing before first use: the target is 100%.
  • The number of "to be completed" lines in the notice and the register: the target is none before launch.
  • The response time to employees' access requests, against the one-month legal deadline.

What Tsuno publishes on its Trust page

Tsuno San is the smart voice assistant for teams, connected to the Tsuno environment. San Voice, the voice interaction, is available as a demo; San Audio, the hands-free mode, is being explored and tested. Each person opens their session with the location code and their PIN. Here is what we publish on the Trust and data page, with nothing added.

  • Hosting: the app and its data are hosted in the European Union, in Germany.
  • Subprocessors: the AI models (assistant, voice note transcription) come from providers based in the United States. The app's privacy policy details these transfers and their safeguards.
  • Roles: for your teams and your customers, Tsuno acts as a processor, on your instructions. You remain in charge of your data.
  • Access: every read and every write checks the brand, the location and the role on the server. Removed access takes effect right away and ends open sessions. From a connected AI assistant, reads are capped.
  • Actions: Tsuno prepares, sums up the effect, you confirm, then it runs and logs. Access is checked again when you confirm. The log keeps who, when and what, with the state before and after. Payroll, invoicing, permanent data deletion and legal acts stay out of reach for the AI assistant.
  • Published periods: clock-in check photos are deleted seven days after they are taken. When your subscription ends, your data stays available to you for thirty days, in a standard format.
  • Refusals: Tsuno does not score a person in place of their manager, predict resignations, recommend who to dismiss, infer someone's motivation or mood from their messages, or cross-reference sensitive data to profile someone. The app shows a monthly leaderboard (points earned through training, punctuality and completed tasks); it is never passed to the assistant, and San assesses no one.
  • Demo: the public demo runs on made-up data; questions and transcripts are not kept there.

For San's voice in real use (what is kept of the audio and the text, and for how long), the Trust page gives no retention period. Ask us in a meeting and we will answer in writing. The assistant is presented on the Tsuno San page.

Every article on this topic

This guide is the entry point to the topic. The 18 articles that go deeper are grouped below by use.

Our principles and our limits

What the law says

Employees and their data

Allergens, procedures and liability

The four other reference guides

Key takeaway

A voice AI is accepted at work when its rules fit on one page an employee can read before speaking to it: purpose, data, retention, rights, refusals. Get those answers in writing from the vendor, and launch nothing while a line still says "to be completed".

Frequently asked questions

Can an employer use a voice AI with staff in France?

Yes, if the tool serves a specific work purpose, each employee is informed before it goes live, data is limited to what is needed and the setup stays proportionate. In French companies with at least 50 employees, the employee representative body (the CSE) is informed and consulted before the decision.

Which laws govern AI at work in France and the EU?

Mainly the GDPR, the French Labour Code (articles L1121-1 and L1222-4 on proportionality and prior notice, L2312-8 and L2312-38 on consulting the CSE) and the EU AI Act, whose Article 5 has banned emotion inference in the workplace since 2 February 2025.

Can an employee find out what an AI assistant has kept about them?

Yes. The GDPR gives them a right of access to data concerning them. The employer must answer within one month, which can be extended by two months for complex requests, provided the employee is told about the extension.

What should I ask a voice assistant vendor before signing?

What is kept (audio, text, actions), for how long, where, who can access it, which subprocessors are involved, what managers can see of each person's questions, what the tool refuses to do, and how you get your data back when the contract ends. Ask for written answers.

Can an AI assess employees based on the questions they ask?

It should not. That would stray from the stated purpose, is often disproportionate, and is the quickest way to make people stop using the tool. A serious tool refuses and leaves judgment to the manager.

Is employee consent enough to roll out a voice AI?

No. Because of the employment relationship, an employee's consent is almost never freely given. The employer relies on another legal basis, usually legitimate interest or performance of the contract, and records it in its register.