BEATRICE

Waitlist and services

Privacy notice

This notice explains which personal data we process when you join the waitlist and when you use Beatrice, under article 13 of Regulation (EU) 2016/679 (GDPR).

It covers beatriceai.it, sign-in on accounts.beatriceai.it, the chat on chat.beatriceai.it, the console on platform.beatriceai.it and the API on api.beatriceai.it. For each processing you will find which data, why, on which legal basis and for how long.

1Data controller

STKZ S.r.l., registered office Via Quintiliano 24/A, 20138 Milano (MI), Italia, VAT number IT12782150960. The platform is developed by Matteo Luigi Feroldi.

Until 18 October 2026 the controller was Matteo Luigi Feroldi, as a natural person. From 19 October 2026 it is STKZ S.r.l., which carries on the same processing, with the same data, on the same servers and with the same retention periods described here.

Privacy contact: privacy@beatriceai.it.

2Waitlist

Data: your email address, the profile you choose (individual, professional, company) and, if you write it, what you would use Beatrice for; along with the language of the page, the date of the sign-up and the version of this notice. Your IP address is only used to limit abuse of the form: it stays in the server’s memory for 10 minutes and is never stored.

Why and on which basis: to run the list, decide the order of invitations and write to you about beta access, at your request (pre-contractual steps, art. 6(1)(b) GDPR).

How long: until 30 days after the invitation, and in any case no longer than 12 months after signing up. Then the data is deleted.

3Account and sign-in

Data: your email address and, if you gave them to us with the invitation, your first and last name; an internal account identifier; sign-in codes, which expire after a few minutes; sign-in sessions, each with the kind of browser and system it came from (for example “Chrome · macOS”), never the IP address. We use no passwords.

Legal basis: the contract (art. 6(1)(b)): they are needed to let you in.

How long: as long as the account exists; each sign-in session lasts at most 90 days, and you can end them whenever you want from the account page. If you delete the account, from the account page on accounts.beatriceai.it or by writing to us, conversations with their files, projects, memory, skills and API keys are deleted at once and the account leaves the sign-in system; usage data, check outcomes and the acceptance of the terms stay for the periods given below. The sign-in system does keep a history of changes to the account, including email and name, for the security of sign-ins.

Providing your email is necessary to have an account; everything else is optional.

4Chat

Data: your conversations (titles, messages, the model’s answers and reasoning), the files you attach, with the text your browser extracts from them, and your projects: name, instructions and reference documents, which the model reads again in every chat of the project. When a conversation grows very long, the model writes a summary of it and from then on reads that instead of the older messages: the summary stays with the conversation, protected and deleted like it.

In Work chats Beatrice writes Python code that runs in your browser, in an isolated environment without internet access, not on our servers. The code, its results and the files it creates, such as charts, Excel sheets or documents, are saved with the conversation, like attachments. Your browser downloads the Python environment and its libraries from our domain. If the code asks for a library we do not keep on our domain, our server downloads it for your browser from the Python Package Index (PyPI, run by the Python Software Foundation, in the United States): PyPI receives only the name of the library, from our server, not your IP address, your account or the conversation. The same holds for the desktop app, which downloads from our domain the libraries it does not carry.

To answer about laws and regulations, Beatrice can consult Normattiva, the database of Italian laws of the Istituto Poligrafico e Zecca dello Stato, and the database of European Union acts of the EU Publications Office (EUR-Lex). Our server sends only the reference of the act, for example “art. 3 of law 241/1990”, or the words searched, without your IP address or your account; Beatrice is instructed not to put names or personal data in searches. The texts received stay in the server’s memory for at most 24 hours, so the same requests are not repeated.

For recent facts, or facts the model does not know, Beatrice can search the web and read public pages, PDFs and data files, in Chat and in Work. Searches go through a metasearch engine of ours (SearXNG) on our servers, which passes them to search engines such as DuckDuckGo and Bing, some based outside the European Union; pages are downloaded by our server, which introduces itself as “Beatrice-User”. Engines and sites receive only the words searched or the address of the page, from our server: not your IP address, your account or the conversation. Beatrice is instructed not to put personal data of the person writing into searches. We do not record the words searched or the addresses read; results and pages stay in the server’s memory for at most 15 minutes. The conversation keeps the titles and addresses of the sources, and in Work the data files downloaded for Python, like the files Python creates.

If you dictate a message with the microphone, the browser records your voice only while the microphone is on and sends the audio to our servers, where a transcription model of ours turns it into text. The audio passes only through the servers’ memory: we do not store it, it does not end up in logs and it is not used to train models. The text comes back into the message field, where you can correct it before sending; once sent, it is a message like any other. Each transcription shows in the usage data like a request to the model: number of tokens, times and outcome, never the content.

In the Chat’s voice conversation the microphone stays on while the conversation is open, also while Beatrice speaks, so you can interrupt her; you turn it off with its button or with “End”. The page notices by itself when you stop speaking and sends only that stretch to our servers, where our transcription model writes it. Beatrice’s answer is read aloud by a speech model of ours, on our servers, one sentence at a time: it is a voice generated by a model. Neither your voice nor Beatrice’s is stored: they pass only through the servers’ memory, do not end up in logs and are not used to train models. What stays written, as in every conversation, are your transcribed messages and the answers. Each sentence read aloud shows in the usage data like a request to the model: number of tokens, times and outcome, never the content.

How we protect them: they are stored encrypted (AES-256-GCM) on our servers in Germany, with one key per organization held by us. The key protecting those keys is never in the backups. In the app each conversation is visible only to its author, owners and administrators of their organization included.

Legal basis: the contract (art. 6(1)(b)).

How long: until you delete them. Owners and administrators of the organization can choose automatic deletion after 30, 90 or 365 days without use. Attachments never sent, and files created in a Work answer that was not saved, are deleted after 24 hours. Projects, with their instructions and documents, stay until you delete them. Backups, in which everything stays encrypted, are overwritten within 3 weeks.

While Beatrice answers, the turn in progress (your question, the answer so far, the steps taken and the state of the work) is saved on our servers, encrypted like the conversation, so the work goes on from where it was after a restart of our servers, the app closing or a long absence. It is deleted when the answer is saved; a turn nobody takes up again for 7 days is closed, and the answer so far saved as interrupted. The desktop app sends with the turn a random identifier of its installation, to take up only its own turns.

You can export your conversations from the chat’s Settings.

5Chat memory

It is optional: the chat offers it once and it stays off until you turn it on. You turn it off whenever you like in the chat’s Settings.

Data: short facts about you that Beatrice draws from your conversations or that you write yourself, such as your name, your job and how you like your answers. They may include special categories of data, such as health or beliefs, if you talk about them in your chats. To draw them, the model reads the new messages of your conversations again on our servers, shortly after you write them.

If you also turn on the search in past chats, Beatrice searches your conversations when useful and reads the passages it finds to answer you. It never reads other people’s conversations.

Legal basis: your consent (art. 6(1)(a) and, for special categories, art. 9(2)(a)). You can withdraw it at any time by turning the memory off; withdrawal does not affect what was done before.

How we protect them and for how long: they are encrypted like the conversations and only you see them. They stay until you delete them, one by one or all together in Settings, even if you delete the conversation they come from. Turning the memory off stops using them and adding new ones, without deleting them. They are deleted with the account; backups are overwritten within 3 weeks. The memory’s readings show in the usage data like the other requests to the model.

6Work skills

Data: the skills you write, upload or ask Beatrice to write in a Work conversation (name, title, description, instructions and other files), their last 20 versions, how many times and when Beatrice used them, and your choices: which of Beatrice’s skills you turned off and whether learning is on.

Learning is optional: Work offers it once and it stays off until you turn it on. When it is on, after a Work answer with many steps, or when you correct Beatrice, the model reads the last exchange of that conversation again on our servers, with the code and its results, and writes or improves a learned skill. Learned skills describe how to do a kind of work, not the content of the conversations: the model is instructed not to put personal data in them, and every learned skill shows under the answer, where you can undo it, and on the Skills page, where you can edit or delete it.

Every skill text that is saved goes through the red-line check described below.

Legal basis: the contract for the skills you write or upload (art. 6(1)(b)); your consent for learning (art. 6(1)(a)), which you withdraw at any time by turning it off on the Skills page.

How we protect them and for how long: they are encrypted like the conversations and only you see them. They stay until you delete them, even if you delete the conversation they come from; learned skills never used for 60 days are archived, that is switched off, not deleted. They are deleted with the account and included in the export; backups are overwritten within 3 weeks. The learning’s readings show in the usage data like the other requests to the model.

7Shared skills

Beatrice also learns shared skills, the same for everyone using Work: short instructions on how to do a kind of work (for example a Svelte component or a spreadsheet), at most 3,000 characters each, only one per topic. The model opens by itself those that help with the work you ask for, except when you choose the skills for a message.

How they come about: after a Work answer with several steps, or when you correct Beatrice, the model reads again on our servers the latest messages of that conversation (at most four), with the code and its results, and decides whether to write a new shared skill or improve an existing one. It stores nothing of the work it read: a shared skill must describe the method, never the content. Before it goes out, every text passes an automatic check that refuses email and web addresses, phone numbers, IBANs, tax codes, VAT numbers, IP addresses and access keys, the red-line check described below and a second reading by the model. Each organization contributes at most 5 times a day.

People using Beatrice do not see the text of the shared skills, not even those who contributed: under the answer and in the skills menu only the name, title and description show. We can read the text, to check its quality, go back to an earlier version or withdraw a skill.

Data: the last 20 versions of each shared skill and, for each one, the organization it comes from, whether from Work, the date and the reason for the change written by the model, which passes the same checks; how many times and when the shared skills are used. The model’s readings for the shared skills are recorded in the usage data like the other requests, without counting against your quotas.

Contributing is on for everyone, and Work presents it with a notice the first time. You can turn it off whenever you want from the notice or from Work’s Skills page: from then on your work is no longer read again for the shared skills, which you still receive. Turning it off does not remove what was learned before.

Legal basis: the legitimate interest in improving Beatrice for everyone who uses it (art. 6(1)(f)), with your right to object by turning contributing off (art. 21).

How we protect them and for how long: the instructions are stored encrypted. Shared skills stay as long as they serve Beatrice; nothing remains of versions older than the last 20. If you delete the account or the organization, the versions lose the link to the organization they come from. Backups are overwritten within 3 weeks.

8API and console

API requests and the model’s answers only pass through the servers’ memory: we do not store them and they never reach the logs. This also applies to the requests the chat and the console’s playground make to the API.

Usage data: for every request we record organization, API key or account, model, number of tokens, timings and outcome. It is used for quotas, to measure costs and to detect abuse, on the basis of the contract and of our legitimate interest in the security of the service (art. 6(1)(b) and (f)). We keep it for 12 months. In the console the members of your organization see it, without the text: of requests made from the chat or the playground by other members they see only that they come from another member.

Of API keys we store only a fingerprint (SHA-256): you see the full key only once, when you create it.

9Red-line checks

An automatic classifier, running on our servers, checks the user’s latest messages, the model’s answers and the text of the skills that are saved against the red lines of the acceptable use rules. If a request crosses them, the service blocks it.

Of each check we record only the outcome: red line, score, action, date, organization and account. Never the text. We keep it for 12 months.

Legal basis: our legitimate interest in preventing seriously unlawful uses of the service (art. 6(1)(f)) and, where the law requires it, the obligation to report them to the authorities (art. 6(1)(c)).

The check can make mistakes: if you think a legitimate request was blocked, write to supporto@beatriceai.it.

10Verification for Beatrice Flash Extreme

Beatrice Flash Extreme, the model without the automatic red-line check, is for verified students, professionals with a VAT number, companies, organizations and public bodies. The owner or an admin of the organization asks for the verification in the console.

Data: the category and the reference you type, that is the VAT number, the tax code of the organization, the IPA code or, for students, only the domain of the university email. For a VAT number we also record the outcome of the check on VIES, the European Commission’s register of VAT numbers, and the name VIES returns, which for a professional is their name. We also record who asked for the verification, our decision, the date and the reason for a refusal, if any. We do not ask for photos of documents.

Students get a 6-digit code at their university address, from our own mail server. The address is only used to send the code and is not stored; of the code we keep a fingerprint, deleted when you type the code or after a day.

Legal basis: the contract, since the verification is the condition for using that model (art. 6(1)(b)), and our legitimate interest in reserving a model without automatic checks to identified people and organizations (art. 6(1)(f)).

How long: as long as the verification is valid; a student’s expires after 12 months. Refused requests and revoked or expired verifications are deleted after 12 months. Deleting the personal account deletes them at once; backups are overwritten within 3 weeks.

11Acceptance of the terms

When you accept the terms of service and the acceptable use rules and declare that you are at least 18, we record the account identifier, the accepted version and the date.

Legal basis: our legitimate interest in being able to prove the acceptance (art. 6(1)(f)). We keep it for the life of the account and the following 10 years, the ordinary limitation period.

12Technical logs and security

To run and protect the services we record technical data about requests: IP address, date and time, the address of the page or API, outcome and browser type. Never the content of requests.

Web server logs are kept for 14 days. Application logs rotate by size: beyond 100 MB per service the oldest are deleted. Limits on sign-in attempts, by IP address and by email, live only in the servers’ memory, for at most 24 hours.

Legal basis: our legitimate interest in the security of the services (art. 6(1)(f)).

13Where the data is and who processes it

The servers are in Germany, at Contabo GmbH, acting as processor.

The model and the red-line classifier run on our own hardware in Italy. The link to the servers goes through a private network encrypted end to end (Tailscale): whoever runs the network cannot read the traffic.

Emails with sign-in codes and with students’ verification codes are sent by our own mail server. Emails you write to us reach our own mailbox on the same server, in Germany, and we keep them as long as needed to answer and to document the answer.

The status page (status.beatriceai.it) is a Better Stack service, which checks from the outside that our services respond. We pass it no user data; the data of people visiting that page is processed by Better Stack under its own privacy policy.

We do not transfer your data outside the European Union and we do not pass it to other artificial intelligence providers. The exceptions are the words of web searches, which our server sends to search engines, some outside the European Union, and the names of the Python libraries Work asks PyPI for, always without your IP address or your account.

14What we do not do

We do not use your content to train models. We do no advertising or commercial profiling, and we do not sell or share your data. The chat memory and the learning of personal skills, if you turn them on, only serve Beatrice to work for you; the shared skills, described above, gather ways of working, not your content.

15Your rights

You can ask for access, rectification, erasure, restriction, objection and portability (arts. 15–21 GDPR) by writing to privacy@beatriceai.it. We answer within one month.

From the account page on accounts.beatriceai.it you can download all your data in one file and delete your account yourself, at any time. Chat conversations you can also export and delete from the chat.

You also have the right to lodge a complaint with the Italian data protection authority, the Garante per la protezione dei dati personali (garanteprivacy.it).

16Cookies

beatriceai.it uses no cookies. The services behind sign-in use only technical cookies, needed for them to work, which require no consent:

  • __Host-zitadel.useragent: recognizes the device for the security of sign-ins, for 1 year;
  • __Host-bea_login (15 minutes) and __Host-bea_sso (30 days): the sign-in in progress and single sign-on across the apps;
  • __Host-bea_chat and __Host-bea_console (30 days), with __Host-bea_chat_login and __Host-bea_console_login (10 minutes): the chat and console sessions;
  • PARAGLIDE_LOCALE: the language you chose, if you change it, for 400 days.

No profiling or statistics cookies.

17Changes

We publish every new version here, with its date. If a change affects your data in an important way, we tell you beforehand, by email or at sign-in.

Version of 31 October 2026. The Italian version prevails.