For IT and privacy officers
In Bee Flow a prompt goes from your own environment to the AI model you chose, and the answer comes back the same way. This page follows prompts, files and answers step by step, for a self-hosted install and for Bee Flow Cloud in the EU.
Send us your security questionnaire
Technical details per readerA prompt leaves your environment only towards the AI model you chose.
You type a message. It goes to Bee Flow, on your own server or on Bee Flow Cloud in the EU. With the guard service installed, the personal-data filter checks it there, on CPU inside that environment. Then the text goes to the model provider you chose, or to a model on your own hardware. The answer comes back the same way. Conversations are stored in your environment. Connected apps receive only what an assistant is allowed to send.
By default, a self-hosted install sends nothing to Bee Flow.
The path of one message
The same path as the picture, in words. Each step names its licence and its condition.
With the guard service installed, a chat message is checked for 21 kinds of personal data inside your environment, on CPU (Free licence). No text goes to a detection service. If the check is unavailable, the chat message is not sent. A core-only install has no guard service and checks nothing.
On the Pro licence the values that were found are replaced with placeholders before the text leaves, for example a name, a BSN or an IBAN. On Free the message can be blocked. Detection uses machine learning and can miss a value.
Only to a provider an administrator connected, or to a model on your own hardware. Text that was not replaced can be read by that provider.
The answer returns along the same path. On Pro the placeholders are put back inside your Bee Flow environment, so the reader sees the real values and the provider never received them.
The conversation is stored in the database of your environment, under managed keys on every licence. Managed keys protect against a stolen database, not against whoever operates the server.
Two ways to run it
The path of a message is the same. Who operates the server differs, and with it who can reach the data.
Zero-knowledge encryption (Enterprise) covers private chats, with exceptions that the technical page lists.
| Your own server | Bee Flow Cloud in the EU | |
|---|---|---|
| Who operates the server | You, or your IT partner | Bee Flow B.V. |
| Who can read data that is not zero-knowledge encrypted | Whoever operates your server. That is you. | Bee Flow, as the operator. That access is bound by the processing agreement, which you can request from us. |
| Where the data is stored | On the server you chose | In the EU. The data centre and the subprocessors are part of the Bee Flow Cloud agreement. |
| Where backups are | Where you make them. The backup is yours to arrange. | Part of the Bee Flow Cloud agreement. |
| What Bee Flow receives | By default nothing. The licence is checked offline on your server and no usage data is sent, unless an administrator switches on licence refresh or the module Hub. Installing pulls the images from a registry. | Bee Flow runs the environment, so the data is on servers we operate. What we may do with it is in the processing agreement, which you can request from us. |
| Which model providers receive text | The ones you connect, under your own contract with them | The ones chosen for your environment. Which contract covers them is part of the Bee Flow Cloud agreement. |
Text leaves your environment only towards a model. You decide which one.
Anthropic, OpenAI, Google, Google Vertex AI, Azure OpenAI, Mistral, Scaleway, IONOS and EU GPT. Only a provider that an administrator connected receives text. EU GPT cannot use tools.
One switch limits every level to a model hosted in the EU. It covers models. It does not cover connected apps such as mail or web search.
With a local runtime such as Ollama, vLLM, llama.cpp or LM Studio, the text stays on your network. A local language model needs a capable server.
Text that was not replaced is readable by the model provider. Whether the provider keeps it depends on that provider and on your contract with them.
Records stay in the database of your own environment. Tokens, API keys and the personal data the filter found are never logged.
Every outgoing call of a tool or an automation is logged with host, company, country, status and duration, without content, on every licence. The report with the map is part of Advanced monitoring (Pro). Calls to AI models are not on it.
Every sign-in, refusal and change of access is recorded with IP address, method and reason. A failed attempt stores a fingerprint, never what was typed. The viewer and the export are part of the Compliance Center (Enterprise).
Every run of an automation is kept step by step, on every licence, so you can see what each step received and returned.
Per call to a model: the model, the number of tokens and the cost, without the text. What the filter did is recorded as kind of data and action, never the value.
The exceptions
The places where the description above does not hold.
The Swarm level in direct chat. Typed messages in template chat. Typed messages in the web page builder chat. Voice transcripts. The learning coach. The builder chats for automations, App Studio and the CMS. AI chat components inside App Studio apps. AI on publicly shared and previewed web pages.
Chat messages are held. Tool results, memory and knowledge passages pass through unchecked, and the unchecked part of a large attachment is cut off unless you set it to block.
A connected app such as mail, files or web search receives what an assistant sends to it, wherever that app is hosted. The EU switch does not change that. An administrator sets per app and per action what an assistant may do.
No conversation, title or memory. Usage and cost rows, guardrail metadata, the tool log and, where an administrator switched it on, the knowledge search log remain.
FAQ
Not by default. The licence key is a signed token that is checked on your own server, and no usage data is sent to Bee Flow. Two things contact us only when an administrator switches them on: licence refresh and the module Hub. Installing and updating pull the images from a registry.
Whoever runs the server can read data that is not zero-knowledge encrypted. On Bee Flow Cloud that is Bee Flow, and that access is bound by the processing agreement. Ask us for it at info@beeflow.nl. On your own server, you are the operator.
That depends on the provider and on your contract with them. Bee Flow sends the text and does not control what the provider does next. What you control: which provider, the EU switch, replacement of personal data before sending (Pro), or a model on your own hardware.
Once installed, an install on your own server keeps running without internet, with a model on your own hardware. Installing and updating need a registry to pull the images from, and a hosted model needs a connection to its provider.
For Bee Flow Cloud there is a processing agreement, which you can request at info@beeflow.nl. The subprocessors are part of the Bee Flow Cloud agreement. By default, an install on your own server sends us no data.
The pages and articles behind this one.
What an install needs, and what running it yourself involves.
The same subject for the partner or director who has to decide.
Where your AI data goes and who can reach it, beyond Bee Flow.
Mistral, Scaleway, IONOS and EU GPT compared on what they host and where.
Send it to info@beeflow.nl. We answer it in writing and show where each answer comes from. Vulnerabilities go to security@beeflow.nl.
Send us your security questionnaire
Technical details per reader