Most vendors promise sovereignty. We give you the number.

Data sovereignty

Measured, not promised

Every AI platform says your data is safe. Almost none will tell you which country each request actually went to. Bee Flow records the real destination of every outbound call and turns it into a score you can put in front of an auditor — alongside the four mechanisms that make the number good in the first place.

See the shield working Run it yourself

Four mechanisms

Sovereignty is an architecture, not a clause in a contract

Each of these is something the software does, not something we undertake to do.

Zero-knowledge encryption, keyed per person

Every account gets its own data-encryption key, wrapped with Argon2id and AES-256-GCM and backed by a one-time recovery key. There is no master key and no support backdoor: an administrator resetting a password destroys that data rather than revealing it, which is the honest consequence of the design. It covers your chat and notebook conversations. Knowledge-base documents and meeting transcripts are stored server-readable — we would rather say which is which than let you assume.

Personal data replaced before the prompt leaves

A detector running on your own hardware finds 21 categories — including BSN, VAT and health-insurance numbers — and swaps them for reversible tokens, restoring the real values in the answer. If the detector is installed but unreachable, the request is refused rather than sent unchecked.

Any model, including one on your own network

Adapters for Anthropic, OpenAI, Azure, Google, Google Vertex and Mistral, plus any OpenAI-compatible endpoint. Point it at a model on your own network and Bee Flow treats it as internal — those prompts skip the egress path entirely, because nothing leaves the building.

A score, computed from where the packets actually went

Bee Flow records the destination IP of every integration call at the moment the socket connects — not the hostname it meant to call, the address it reached. That becomes an organisation-wide sovereignty score out of 100, broken down by person, by app, by assistant and by category of personal data, with a per-call log underneath it. When something leaves the EEA you find out because it is on the dashboard, not because someone asks.

What the score is, precisely

It is the share of integration traffic that stayed inside the EEA or never left your network, weighted against calls that carried personal data somewhere else. It is organisation-wide, with four ways to slice it. It is not a per-step figure inside an individual automation — the data is recorded per step, but the score is not computed that way, and we would rather tell you that than let you discover it in a demo. Integration activity monitoring has to be switched on, and the dashboard is part of the Enterprise plan. You can read the live version of it on the Privacy Shield page: the "What happened" tab there is this score, with the destinations it was computed from underneath.

What "European" means here, and where it stops

Bee Flow B.V. is a Dutch company. The managed service runs on Scaleway in EU data centres, so your data is processed and stored in the EEA unless you choose a model or a feature hosted elsewhere. That exception is real and worth reading twice: several of the strongest models are American, and where you pick one, the transfer is covered by standard contractual clauses rather than by geography. Switch on EU mode and every model tier routes to an EU-hosted model instead. Either way the platform records where each call actually went, so the answer is a measurement rather than an assurance.

Live demo · in development

Sovereignty, graded against your own installation

The Compliance Center reads the workspace you are running - where the models are, who can see what, whether the retention job runs - and grades it against GDPR, the EU AI Act and ISO 27001. Switch frameworks on the left, open a failing check to see what it read, or open the Statement of Applicability and read a control exclusion. This is what turns "where did our data go" from a question into a document you can hand to an auditor.

The Compliance Center is still in development - what you see here is real and running, but it is not finished and the details will change. The organisation is invented, the findings are invented; the checks and the Annex A catalogue are the real ones. Nothing is saved and the PDF exports are switched off because the demo has no server behind it.

Open the real thing

In practice

What this looks like on a Tuesday

Someone pastes a customer email into a chat

The name, the address and the account number are found and replaced before the request leaves your infrastructure. The model reasons over placeholders. The reply comes back with the real values restored, and the person never had to think about any of it.

Mail sanne@voorbeeld.nl over factuur NL91 ABNA 0417 1643 00 → Mail [email_1] over factuur [iban_1]

An automation reads a mailbox at 07:30

Every tool call it makes is logged with the destination IP, the country, and whether personal data was involved. Nothing about that is retrospective reconstruction — it is captured as the connection is made.

gmail_search → 142.250.x.x (IE) · EU · no PII detected

Your DPO asks where the data went last quarter

You open one dashboard, filter to the period, and read the score with its breakdown and the calls behind it. The answer is the same whether the person asking is a colleague, a customer or a regulator.

Sovereignty score, broken down · by person, app, assistant, PII category

The awkward questions

Asked in the order a sceptic asks them

Is everything zero-knowledge encrypted?

No, and anyone claiming that about an AI workspace is worth a second look. Your chat and notebook conversations are, with a key only you hold. Knowledge-base documents and meeting transcripts are stored server-readable, because search and diarisation need to read them. We publish which is which rather than letting the strongest claim cover the whole product.

What if I do not install the PII detector?

Then nothing is detected and prompts go out as typed. It is an optional service, and on a self-host you have to deploy it. Once installed it is on by default across all categories, and if it becomes unreachable Bee Flow fails closed and refuses to send rather than quietly passing your data through unchecked.

Can I use a model that never leaves my network?

Yes. Any OpenAI-compatible endpoint works — point it at a model server on your own network and Bee Flow recognises it as internal, so those prompts bypass the egress path entirely. There is no dedicated integration for a particular local runtime; it is a URL, an optional key, and a test button.

Does the score work if I self-host?

The recording does, because it is part of the platform. The dashboard that turns it into a score is an Enterprise feature. Usage caps, for what it is worth, are not enforced off our cloud at all.

Is Bee Flow open source?

Source-available, and we would rather be precise than borrow the word. The server and frontend are published under the Sustainable Use License v1.0: read it, modify it, run it internally without paying — what you may not do is resell Bee Flow as your own hosted service. It is not an OSI-approved open-source licence. We are working toward AGPL-3.0-or-later; today the Nextcloud connector is the only part that carries it.

What holds this up

Ask us the question you cannot get answered elsewhere

Bring the one your security team keeps asking. If the answer is no, you will get a no.

Start a conversation Read the architecture