See sensitive data before you send it.
BeeSensible marks names, IBANs, BSNs and 65 data types while you type, in AI tools, email, and chat. You decide what to do with each one. It never changes your text on its own and never blocks sending.
ClaudeThe method
The same four steps, now for sensitive data.
Every BeeSensible module follows this cycle. Here is what it looks like for the data your people type and paste.
Where sensitive data shows up
Which kinds of data pass by, in which apps, and how often. Aggregated, never the text itself and never per person.
What counts as how sensitive
Per detection profile you set which of the 65 data types are recognised, and whether something is standard or critical. You attach that profile to the apps where it applies.
A marking while you type
What is sensitive gets a marking as someone types. Replace, mask, or remove, in one click. The text never changes by itself.
What happens to the markings
How much gets fixed before anything leaves, and what is sent anyway. The figures you substantiate your policy with.
and again How the cycle works
See where sensitive data shows up.
Aggregated, never individual. Which platforms create the most exposure, how much your team handles before sending, and when it happens.
Set per app what gets recognised
Once you see which data shows up where, you can decide about it. In an environment you set up and assessed yourself, someone may well be allowed to work with client data; in a free chatbot that reads differently. A detection profile records that difference: you create a profile, pick the apps it applies to, and set per data type whether it is recognised and whether it counts as standard or critical.
One profile can cover a single app or a whole group at once, in the browser and in the desktop app. Apps you do not attach keep following the default profile. There is also a profile without detection, for the apps where BeeSensible has nothing to do.
Replace, mask, or remove
The profile sets what gets recognised. What happens next is the employee's call: each marking becomes a choice in the panel. Replace a name with a realistic stand-in so the prompt still reads naturally, mask it behind a placeholder, or remove it entirely. For that one occurrence, or for every occurrence in the field.
What happens after a marking
Every marking is counted with its outcome: replaced, masked, removed, or sent despite the warning. That last figure is the honest one: it shows where the advice is not followed, per app and per data type. It is what substantiates data minimisation at the moment of input, and it shows where a profile should be stricter or can be calmer.
Bombus, our best engine yet.
Bombus 2.1 is the engine behind BeeSensible. A smart combination of fast AI techniques and hard rules, working as one judgement: sharp on structured data like IBANs and medical codes, and just as good on names and addresses that depend on context. On our internal benchmark it scores roughly 96% F1, across all 65 data types, in Dutch and English.
How it works
Checked in the EU, kept nowhere
Detection runs on our own servers at European providers: Scaleway in Amsterdam, the models at Hetzner in Germany. European companies rather than the European region of an American cloud, so there is no US parent that can be ordered to hand data over. The text is checked in working memory and discarded right away, and no external AI service is involved.
Dear Eva Eriksson,
Someone types
in ChatGPT, email or a document
European servers
at Scaleway and Hetzner, in working memory and wiped right after
Check your policy
aggregated numbers for analysis and compliance
Across the apps you already use
AI tools like ChatGPT and Claude, email in Gmail and Outlook, chat in Slack: the apps you already work in, right in the browser. The desktop app adds Outlook, ChatGPT, Claude, and Copilot as native apps on macOS and Windows.
Three actions, and the choice stays with the employee
Replace with a realistic stand-in, mask with a placeholder, or remove it. For that one occurrence, or for every occurrence in the field.
Nothing is stored
Detection runs on our own servers at European providers, in working memory. The text itself is never stored.
Try it on your own text.
Install the extension, open ChatGPT or Gmail, and see what gets marked before anything is sent.
