Recognise sensitive data before it's sent.
An ID number in a prompt, a bank account number in an email, a diagnosis in a chat message. Bombus, the detection engine behind BeeSensible, recognises details like these as people type, in English and Dutch. People can then mask or remove them before the text goes out. Further down, we explain how we test the detection.
Hi team, Lucía Ortega will be joining the project. You can reach [EMAIL1] directly; her bank account is NL91 ABNA 0417 1643 00.
Which data the detection recognises
From a Dutch citizen service number (BSN) or an IBAN to a diagnosis, an address or an API key: each of the 65 data types belongs to one of the 11 categories below. The levels shown here are examples. You use detection profiles to set what counts as critical in your organisation, and which categories are active in each app.
- · Citizen service number (BSN)
- · IBAN
- · Passport number
- · Driver's licence
- · Case number
- · Medical code
- · Diagnosis
- · Medication and dose
- · Health insurance ID
- · Street address
- · Postal code
- · City / Region
- · GPS coordinates
- · API key
- · IP address
- · Device ID
- · Username
- · Credit card number
- · CVV and expiry date
- · Crypto wallet
- · Amounts
- · Gender
- · Language
- · Marital status
- · Nationality
- · Email address
- · Phone number
- · Web address (URL)
- · Flight number
- · Ticket number
- · Date of birth
- · Event date
- · First name
- · Last name
- · Company name
A healthcare profile switches on medical codes; a finance profile switches on IBANs and card data. Start from the default set or from scratch, and adjust as you go.
Decide what counts as sensitive in your organisation
Choose which of the 65 data types a profile recognises, and give each type a level: standard or critical. In most organisations a BSN counts for more than a street name, and a date of birth is just a date in one team but patient data in another.
- · Toggle a single data type without rebuilding the profile from scratch.
- · Every type comes with an example value, so you can see what it covers.
Where your text is checked, and what happens to it afterwards
The check runs as someone types, and on documents before you share them. The text is processed in working memory and discarded immediately. There is no content log, and the extension doesn't ask for clipboard access.
European infrastructure
On our own servers, hosted by European providers.
By default, detection runs on BeeSensible's own servers, hosted by European providers: Scaleway in Amsterdam, with the models at Hetzner in Germany. Both are European companies without a US parent, and both are ISO 27001 certified. The text isn't stored or logged; only counts reach the dashboard.
Profile-scoped
Each app gets its own profile.
You choose which detection profile applies to each app. A strict profile for ChatGPT doesn't have to apply to Slack: each app follows the profile you assign to it, and the default profile covers the rest.
Two levels
A BSN stands out more than a name.
Critical types (BSN, IBAN, credit card number) get a bright red highlight. Standard types (names, email addresses) get a yellow one. You set the level for each type in the profile.
How reliable the detection is, and how often it raises a false alarm
Accuracy
Every data type is tested on its own.
On our internal held-out benchmark, Bombus reaches an F1 score of 0.96 across all 65 data types, in English and Dutch. We measure by category and by level, so a good average can't hide a weak class. It's our own test, not an independent result.
Low noise
No highlights for everyday words or stray numbers.
Precision is tuned for each data type. An everyday word or a random number doesn't trigger a highlight; a pattern has to match in both structure and context first.
English and Dutch
English and Dutch, both tested.
Dutch-specific types, such as the BSN, Dutch IBANs and Chamber of Commerce (KvK) numbers, are tested on Dutch-language samples, and English types on English-language samples. Both languages are held to the same standard.
Continuously improved
Your text is never used for training.
The BeeSensible team builds and refines the detection engine itself. What your users type never ends up in a training set.
In our own testing, BeeSensible recognised 95 out of every 100 sensitive details.
People no longer have to strip sensitive details out of their text by hand. BeeSensible finds them: names, ID numbers, account numbers, addresses and medical details, in Dutch and English. One click masks them, or masking happens automatically when someone pastes or sends. Your team can use AI to the full and sees straight away what is sensitive.
How we measure this
Measured with Bombus, BeeSensible's detection engine, on 824 synthetic test documents covering all 65 data types, in Dutch and English. The test set is fixed and was not used for tuning. The F1 score is 0.96. Without BeeSensible, nothing is recognised automatically. The figure shows what the detection finds, not how many data breaches are prevented.
Start with the apps your team uses most
Pick the AI tools and email apps your team uses most, assign a profile to each one, then add the rest when you're ready.
