Banning does not work
Shut AI down and usage moves to personal devices and personal accounts, where you see nothing at all. That is not a security gain, it is a relocation of the risk.
AI governance for security and IT
The risk is not a file any more. It is a text box.
See which AI tools are running across your organisation and decide per tool what is allowed. Cover the moment of input, where coverage usually depends on browser, device onboarding and licence tier, and back it up with figures.
14 days free, no credit card. Rolled out across your organisation within the hour.
"Awareness fades when you're busy. That's why we use BeeSensible." Debby Elfrink, Operations director, Speck Pompen
Your organisation 
Microsoft 365 Copilot Allowed 
Claude Microsoft Security Copilot Allowed 
ChatGPT 
DeepSeek
Microsoft 365 Copilot Allowed Claude Microsoft Security Copilot Allowed CrowdStrike Charlotte AI
Tracecat
ChatGPT Dropzone AI No decision yet
DeepSeek MiniMax / Hailuo Not allowed
Incident report Vulnerability report Log file Network diagram Employee name
Your organisation Microsoft 365 Copilot Allowed Microsoft Security Copilot Allowed MiniMax / Hailuo Not allowed CrowdStrike Charlotte AI
Incident report
Vulnerability report
Log file
Network diagram
Employee name
Plus everything else that is running
Microsoft 365 Copilot Allowed Microsoft Security Copilot Allowed MiniMax / Hailuo Not allowed CrowdStrike Charlotte AI No proxy, no agent on the network. No prompts stored. Everything processed inside the EU.
What this is about
Your stack was built for files. This leaves as text.
Classic data loss prevention was built for files on a drive, email through a gateway and traffic over a network you control. Your stack is strong there, and it stays that way.
But someone pasting a client file into a chatbot does it over encrypted traffic, in a browser tab, often on a tool nobody requested. No file leaves. Inline browser controls do exist, but in practice they depend on which browser, whether the device is onboarded and which licence tier you hold. Outside those conditions, nothing is watching.
Your existing stack misses this
Coverage at the moment of input exists, but it is conditional: often one browser only, only on onboarded devices, and only in a higher licence tier. Precisely the combination that is rarely complete everywhere.
You can only steer on what you see
A rule about approved AI tools is quickly written. Whether it is followed, and which tools arrive meanwhile, appears in no dashboard.
Where this fits
Not a replacement for your DLP. The layer missing beside it.
BeeSensible sits at the moment of typing, in the browser and in the apps on the device. No proxy, no change to your network, no device onboarding. The extension runs in Chrome and Edge, the desktop app on macOS and Windows.
After that comes the same round as any other risk: see, decide, help, substantiate.
01 See 02 Decide 03 Help 04 Substantiate
01See
Know what is running, not what was requested.
BeeSensible sees when AI tools are opened and shows, at organisation level, which ones appear, how usage develops and which are new. Without reading prompts and without building a file on anyone.
- Which AI services are running, including the ones that never passed IT?
- Which appeared for the first time this week?
- Which of them host outside the EU, or have no processing agreement?
AI adoption
68%
81 of 120 seats AI-active this month
Prompts sent
1,596
22% vs previous 30 days
AI tools in use
7
2 newly discovered this period
Safe adoption
71%
of AI use via allowed tools
How often AI is used
Daily3428%Weekly3025%Occasionally1714%Not active3933%
Computed on-device and reported as anonymous counts, never per person.
What AI is used for
Analysis & knowledge354
Summarizing353
Writing communication258
Detections over timeLast 30 days
12,438+18% vs the previous period
Detections by platform
78% adjusted
8,124
81% adjusted
3,210
74% adjusted
812
Most encountered data types
4,212
2,930
1,486
1,104
612
SensitiveHighly sensitive
An impression of the AI tools catalogue. The tools and their risk scores come from the product's own catalogue; the verdict in the last column is your organisation's to set.
The catalog
865 AI tools assessed for risk, and more arrive every week
You should not have to research a tool before you can decide about it.
Before you can decide anything about a tool, you need to know what it does with your data: where the vendor sits, whether it trains on what you type, whether incidents are known. We worked that out for 865 tools, and more arrive every week. What is left is the question only you can answer: is this allowed here, and under what condition?
The GDPR asks for a record of processing, NIS2 for a view of your suppliers, and the EU AI Act for demonstrable AI literacy. Because every observed tool is recorded with its vendor, hosting region and training policy, and carries a decision from your organisation, you can answer at any moment.
See the catalog-
Which tools are in use, and for what
Every observed tool is in the register, with vendor, country, hosting region, training policy and the kind of work it is used for.
-
Which of them fit your policy
One decision per tool: allowed, limited use, not allowed, or no decision yet. Nothing is silently approved; the decision is always yours.
-
What you do to uphold that policy
Anyone opening a tool you have decided something about sees it at that moment, right in the page. That the notice appeared is counted, without content and without a name, so you can show the mechanism actually reached people.
The overview does not establish that you comply with the law. It shows which measures you can substantiate with measurements, and which you cannot. The judgement stays with you.
02Decide
A risk score you can retell.
Per tool you see the vendor, the country, the data processing agreement, certifications, documented incidents with their sources, and a risk score built from six dimensions: data governance, jurisdiction, compliance, incidents, integration scope and vendor trust. Beside it sits a separate EU AI Act reading.
So you see not only a number but where it comes from. The verdict stays yours, in one of four states.
- Allowed
- Fits our boundaries.
- Limited use
- Use it, but not with company or client data.
- No decision yet
- Newly found, not yet assessed.
- Not allowed
- With an approved alternative alongside it, where one exists.
03Help More about document anonymisation
A notice in the moment, not a report afterwards.
When someone opens a tool you have decided about, they see that decision with the reason behind it, and the alternative that does fit. When they type sensitive data, it gets marked as they write: credentials, keys, tokens, IP addresses, customer details.
BeeSensible does not block. That is deliberate: a control that stops the work gets routed around, and then you are worse off than before. It does count how often something was sent despite a marking.
BeeSensible notice
Not allowed DeepSeek is not allowed here
Your organisation does not allow this tool for work. Use an allowed alternative.
Use instead
Open Microsoft 365 CopilotAlso allowed: Microsoft Security Copilot ·Claude
Hold to continue anyway
BeeSensible notice
Limited use DeepSeek is allowed here, with limits
Your organisation allows this tool with limits. Only use it with anonymised data.
Use instead
Open Microsoft 365 CopilotAlso allowed: Microsoft Security Copilot ·Claude
Hold to continue anyway
BeeSensible notice
No decision yet No decision yet on DeepSeek
Your organisation has not made a decision on this tool yet. Do not share personal data or confidential business information here.
Use instead
Open Microsoft 365 CopilotAlso allowed: Microsoft Security Copilot ·Claude
Continue anyway
Only enter public information here
GeminiDefault1 / 3Draw box
Postmortem - phishing campaign
First click at 09:14 by d.hofman, from 145.94.28.117. The message arrived at [email protected].
The analysis showed the submitted password Zomer2026! was also in use on two other services.
4 items will be removedPreviewAnonymise (4)
04Substantiate
A control you can show, not only describe.
How often was a marking acted on? How often did someone take the approved alternative? What share of AI usage now runs through tools you said yes to? And how often was something sent despite a marking?
That last figure exists to test your policy, not to confront a person: it carries no content and no name. Exactly what an audit or a board wants to see, and nothing more.
Warnings about tools
Shown the moment someone opens a tool you have not decided on, or have limited or disallowed. About the tool itself, not about what gets typed into it.
148
warnings shown
61%
chose an alternative
Not allowed52 times
38 alternative9 held through
Limited use44 times
21 alternative15 held through
No decision yet52 times
31 alternative14 clicked through
alternative openedcontinued anyway
Browser use only. Counted at most once per tool per day. Group figures only, never who.
Sent despite warning
1,120
Detected, not adjusted, and still sent.
Critical sent despite warning
214
14% of critical detections
OutcomesLast 30 days
9,702
adjusted before sending
78% of 12,438 detections
Masked7,428(60%)
Removed2,274(18%)
Sent despite warning1,120(9%)
No action recorded1,616(13%)
Accountability GDPRNIS2
- Substantiated
Data minimisation at input
Sensitive data marked and handled before sending
- Substantiated
Register of AI services
Vendor, classification, and a decision per tool
- Partial
Appropriate technical measures
Detection active on the supported applications, handling still below the norm
- Not substantiated
Personal data breach notification
BeeSensible provides the signal, not a reporting process to the regulator
A Tuesday across the organisation
Risk rarely looks like risk.
09:14
Support
Someone pastes a customer ticket into a chatbot to draft a proper reply. Name, email address and account number travel along.
11:30
Engineering
A configuration file goes into an AI tool to trace a bug. It still holds a live key.
13:55
Marketing
Somebody tries a new AI tool for imagery. The vendor hosts outside the EU and has no processing agreement.
16:05
Executive
A board paper gets summarised with AI, on a personal account, because it had to be quick.
None of these people is dodging your policy. They are working. Which is why a marking in the moment beats a rule nobody reads at that moment.
The rules you already sit inside
Fits obligations that already apply.
GDPR
Appropriate measures, demonstrably
The GDPR asks for appropriate technical and organisational measures, and for being able to demonstrate that you meet the principles. The moment of input is precisely where that is hardest to prove.
BeeSensible covers that moment and supplies the figures to back up its working. The lawful basis and the DPIA stay with the organisation.
NIS2
Duty of care and board responsibility
In the Netherlands the Cybersecurity Act has applied since 15 August 2026 to organisations in scope. It asks for risk management, including across the supply chain, and puts responsibility explicitly with the board.
An overview of AI services with vendor, country, processing agreement and documented incidents belongs in that conversation.
EU AI Act
Knowing what is running
Different applications can carry different duties, and AI literacy applies to everyone working with AI. The catalogue holds an EU AI Act reading per tool, separate from our own risk score.
A current register of what is in use is the start. It does not make you compliant with the regulation.
The question that always comes
Sight of the risk, not surveillance of your colleagues.
“If you can see what gets typed into AI tools, what do you keep?”
Very little.
A tool that logs prompts to manage risk creates a second risk: a database full of exactly the data you set out to protect. That is why the product is built so the content never settles anywhere.
The full story in the Trust Center- Detection runs on our own servers at European providers, in working memory, and the text is discarded straight after.
- No external AI service is involved: nothing goes to ChatGPT, Google, or any other party.
- Prompts do not reach the dashboard. Not shortened, not as an example.
- The usage data holds no user id. What an individual colleague did cannot be reconstructed afterwards.
- Reporting is counts and patterns, never a per-employee view.
How it is put together
One process, four parts.
The first three can be bought on their own, so you can start with one.
Before you look further
The questions people ask first.
Does BeeSensible read along with what colleagues type?
No. The text is checked to mark sensitive data, and then not kept. It goes to our own servers at European providers, is checked in working memory and discarded straight after. No external AI service is involved.
Can administrators see which colleague got a marking?
No. For markings, documents and notices, who it was is not recorded, so it cannot be looked up afterwards either. The dashboard has no per-employee view of any kind. At most the department is stored, and only if you switch those insights on, with a minimum group size.
Does BeeSensible block AI tools?
Never. It also never changes what someone wrote by itself. Your colleague sees what was found and what your organisation decided about that tool, and chooses. If something was sent anyway despite a marking, that gets counted, without content and without a name.
Does this replace our existing DLP?
No, and it should not. Classic DLP is strong on files, email and network traffic. BeeSensible sits at the moment of typing, in the browser and in the apps on the device. They act at different moments in the same chain and do not compete.
What does this ask of our infrastructure?
Little. No proxy, no change to your network, no device onboarding and no new certificates. The extension runs in Chrome and Edge and deploys through Intune; the desktop app installs silently in the background.
Where does BeeSensible run?
On our own servers at European providers: the application and the database at Scaleway in Amsterdam, the models at Hetzner in Germany. European companies rather than a European region of an American cloud, so there is no US parent that can be ordered to hand data over. The text is checked there in working memory and discarded straight after. What reaches the dashboard are counts and attributes, never content.
How much work is the rollout?
Not much. No proxy, no change to your network. The extension runs in Chrome and Edge and deploys centrally through Intune; the desktop app installs silently in the background. With a managed rollout your organisation is up inside an hour.
Where does the risk score come from?
From six dimensions: data governance, jurisdiction, compliance, incidents, integration scope and vendor trust. Per tool you see the breakdown, the reasoning and the sources behind documented incidents, so you can retell the score instead of having to believe it.
Start by seeing
Cover the place you currently cannot see.
Within two weeks you can see which AI tools are in use across your organisation, and build from there: decide, help, substantiate.
14 days free, no credit card. Rolled out across your organisation within the hour.
