On-device or cloud: what the difference actually buys you
Processing on the device sounds like the answer to every privacy question. It is a real difference with a real trade-off, and not every task fits inside it.
Read the article Topic
AI data leakage
Where company data slips into AI tools, and how to see it before it leaves.
Quick answer
AI data leakage happens when someone pastes customer records, contracts, or credentials into AI tools that may store or expose them. European regulators increasingly treat this as a reportable data breach, and the practical fix is to catch sensitive data in the text field before it's sent, not to ban AI.
Most AI data leaks don't start with a breach. They start with a paste. Someone drops a customer email, a contract clause, or a spreadsheet into ChatGPT to save twenty minutes, and that data now sits in a system no one reviewed.
This is no longer hypothetical. The Dutch data protection authority (Autoriteit Persoonsgegevens) reports a rising number of data-breach notifications caused by staff using public AI chatbots, often free accounts, on their own initiative. In one case at a Dutch municipality, a single month's sample turned up CVs, care files, and internal reports uploaded to a chatbot.
The exposure depends less on the tool than on the account and the context. A personal ChatGPT login, a free tier woven into a workspace, or a prompt that happens to include a customer's bank details all carry different risk. These guides break down where each AI assistant sends your data, what it keeps, and which settings actually change the outcome.
Blocking the tools rarely works, because people move to their phones. The more durable approach is to make sensitive data visible in the text field, so the person typing can remove, replace, or mask it before they send.
Articles in this topic
What is prompt injection, and why is it not a hacker story?
A model cannot separate your instruction from text it encounters along the way. Once it can also act, that stops being a curiosity and becomes a governance question.
Read the article What is an MCP server, and what are you connecting to it?
Integrations between AI tools and your own systems now take a few clicks, often from the employee's own laptop. Here is what actually happens.
Read the article
Why an AI agent needs permissions
An agent acts on your behalf, so it acts with your permissions. Why least privilege is harder for agents than for people, and what holds up instead.
Read the article What is an AI agent, and what can it do in your organisation?
The difference between a chatbot that answers and an agent that acts. Once a model can do something rather than say something, the question changes.
Read the article Browser extensions are AI tools too
An extension with read access to every site sees more than most AI services do. We are one ourselves, so this article is about us as well.
Read the article
Who decides what an AI assistant remembers?
Memory became a setting, and settings have owners. Sometimes you, sometimes your administrator, sometimes the vendor deciding for both of you.
Read the article
Before you paste a PDF into ChatGPT, redact it
Advice on safe AI use is almost always about what you type. But a full case file, contract, or report uploaded wholesale skips that advice entirely. Here is what to do about the document, not just the prompt.
Read the article
Free vs business AI accounts: the privacy difference
The biggest privacy variable in AI tools is not which tool your team picks. It is which account tier sits underneath it. A free login and an enterprise licence process the same prompt under very different terms.
Read the article
Microsoft Copilot inherits your access - including the files you forgot you could open
Copilot does not break permissions - it respects them perfectly. That is the problem. It makes every overshared folder, every stale 'anyone with the link', and every forgotten access right instantly findable by asking a question in plain English.
Read the article
DeepSeek at work: the question isn't the model, it's where your data lands
DeepSeek's model is impressive and openly licensed. But the app most people use sends every prompt to servers in China, under a legal regime with no EU adequacy. The privacy question is not how good the AI is - it's where your data sleeps, and who can ask for it.
Read the article
The two Geminis: which one your team uses decides your privacy risk
Gemini is not a website you visit - it's woven into Gmail, Docs, Drive, and your Android phone. The privacy question isn't only what an employee pastes into a prompt. It's which Gemini they are using, what it can already see, and what it's allowed to act on.
Read the article
Grok at work: the chatbot that publishes
Grok lives inside X, a platform built for publishing - and in 2025 its share button quietly turned hundreds of thousands of private conversations into Google-indexed web pages. When the AI sits inside a social network, 'share' and 'publish' blur, and so does the line between a work tool and a personal account.
Read the article
Meta AI is the assistant inside your social apps - and that's the risk
Meta AI is not a separate tool you choose to open. It is built into WhatsApp, Instagram, Messenger and Facebook - the apps people use personally and, in many small businesses, for work too. It appears where the line between private and professional is already blurry, and it is a consumer product with no business protections.
Read the article
Le Chat is the European option - and it still can't fix the paste
Mistral's Le Chat is the GDPR-friendly choice: a French company, EU data residency, no US CLOUD Act exposure by structure. All of that is real and genuinely valuable. None of it changes what an employee chooses to type into the box.
Read the article
Perplexity feels like search - so people paste like it's private
Perplexity looks and behaves like a smarter Google, so employees treat it like one: typing in client names, internal context, and confidential questions as if it were a private search box. It isn't. It's an AI service that retains data by default - and whose crawlers have been caught ignoring the rules.
Read the article
What happens to your data in Claude? The privacy risks, explained
Claude can be a useful work tool, but account type determines whether Anthropic trains on your conversations. The gap that settings cannot close is the sensitive context employees paste before pressing Enter.
Read the article
How AI memory works (and how to manage it)
AI tools now remember things between conversations. Useful for preferences, risky for work: a client name or case detail you typed once can resurface in an unrelated chat weeks later. Here is how to view, manage, and delete memory in each tool.
Read the article What happens when Copilot can reach SharePoint?
Over-broad permissions were harmless for years because nobody could find anything. An assistant that searches well changes that in a single query.
Read the article
ChatGPT and work data: the privacy risk is the paste
ChatGPT can be used safely at work, but the dangerous moment is often the first paste: customer records, HR notes, contracts, screenshots, and files copied into the wrong workspace.
Read the article How AI tools reach files you never uploaded
You do not have to share anything to share something. Once an assistant is connected to your account it works with your permissions, and those reach further than you remember.
Read the article What is RAG, and how does it change your data risk?
Once an AI tool can search your own documents, the question moves from "what am I pasting in" to "what can it reach". That is a different kind of risk, owned by different people.
Read the article Why AI makes things up
Hallucination is not a fault in the system, it is a property of how the system works. And once the invention concerns a real person, it becomes a GDPR question.
Read the article Where is your prompt while you type it?
Between the first keystroke and the send button, your text exists in more places than the interface suggests. Drafts, sync, extensions, keyboards.
Read the article What happens when you upload a PDF to a chatbot?
Step by step, from the moment you drop the file to the moment a summary appears. And why "I only asked it to summarise" describes your request, not your transmission.
Read the article Why encryption does not protect your prompt
Encryption protects text from being read in transit. The recipient has to read it, or it cannot answer. That is not a gap in the design, it is the design.
Read the article Prompt redaction guide: remove, replace, or mask sensitive data before using AI
Concrete examples of risky prompts and safer alternatives for support, finance, legal, HR, healthcare, and product teams.
Read the article What a chat history actually is
Where your conversations live, who can reach them, what deletion does in the database, and what sits in the backup. Four layers, each with its own clock.
Read the article AI data leakage prevention: how to stop sensitive data entering AI tools
A practical guide to reducing AI data leakage with policy, prompt hygiene, browser warnings, and evidence-backed rollout steps.
Read the article
What is a context window, and why does AI forget things?
A long conversation loses its beginning. That does not mean the data is gone. The difference between what the model forgets and what the service retains.
Read the article
What "we don't train on your data" actually means
The promise covers training. It does not cover storage, human review, logging, or retention. What happens to text that is not used for training.
Read the article What you can and cannot share with AI
A practical checklist of data you should never put in an AI tool, what is usually fine, and the rule of thumb that helps you decide in borderline cases. It applies beyond chatbots, to meeting AI, email assistants, and transcription too.
Read the article
Does AI train on my data, and how do I turn that off?
Whether an AI tool uses your conversations for training depends on the account and the settings. How to turn off training, how to request deletion, and why that does not fix everything.
Read the article
Is AI safe for company data?
Your board wants a yes or a no. The honest answer is that this is four questions wearing one coat, and three of them were settled before anyone started typing.
Read the article FAQ
Common questions
What counts as AI data leakage?
Any time sensitive information, such as customer data, financial details, health records, or credentials, is entered into an AI tool that may store or process it outside your control. Under the GDPR, that can qualify as a personal data breach.
Is ChatGPT safe to use at work?
It depends on the account and the settings. Business, Enterprise, and Edu plans exclude your data from training by default; personal and free accounts often do not. The larger risk is what people paste before anyone checks.
How do you stop data leaking into AI tools?
Combine a clear policy, the right account tier, and a signal in the text field itself, so the person can act on sensitive data before it's sent. Awareness at the moment of typing tends to hold up better than a yearly training.
See how BeeSensible works
Detect sensitive data before it leaves your team, in any app, in real time.