The Paste That Cost a Fortune
AI tools have quietly become part of how everyone works. Your sales team drafts proposals in ChatGPT. Your accountant asks Copilot to explain a spreadsheet. Someone in marketing feeds a client brief into a chatbot to "tidy up the wording." None of it feels risky. It feels like using a smarter version of spellcheck.
But here is the part nobody mentions in the excitement: every time a staff member pastes something into a public AI tool, they are sending that information to a company on the other side of the world. Depending on the tool and the settings, that company may store it, have humans review it, and use it to train the next version of their model. Your confidential data becomes someone else's training material.
This is not a hypothetical. In 2023, engineers at Samsung reportedly pasted confidential source code and internal meeting notes into ChatGPT to get help with their work. Within weeks the company banned staff from using public AI tools on work devices. The information was already out the door. You cannot un-paste something.
Why "Free" AI Is Rarely Free
There is an old rule of thumb on the internet: if you are not paying for the product, you are the product. It applies neatly to AI.
Most consumer-grade AI tools, the free versions you sign up for with a personal email, reserve the right to use what you type to improve their models, unless you actively turn that setting off. That is the deal you accept when you click "I agree" without reading. Your prompts, your uploaded documents, and the AI's answers can all become part of how the system learns.
Paid business and enterprise tiers are a different animal. Microsoft 365 Copilot, ChatGPT Team and Enterprise, Claude for Work, and Gemini for Google Workspace all commit, in writing, not to train their models on your business data. That single contractual promise is the main reason these tiers exist, and it is why we steer clients onto them rather than the free consumer versions.
The lesson is simple. The tool your team reaches for by default is usually the riskiest one. The safe version almost always costs a few rand per user per month.
The POPIA Angle You Cannot Ignore
If you operate in South Africa, there is a legal layer on top of all this. The moment a staff member pastes a customer's name, ID number, medical detail, or banking information into an AI tool, you have shared personal information with a third party. Under POPIA, that third party is an "operator," and you are responsible for how they handle that data.
If you have no agreement with that AI provider governing how they protect the information, and no lawful basis for sharing it, you may have just created a compliance problem without realising it. A well-meaning employee trying to work faster can quietly expose you to the exact penalties POPIA was written to enforce. The Information Regulator does not accept "we didn't know the intern was using ChatGPT" as a defence.
The Golden Rule: Never Paste These
You do not need a 40-page policy to dramatically cut your risk. Most of the danger disappears once your team internalises one habit: stop and think before pasting. Some things should simply never go into a public AI tool.
- Client personal information: Names tied to ID numbers, addresses, medical details, or anything you would not want printed in the newspaper. This is the POPIA landmine.
- Banking and payment data: Account numbers, card details, and payment instructions. Obvious, yet it happens constantly.
- Passwords and access keys: Never paste credentials, API keys, or connection strings into a chatbot to "help debug." That is handing someone the keys to your systems.
- Unreleased financials and strategy: Board packs, acquisition plans, pricing models, and anything market-sensitive. Samsung learned this one the hard way.
- Source code and proprietary methods: The thing that makes your business yours. If it is your competitive advantage, keep it off public tools.
- Legal and HR documents: Contracts under negotiation, disciplinary records, and confidential correspondence carry obligations to other people, not just you.
The Practical Do's
Risk management is not about banning AI. The productivity gains are real, and a blanket ban just pushes people to use it secretly on their phones, which is worse. The goal is to make the safe path the easy path.
- Pay for the business tier. Put your team on an enterprise plan with a no-training guarantee and proper admin controls. It is the single highest-value move you can make, and it costs less than most people expect.
- Turn off training in any consumer tool. If staff insist on using a free version, switch off chat history and model-training in the settings. It takes thirty seconds and changes the default behaviour.
- Redact before you paste. Teach the habit of swapping real names and numbers for placeholders. "Client A owes R45,000" works just as well as the real details for getting a useful answer.
- Write a one-page AI use policy. Spell out which tools are approved, what may never be shared, and who to ask when unsure. People follow rules they can actually remember.
- Check where the data lives. For regulated industries, data residency matters. Know which country your chosen tool processes and stores information in.
- Be wary of unofficial apps and extensions. The internet is full of "free AI" browser plug-ins and apps that quietly harvest everything you type. Stick to the official tools from established vendors.
Building a Private AI Setup
For businesses that want the full benefit of AI without the exposure, there is a better path than relying on public chatbots and good intentions. You can give your team an AI assistant that runs on tools you control, grounded in your own documents, with your data staying firmly inside your environment.
That means your staff get fast, genuinely useful answers drawn from your actual policies, contracts, and knowledge base, while nothing sensitive ever leaves to train an outside model. It is the difference between handing your filing cabinet to a stranger and hiring an assistant who works in your own office under your own rules.
The Bottom Line
AI is not the enemy here. Carelessness is. The same tool that leaks your client list in the wrong hands can save your team hours a week in the right setup. The gap between those two outcomes is mostly settings, a clear policy, and a few good habits.
If you are not sure what your team is currently pasting into which tools, that uncertainty is itself the risk. A short audit and a one-page policy will tell you where you stand, and usually costs nothing to start. The expensive option is finding out the hard way, like Samsung did.