What not to paste into AI: a one-page set of data rules for a small company
The biggest AI risk in a small company is not in the model. It is in the employee who, at ten to five, pastes a client contract into a personal chat account to “summarise it quickly”. Here are rules you can introduce in a week.
Why it is a problem
AI tools are convenient and free from any browser. An employee does not have to ask anyone before pasting text and getting an answer. They do not do it out of bad faith. They do it because it works. The trouble is that on a free, personal account the vendor’s terms may differ from a business plan: data may be kept longer, sometimes used to improve models, and nobody in the company knows what has already been sent.
The answer is not a ban. A ban moves AI use onto personal phones. The answer is a few simple rules and business accounts on which those rules can actually be enforced.
Data categories
Instead of a long list, four categories everyone can remember:
| Category | Examples | Rule |
|---|---|---|
| Never | Passwords, API keys, tokens, card details, national ID numbers, health data, data about children | Not pasted into any tool, on any account |
| Anonymised only | Customer and employee data: names, addresses, phone numbers, personal tax IDs | Replaced with roles or numbers before pasting |
| Business account only | Contracts, proposals, unpublished financials, customer correspondence, source code | Allowed in a company tool with training switched off; never on a personal account |
| Freely | Public texts, your own notes without personal data, general drafts | Any tool |
The examples are there to be adapted to your industry. In an accountancy practice the “never” category is longer. In a marketing agency more things are “freely”.
Personal versus business accounts
This is the most important decision and the cheapest change. The company opens accounts on a business plan with one or two vendors (ChatGPT, Claude, Microsoft 365 Copilot, Gemini) and hands them to staff. Business plans and the API state in their terms that data is not used to train models, and they give an administrator settings: retention period, a log, features that can be switched off. Personal accounts for work go away.
From then on the rule is simple: for work, company accounts only. Nobody has to remember which tool is safe. They have to remember that the personal one is not.
Anonymisation in practice
Anonymisation sounds heavy, but in practice it is three moves before pasting:
- Names become roles. “John Smith from ABC Ltd” becomes “the client’s finance director”. The model does not need a name to summarise an email.
- Numbers go. ID numbers, tax numbers, bank account, phone, address. Replace with “[number]” or remove. If the task needs the number, it is not a task for a chat window.
- An example instead of the original. If you want AI to write a template reply to complaints, paste one anonymised complaint, not twenty originals.
In implementations where AI runs automatically (email triage, reports), code does the anonymisation before anything reaches the model, and personal data stays in your systems. The employee does not have to remember.
Vendor settings
In every company tool we check and record four things in the documentation: whether training on your data is switched off (on business plans it is by default, but we check), how long conversations are retained, whether there is an administrator log, and where the data physically sits (an EU region, if offered). These settings change along with vendor terms, so a review every six months goes in the calendar.
A one-page policy
You do not need a twenty-page document. You need one page everyone will read:
- Which tools we use and on which accounts. A list with names. Personal accounts for work: no.
- What we do not paste. The four-category table, adapted to you.
- How we anonymise. The three moves from this article, with one example from your own work.
- What a person approves. An email to a customer, a proposal, a contract, a publication: AI writes the draft, a person reads and sends.
- Where to report a slip. Someone pasted something they should not have. One address, no penalty for reporting. Better to know.
- Who reviews, and when. A name, every six months: vendor settings, the list of tools, the category table.
Agree the GDPR aspects of the policy with your lawyer. This page is an outline from the technical and organisational side, not legal advice.
If you want AI to work on your email and documents with these rules built into the mechanism rather than into staff memory, this is how an email and documents project goes.