Last week more than 100 technology companies, among them OpenAI, Anthropic, Google and Microsoft, signed an open letter warning that cyber attacks carried out with artificial intelligence will become far more frequent and far better organised (TechCrunch). It sounds like a problem for hospitals and water plants, not for a company with five people. Yet in your own office the risk starts in a much more ordinary way: with whatever your team types today into an AI chat opened on a personal phone.
The numbers show how common the habit is. In a survey of 1,250 employees at large companies, 66 percent had used AI tools at work even though they believed they were not allowed to, 88 percent had put work information into them, and 34 percent had entered customer data (PagerDuty). Small companies are no different, with one twist: you have no IT department to notice. Below are six rules you can write down in half an hour, and they give you back control.
1. Spend 15 minutes listing the tools already in use
Send a message to the company group: "which AI tools do you use for work, including on your phone?". No blame, otherwise nobody answers honestly. You will usually find three or four things you did not expect: an assistant taking notes in meetings, a browser extension rewriting emails, a personal chat account used for quotes. You cannot set rules for something you do not know exists, and this list is the only one that matters.
2. Pick one or two approved tools, on a company account
You do not need ten subscriptions. Choose one for writing and summarising, perhaps one for accounting or invoices, and buy them on a company account rather than on someone's personal account. You gain three things at once: you pay once instead of five separate subscriptions, you can cut off access the day a colleague leaves, and on business plans your conversations are not used by default to train the model. Check that last point in the contract before you sign.
3. Write down what never leaves the company
A short list pinned in your team chat works better than a ten page policy nobody reads. The list should include customer personal data (full name, national ID number, address, phone), signed contracts, individually negotiated prices, passwords and access keys, personnel files, anything health related. The rule is easy to remember: if the information would upset you in an email sent to the wrong person, it has no place in a public AI chat either.
4. Anonymise before, not after
Most tasks work perfectly well without real data. Need a reply to a complaint? Replace the name with "the customer" and the firm with "company X". Need a contract summary? Remove the parties, the exact amounts and the identifying details, keep the clauses. Need a quote? Ask for the structure and the tone, then fill in the figures yourself at the end. It takes thirty extra seconds and removes almost all of the risk, without any loss in the quality of the answer.
5. Keep a human in the loop for anything the customer sees
No quote, invoice, draft contract or official reply leaves the company before a person reads it in full. AI will certainly get things wrong on numbers, on deadlines and on context it simply does not have. Decide who signs off on what: the owner checks the quotes, whoever handles marketing checks the website text, anything with legal weight goes past your accountant or lawyer. This rule saves you from most real problems, because real problems are the ones that reach the customer instead of staying in the chat.
6. Tighten access rights and review them monthly
Today's AI tools ask for access to your email, your files, sometimes your browser. Give them the minimum: one specific folder rather than the whole drive, a dedicated mailbox rather than the owner's inbox. Put a quarter of an hour in your calendar every month to check who still has access, which subscriptions renew automatically, and whether a new tool has appeared. This is where the warning from those 100 plus companies fits in: the cheapest moment to tighten access is before you have a problem.
Data caution in short, and why GDPR matters
When an employee pastes your customer list or a contract into a public chat, your company has transferred personal data to a supplier, often outside the European Union, with no legal basis and with nobody aware of it. In front of the data protection authority it is the company that answers, not the employee who pressed copy and paste. The good news is that prevention is cheap: company accounts, a list of what is never sent, anonymising before sending, and a short note in the job description. You do not need expensive consultancy to take these steps.
One order of magnitude, even though it comes from companies far larger than yours: this year one in four malicious data breaches was carried out with the help of AI, at an average cost of around 6 million dollars (IBM). In a small company the damage is not measured in millions but in lost customers and in workdays thrown off track. The point stays the same: the mistake costs more than the rule you write down in advance.
Your step for today
Pick one item from the list above and do it before this evening. My recommendation is rule number one: send the message to your team and ask for the list of tools in use right now. It costs nothing, it takes fifteen minutes, and in most of the companies I have worked with the answers completely changed the conversation about what to do next.