A one-page AI policy for a small company
· By Bee Flow · 5 min read
On 27 January Spain's data protection authority told people who use AI at work to follow their organisation's policy. Many small companies do not have one yet. What goes on one page, how to introduce it in one team meeting, and how to keep it current.

On 27 January 2026 the AEPD, Spain's data protection authority, published ten recommendations for people who use AI tools, under the title "Cuidado con lo que le confías", roughly "careful what you confide in it". Most of the advice is written for citizens. Point four is about work. If you use AI in your job, says the leaflet, follow your organisation's information and security policies, and do not enter confidential information about the organisation, its staff or its clients, such as contracts, reports or strategies.
That advice assumes there is a policy to follow. In plenty of companies with ten or forty people there is none. Everybody uses something, often on a personal account, and nobody quite knows what. January is a good month to fix that, with one page and one team meeting.
Why one page is enough
Nobody reads a twelve-page policy. A single page can go on the wall, into the folder for new starters, and through a meeting in five minutes. The AEPD did write a long one for its own staff, a general policy for generative AI in its administrative processes (version of 27 November 2025). That is far more than a small business needs, but one idea in it carries over well: the risk depends mostly on whether personal data or confidential information is involved.
There is a legal side too. Article 4 of the EU AI Act asks organisations that deploy AI systems to ensure a sufficient level of AI literacy among their staff. That duty has applied since 2 February 2025. According to the European Commission's questions and answers, no certificate is needed, and an internal record of training and other initiatives is enough.
In November 2025 the Commission proposed to simplify parts of the AI Act. On 21 January 2026 the EU data protection bodies, the EDPB and the EDPS, recommended in a joint opinion that the duty for providers and deployers to ensure AI literacy among their staff be kept. How the proposal ends is not known yet. Either way, a page of agreed rules and a meeting where you explain them is a sensible start. This article is not legal advice.
What goes on the page
Four questions, each with a short answer.
Which tools are allowed? Name them, together with the kind of account. A business account the company manages is a different thing from an employee's free personal account. In point five the AEPD advises reading a service's terms before using it, and choosing services that collect no more than they need to provide it. Add that any new tool goes past one named person first.
What never goes in? Be specific, because "no confidential information" is too vague to act on. Write instead: no names, addresses, phone numbers or national ID numbers of customers or colleagues, no salary data, no contracts, no quotes with prices, no medical information. The Spanish leaflet has a useful tip here. Describe a made-up case instead of the real one.
Who do you ask? One name, with a phone extension or a chat channel. That includes the message "I pasted something in by mistake". People need to be able to say that without embarrassment. Then whoever handles privacy in your company can look at it with them and decide whether it is a data breach that has to be reported.
What gets checked before it leaves the building? An AI answer can sound convincing and still be wrong, the AEPD writes in point seven, and checking it is your own responsibility. Agree that a person reads every text before it goes to a customer, supplier or public body, and that amounts, names and commitments are checked.
An example text
The company below does not exist. Adapt the text to your own tools and people.
AI rules at Example Installations Ltd (a made-up example), January 2026
- We use AI for drafts of emails and quote texts, for summaries and for explanations. Only through the business account Sanne set up. No personal accounts for work.
- These never go in: names, addresses, phone numbers and email addresses of customers, anything about colleagues, national ID numbers, salaries, price agreements with suppliers, and photos with people in them. Describing a situation is fine, as long as nobody can tell who it is about.
- Anything that goes outside, you read yourself first. You check amounts, dates, measurements and everything we promise. You answer for it, the AI does not.
- Want to use another tool, or not sure about something? Ask Sanne (office, extension 12). Pasted something that should not have gone in? Tell Sanne the same day. Nobody gets punished for that.
- We review these rules every quarter in the team meeting.
Five points, about a hundred and fifty words. That is all it takes. A fuller starter policy for a firm has its own page: the policy, ready to copy.
Introducing it in one team meeting
Half an hour in the first meeting of the year is enough. Start by asking who uses which AI tool and for what, and let people answer honestly without any judgement. You will hear straight away which tools belong on the page and which tasks come up most.
Then go through the points and show an example at the second one. Take a real customer email, remove the name, address and order number, and show that the question still works without them. That sticks better than a ban.
Close with the name of the person to ask and the date of the first review. Note briefly who was there and what was discussed. That note doubles as your internal record of what you did for AI literacy.

Keeping it alive
A policy discussed in January and then filed away no longer works by April. Put a fixed fifteen-minute item on the team meeting agenda with two questions: are we using anything new, and did anything go wrong. Have new starters read the page on their first day, and ask them after a week whether it was clear.
Update the list of tools as soon as one is added or dropped, and put a date on the page so everybody can see which version applies. When a question comes up that the page does not answer, write the answer on it. After a year you have a page that fits the way your company works.
Where Bee Flow helps
Rules hold up better when the allowed route is also the easiest one. That is what we build Bee Flow for: one AI workspace for the whole team, on your own server or hosted in the EU, where the administrator decides which language models are available. A filter can check messages for personal data before they reach the model. Answers drawn from your own documents point to their source, which makes checking quicker. See private AI chat and Privacy & Security for how it works, or get in touch to arrange an introduction.