Why a ban only makes use invisible
The obvious reflex after the first moment of doubt is a ban. No AI with company data until everything is settled. It looks cautious, but it often achieves the opposite. People who have found a tool useful rarely stop using it. They simply stop using it where anyone can see.
The work then moves to a personal phone, a browser at home, an account tied to a private email address. The company no longer knows which data sits where. It cannot ask either, because officially nobody is doing anything. The risk has not gone away, only the knowledge of it.
A workable policy therefore starts with an honest inventory. Ask the team, without blame, which tools are in use, under which account and for which tasks. The answers show where the rules need to apply. They also show what an approved tool must offer later, so that nobody falls back on private alternatives.
The most important finding from this inventory is rarely about the tools. It is about the data that flows into them.
A company AI policy starts with the data, not the tools
A list of permitted programmes goes out of date quickly. New providers appear, existing ones change their terms, features turn up in software already in use. A classification of data is more stable. It defines what may enter an AI tool at all. For a small team, three tiers are enough.
Free to use. Public information, your own drafts without personal data, general technical questions. Any approved tool may be used here, and nobody needs to ask.
Internal. Quotes, cost calculations, supplier terms, internal workflows. This data may only go into tools that run on a company account. The contract must also make clear that inputs are not used to train the models.
Blocked without individual approval. Personal customer data, health information, job applications, login credentials, draft contracts with third parties. A named person decides case by case. For personal data, a data processing agreement under Article 28 GDPR is also required, meaning an agreement on how the provider handles the data on your behalf.
The advantage of this classification is that staff do not have to read provider terms. They only have to judge which tier a piece of text belongs to. Whether that works is easy to check. Three colleagues asked at random should place a forwarded customer email in the same tier.
Which tools are approved, and why the account decides
Whether a tool is safe often depends less on the provider than on the account. Many providers distinguish between consumer and business accounts. With consumer accounts, inputs may be used to improve the models, depending on the settings. With business accounts, this is usually excluded by contract, and central administration can revoke access.
This leads to a rule that sounds simpler than it is: business data belongs only in company accounts. A privately created account stays private, even if the same tool sits behind it. Training is not the only reason. Whoever leaves the company takes a private account and its history with them, including every piece of customer data ever pasted into it.
The list of approved tools should be short. Two or three tools that cover the most common tasks are better than ten that nobody can keep track of. Each entry states which account applies, which data tier is allowed and who grants access. A request for a new tool should be answered in days, not months. Otherwise nobody will bother to ask.
One case is often overlooked: AI features built into existing software. The CRM, the office suite or the shop receives an update, and suddenly an assistant is summarising customer data. The policy should state that such features are only switched on after review.
Whoever passes on an AI result is responsible for it
Language models write fluently, even when they are wrong. That is precisely what makes review difficult. A clumsy text invites suspicion, a polished one does not. The policy therefore has to define which results are checked before they leave the building, and by whom.
A blanket duty to check everything will not be followed. A tiered approach based on recipient and consequences works better. Internal notes and summaries for personal use need no formal review. Texts to customers are checked by the person sending them, for facts, commitments and tone. Figures, legal questions and product claims need a second person or a reliable source to compare against.
This is where intuition misleads. Risk does not rise with the difficulty of a task, but with its routine. Everyone reads a complicated draft contract carefully. Nobody reads the twentieth reply email of the day that closely. Mistakes happen where AI texts are waved through so often that nobody reads them any more.
A checkable test: can the person who sent a text justify every figure and every commitment in it when asked. If not, the review was a signature, not a review. For customer communication, it is also worth looking at the sources that feed the answers. More on that in the article AI in customer service.
Who is responsible when a question comes up
A policy without a point of contact creates uncertainty rather than clarity. Staff run into cases the rules do not cover. If nobody answers, they decide for themselves, and usually in favour of convenience.
A small company does not need a department for this. It needs one named person with time and the authority to decide. That person answers individual questions, reviews new tools, maintains the list and receives reports when something has gone wrong. It might be the managing director, the person in charge of IT, or someone in operations who knows the tools.
A reporting route for mistakes matters at least as much. Someone who has accidentally copied customer data into a private account must be able to say so without fear of trouble. Only then can anyone assess whether a personal data breach has occurred. Where there is a risk to the people concerned, Article 33 GDPR requires notification to the supervisory authority within 72 hours. A culture in which mistakes are kept quiet makes that deadline impossible to meet.
Since February 2025, the EU AI Act has also required organisations that deploy AI systems to take measures ensuring sufficient AI literacy among their staff. A short briefing on your own policy is an obvious part of this. It should be documented and repeated for new colleagues.
Where a company AI policy turns into paperwork
Many policies exist without changing anything. That is rarely down to the content and usually down to one of four patterns.
It is too long. A twenty-page document is signed once and never opened again. What matters day to day has to fit on one page: data tiers, approved tools, review duties, point of contact.
It prohibits without offering an alternative. If the approved tool is worse than the private one, or a request takes weeks, the private tool stays in use. The policy then contradicts practice, and everyone knows it.
Nobody asks whether it applies. Without an occasional check on which accounts are actually in use, it remains unclear whether the rules take hold. This does not mean surveillance, but a regular conversation within the team.
It is never updated. Providers change their terms, new features appear in existing software. A policy unchanged for a year hardly reflects the actual situation any more. A fixed date each quarter is enough for most small teams.
One limit belongs in any honest account. No policy prevents someone from deliberately breaking it. It creates clarity for everyone who wants to do the right thing. In small teams, that is the large majority.
A policy is an operating state, not a document
The first version is quickly written. It only takes effect once the approved tools are set up, the company accounts are managed and the point of contact actually answers. After that, the ongoing work begins: reviewing new features, maintaining access, adjusting the list as the team grows.
How we look after AI tools and the workflows around them over time is described under AI Operations Care.