In this article
A company AI policy comes down to three management decisions: which tools are approved and with which account, which data never goes into them, and who sees what in Microsoft 365 before Copilot is switched on. They can be taken in a single meeting, fit on one A4 page and prevent the most common data leak of 2026. This guide sets out the right order and what the AI Act already requires of SMEs.
The starting point: AI is already on your computers
According to Verizon's 2026 Data Breach Investigations Report, 45% of workers already use AI regularly on company devices — up from 15% the year before. And 67% do so with accounts that do not belong to the company. So-called shadow AI, meaning tools used without approval, is already the third most frequent non-malicious internal action in data breach records.
The practical conclusion is that banning it does not work. Without an official tool and clear rules, people turn to personal accounts and the company no longer knows what is leaving.
Where the risk lies, in two scenarios
Scenario 1 — sensitive data in personal accounts
Nobody acts in bad faith: someone in accounts pastes a statement to summarise it, a salesperson pastes a client's email, HR asks for help rewording an appraisal.
What decides the risk are the terms of use. OpenAI states that, in its consumer services, conversation content may be used to train models unless the user switches that option off. In business plans and the API the opposite applies: by default, content is not used for training. The difference lies in a company account, managed by the company.
A useful image: using AI with a personal account is like discussing a contract on a crowded café terrace — you do not know who is listening or what is being recorded. The business account is the meeting room, with its own rules and a log of who comes in.
Scenario 2 — Copilot exposes badly configured sharing
Microsoft 365 Copilot does not bypass permissions. Microsoft documents that it only surfaces organisational data to which the user already has at least read access. That is precisely where the problem lies.
DataRoad audits frequently turn up forgotten shares: the board's folder open to everyone, a link for anyone in the organisation created three years ago for the salary file, an old project's SharePoint site with no restrictions. Until now nobody noticed. With Copilot, all it takes is asking about the sales team's salaries — and, if the permissions allow it, the answer appears.
The correct sequence is to review permissions in SharePoint, OneDrive and Teams first, and only then enable Copilot for a pilot group. It is a matter of managed Microsoft 365 configuration, not a purchase.
The rules page, ready to sign
The sheet should be signed by the board, not by the IT department.
| Rule | In practice | Owner |
|---|---|---|
| 1. Approved tools, company account | A short list of authorised tools; personal accounts kept out of work. | Board and IT |
| 2. Prohibited data | Personal data of clients and employees, health data, salaries, passwords, confidential contracts and proposals. | Every user |
| 3. Permissions before Copilot | Review of shares and open links; Copilot first for a pilot group. | IT |
| 4. Human validation | AI prepares, a person checks figures, names and facts before anything goes to a client. | Managers |
| 5. Recorded training | A short session for everyone, with an attendance list — the basis of AI literacy. | Human resources |
These rules sit alongside the internal cybersecurity policies that any company with more than 10 people should have in writing.
Before enabling Copilot, confirm
- MFA Multi-factor authentication active on every account.
- Old accounts Former employees removed and licences recovered.
- Open links Shares with anyone reviewed and, where possible, disabled.
- Sensitive folders Board, HR and finance restricted to those who need them.
- Pilot A group of 5 to 10 people and a deadline for evaluation.
What the AI Act already asks of an SME
The European Artificial Intelligence Regulation — Regulation (EU) 2024/1689 — has been in force since 1 August 2024 and applies in phases. For those using ChatGPT or Copilot day to day:
- 2 February 2025 The prohibited practices and Article 4 began to apply: those who use AI must ensure their staff have sufficient knowledge to use it with judgement.
- July 2026 Regulation (EU) 2026/1744, the Digital Omnibus on AI, postponed the high-risk rules to 2 December 2027 (stand-alone systems) and 2 August 2028 (embedded in products). The literacy obligation remained.
- High risk Includes recruitment, biometrics and critical infrastructure. If you use AI to screen candidates, consult your lawyer.
For those using AI to write emails, summarise documents or prepare proposals, the law does not require certifications. It requires people to know the tool, its limits and the care needed with data, and the company to be able to prove it. Rule 5 covers this.
DataRoad does not provide legal advice. It implements the technical side — accounts, permissions and logs — and helps prepare the training.
DataRoad IT Unlimited
Talk to DataRoad. We assess your operation and propose the right plan.
Example: the same rules in three offices
Velv by OnRising develops software from offices in Porto, Lisbon and Rio de Janeiro and has around 200 workstations under an IT Unlimited contract with DataRoad. International clients account for 45% of its portfolio, and the contract includes firewall management, network security and protection policies.
Here, knowing which client data may go into an AI tool is a daily question, and the answer has to be identical in all three offices. That is why the AI rules are not a separate project: they are part of the existing policies and are applied by the team that already manages accounts, permissions and devices.
Three steps for this week
- Survey — ask the team which AI tools they are using.
- Choice — define one official tool with a business account.
- Review — request a permissions audit before any Copilot pilot.
If MFA is not yet in place, start there — it is the first step of Zero Trust in four steps for SMEs. And remember that AI also serves attackers: prepare the team against phishing and deepfakes.
Frequently asked questions
Does it make sense to ban ChatGPT?
Banning it without an alternative drives people to personal accounts, outside the company's control. It is safer to approve a tool with a business account and define the data that must never go into it.
Can Copilot show files to people who should not see them?
It only shows what the person already has permission to see. If there are folders shared with everyone or old open links, Copilot makes that mistake obvious. Hence the need to review permissions before enabling it.
Does AI literacy also apply to small companies?
Yes. Article 4 of the AI Act has applied since 2 February 2025 to anyone using AI professionally, whatever their size. Short training, written rules and an attendance record are a good start; for the legal interpretation, speak to your lawyer.
Are new licences needed?
Not always. Most of the work is configuring what already exists: MFA, permissions and business accounts on the approved tools. Copilot licences make sense once that foundation is in place.
Permissions reviewed and AI rules in 10 working days
DataRoad reviews the sharing in your Microsoft 365, shows what Copilot would expose and hands over the rules page ready to sign. Security is part of the IT Unlimited Premium plans, with end-to-end managed cybersecurity.
More than 10 workstations · support contract · remote, on-site and telephone support. Request a no-commitment proposal.
Sources
- Verizon — 2026 Data Breach Investigations Report, Executive Summary
- Microsoft Learn — Data, privacy and security for Microsoft 365 Copilot
- OpenAI — How your data is used to improve model performance
- EUR-Lex — Regulation (EU) 2024/1689 (Artificial Intelligence Act)
- EUR-Lex — Regulation (EU) 2026/1744 (Digital Omnibus on AI)
- European Commission — Agreement to simplify AI rules
