AI security and governance
One department uses an assistant, meetings are transcribed automatically and a CRM update has switched on AI features. Often nobody decided any of it, yet it is all already in use. We sort out the tools, the data and who is responsible for each. And before an assistant reads company data, we test it the way someone trying to trick it would.
01 · copy the link to this sectionThe situation
Nobody decided it. It just happened, one department at a time.
One department has connected an assistant to the shared folders. The CRM vendor has switched on new features, and nobody has reread the terms. Nobody can say for sure which tools are in use, what data they read or where that data ends up. Sooner or later the question arrives. It comes from a client with a questionnaire, from the board, or after an incident.
In a large organisation a blanket ban doesn’t last long. Procurement renews the licences and departments switch on the new features. What you need is a short list of approved tools, a named owner for each one and a few technical checks done properly.
Then there is a new risk. An assistant that reads emails and documents can pick up instructions hidden inside that content (prompt injection).
02 · copy the link to this sectionGovernance
In practice, that means a list of the tools in use, a name next to each one and a few rules that everyone can read.
03 · copy the link to this sectionSecurity
An assistant that reads company data is a new way into your systems. You test it the way you would test a new door.
Before an assistant or an agent goes live on company data, we run targeted tests. We do this with written authorisation and, where possible, in a test environment. The tests follow the list of the best-known risks for these systems, published by OWASP, a non-profit foundation dedicated to software security.
At the end, the company receives a report with the issues ranked by severity, the proposed fixes and a summary for management.
Text planted in emails, documents and web pages to trick the assistant. The assistant must not act on it.
Even through indirect questions, no one should obtain what they are not allowed to see.
An agent does only what its task requires. A person confirms the important operations.
Questions, sources and actions leave a record that only the right people can read.
04 · copy the link to this sectionReference
Updated 16 September 2026
The AI Act is the European Union’s regulation on artificial intelligence. It also applies to organisations that do not develop AI systems but use them in their work. An amendment that came into force on 27 July 2026 (the regulation known as the Digital Omnibus on AI) moved some deadlines. This is the updated picture.
| From | What | What it means for you |
|---|---|---|
| 2 February 2025 | Prohibition of certain practices (Article 5) and preparation of the people who use AI (Article 4, which the law calls AI literacy) | None of the tools in use may involve a prohibited practice. People who use AI need suitable preparation |
| 10 October 2025 · Italy | Law No. 132 of 23 September 2025 (Italy’s AI law) comes into force | Among other things, employees must be informed about the use of AI in the cases set out in Article 11 |
| 27 July 2026 | Regulation (EU) 2026/1744 (the Digital Omnibus on AI) comes into force | Staff preparation is still required, with no fixed level to reach. The dates for the most sensitive uses of AI have moved |
| 2 August 2026 | Transparency obligations (Article 50) | People talking to an AI assistant must know they are dealing with AI, unless it is obvious. Realistic fake images and videos must be disclosed as such |
| 2 December 2026 | End of the transition period for marking AI-generated content, for systems already on the market before 2 August 2026 | This mainly concerns the vendors of the tools. It is worth asking them to set out in the contract how they handle it |
| 2 December 2027 | ‘High-risk’ systems (Annex III), for example AI used to select staff or to assess creditworthiness | A competent person oversees the system, logs are kept for at least six months and employees must be informed beforehand |
| 2 August 2028 | AI inside products that are already regulated, such as machinery (Annex I) | This concerns those who manufacture or integrate these products |
| Law | What it requires | Why it matters for AI in your company |
|---|---|---|
| GDPR, Regulation (EU) 2016/679, Article 35 | A data protection impact assessment when the risk to people is high | An assistant that reads a lot of personal data may require one. This is assessed with the DPO |
| GDPR, Article 22 | Limits on decisions taken solely by a machine | Decisions about clients and employees are not left to AI without a person involved |
| Legislative Decree 138/2024 (Italy’s NIS2 law), for companies subject to NIS2 | Management of security risks to systems | Assistants connected to company systems fall under the same risk management |
| Law No. 171 of 21 December 2018, in San Marino | Protection of personal data | It is the reference law for organisations in San Marino |
This box is for information only and is not legal advice. How it applies to a specific case is assessed with the company’s DPO and lawyers. For businesses in San Marino the framework is different and must be assessed case by case.
05 · copy the link to this sectionHow we work
Where we start, who does the work and what the company keeps.
We decide where to start. It might be an inventory, an assistant to test, a questionnaire or a question from the board.
With IT and the departments, we list the tools in use, including those that arrived with updates. We look at the tools, not the people.
With the DPO and the lawyers, we decide what to keep, what to replace and what to switch off.
We test the assistants that read company data, with written authorisation.
The register is updated whenever a new tool arrives. If the company asks, support response times and terms are written into the contract.
06 · copy the link to this sectionPrinciples
An assistant that can read everything needs the same checks as anyone who can read everything.
07 · copy the link to this sectionFrequently asked questions
08 · copy the link to this sectionObservatory
Analysis and guides from the Strategic editorial team.
First decide whether the process is ready to be automated, then decide what to use. Here are the criteria for choosing between an off-the-shelf product, custom software and a mixed solution, with multi-year costs and the clauses that protect your exit.
16 October 2026An organisation that is behind on artificial intelligence does not catch up by buying licences. It catches up through a few leadership decisions, taken in order over ninety days. Here is what they are, who takes them and what gets written down.
8 October 2026In a structured organisation, measuring an AI project is not left to whoever proposed it. Management accounting takes it on, with rules approved in advance, comparisons between sites, agreements on logging and indicators for the board.
2 October 2026The essential parts of this page print as a short summary of a few pages, without images or invitations to get in touch.
09 · copy the link to this sectionContact
A few lines are enough. Tell us which AI tools or projects you have, who looks after them and whether there is a deadline, such as a questionnaire, a launch or a board meeting. We reply by email, usually within two working days.
Please do not put documents, credentials or confidential data in the form.
Or write to us:WhatsAppstaff@strategic.sm
On WhatsApp, messages are answered by Strategic’s AI assistant.
