Glossary · 2026-10-18 · 6 min

Glossary: AI governance, AI security and data rooms

The terms that come up in our articles and products, each in a few sentences. The first sentence is always the definition.

AI register

An AI register is an up-to-date overview of every AI system an organisation uses: which system, for what, by whom and with which data. The EU AI Act prescribes no format, but without such an overview you cannot classify or document a single system. A register kept by hand tends to lag behind actual use.

Risk classification (EU AI Act)

Risk classification is placing an AI system in one of the four risk categories of the EU AI Act: unacceptable, high, limited or minimal risk. The category determines the obligations: prohibited, extensive documentation and human oversight, a transparency duty, or no extra requirements. Even for minimal risk you must be able to justify the classification.

Deployer

A deployer is an organisation that uses an AI system under its own authority, as opposed to the provider that develops it. The EU AI Act places obligations on deployers too. Anyone using an AI assistant or an AI-driven HR tool, for example, has a compliance position of their own.

AI literacy

AI literacy is the knowledge and skill staff need to use AI systems responsibly and understand their risks. Article 4 of the EU AI Act has required providers and deployers to take appropriate measures for it since February 2, 2025.

Shadow AI

Shadow AI is the use of AI tools within an organisation without the knowledge or approval of IT or compliance, such as a personal chatbot subscription or a free AI service that documents are pasted into. The risk is that confidential data ends up with an outside party and that the AI register is incomplete.

RAG (retrieval-augmented generation)

RAG is a technique in which a language model is first handed relevant passages from your own document collection and only then answers. The answer is then based on your own documents rather than on what the model learned itself, and it can cite the source.

Permissions before retrieval

Permissions before retrieval means an AI chat applies the user's access rights in the search step itself, so the model only sees passages from documents this person may read. Filtering after the search looks safe, but gives worse answers and raises the risk of leaks through suggestions or search results.

PII masking

PII masking is making personal data such as names, contact details or national ID numbers unrecognisable before text is stored or passed to a language model. The stricter the masking, the less can leak, but masking too broadly makes answers unreadable.

Data room

A data room is a closed online environment in which an organisation shares confidential documents with a limited group of outsiders, such as investors during a fundraise or buyers in an acquisition. Access runs per person or group, and every action is recorded.

Zero trust document access

Zero trust document access means nobody can see a document unless it has explicitly been shared with them, and the server checks that again on every request. What has not been shared does not exist for the reader: not even the name of a document or folder is visible.

Audit trail

An audit trail is a running log of who did what and when in a system, such as viewing or downloading a document or asking a question. It makes access verifiable after the fact and, in data rooms, is the basis for monitoring and record keeping.

Dynamic watermark

A dynamic watermark is a marking added when a document is opened or downloaded, with details of the reader such as name, email address and time. If a page turns up outside the data room, you can trace who viewed it.

READ MORE
Insights

Field articles on the EU AI Act, shadow AI, AI chat over confidential documents and choosing a data room.

All insights →