Glossary

An agent is given a goal rather than a single instruction and decides its own intermediate steps — a search, a calculation, a lookup. That increases reach and risk together: anything acting on its own needs tightly bounded permissions and a log that records every action.

Article 4 of the EU AI Act requires providers and deployers to ensure that people working with AI systems have sufficient AI literacy. This is not about a certificate but about demonstrable understanding: what the system does, what can be relied on, and how to recognise a failure.

Every model can only process a limited amount of text at once — input and output combined. Anything beyond that is dropped, usually from the beginning. With long documents, this is why a model appears to "forget" the opening.

Anyone processing personal data for you is a processor, and the relationship needs a contract under Article 28 GDPR. With AI vendors, two points belong in it explicitly: the list of sub-processors, and the exclusion of your inputs from training use.

See also: Data residency, Purpose limitation

The place of processing and the place of storage can differ, and both belong in a contract rather than on a product page. Sub-processors count — including for log data. Support access from a third country is processing.

See also: Data processing on behalf

The EU AI Act distinguishes provider, deployer, importer and distributor. Anyone using a bought-in system is a deployer, with markedly lighter duties than a provider — but not none: purpose limitation, human oversight, informing affected people, and retaining logs all remain.

See also: High risk

Under audit, what counts is not the description of a process but proof that it ran that way. A usable log answers four questions without rework: which process, on which model version, over which data, under whose control.

An existing model is trained further on your own examples so it hits a style or format more reliably. For factual knowledge, fine-tuning is usually the wrong instrument — RAG suits better, because sources can then be named and updated.

Language models produce text that sounds probable, not text that is evidenced. Where knowledge is missing, the model fills the gap with a plausible phrasing. The result is linguistically flawless and factually wrong. That is precisely why source citations and a second reader belong to any process whose output leaves the building.

The EU AI Act classes a process as high risk when it materially affects people's rights or opportunities — recruitment, credit decisions, access to education or benefits. The classification triggers risk management, data quality requirements, human oversight and documentation duties.

See also: Deployer

Human oversight means more than someone sitting nearby. What is required is that a person can judge the output, receives the information needed to do so, holds the authority to intervene or stop — and that this intervention is visible in the log.

See also: Evidence

A prompt is the instruction a language model receives. A usable one states four things: the task, the context needed, the required format, and the limits. A missing one of those is, in practice, the most common cause of unusable output.

See also: Context window

The principle applies unchanged when a language model is involved. Customer data collected to perform a contract cannot simply be fed into an analysis — the new purpose needs its own legal basis.

Retrieval-augmented generation: before the model answers, a search runs over your own document collection and the matching passages are attached to the query. The model then answers from those passages rather than from memory. The practical gain is not better prose but the ability to cite a source.