The vocabulary that carries legal weight

A handful of terms in the EU AI Act are not description. They are switches. Each one, once it applies to a system, changes which obligation set attaches. Practitioners who use them loosely in minutes, policies and vendor correspondence create a record that will be read back to them.

AI system

Article 3(1): a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment, and that infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions capable of influencing physical or virtual environments.

Inference is the constant. Adaptiveness is optional, which disposes of the argument that a frozen model is out of scope. Autonomy varies, which disposes of the argument that a system with a human in the loop is out of scope. Neither of those arguments survives a reading of the definition, and both will be made to you.

Provider and deployer

The provider develops a system or has one developed and places it on the market or puts it into service under its own name or trade mark. The deployer uses a system under its own authority in a professional activity. Most regulated firms are deployers of most of their estate and providers of a few systems they built.

The role is per system, not per organisation. A firm is routinely a provider of three systems and a deployer of forty, and a governance programme that assigns the firm a single role will misstate its obligations on every entry that does not match.

Intended purpose

The use for which a system is intended by the provider, as stated in the instructions for use. This is the most consequential phrase in the Act and it is usually written by a vendor's product marketing team. Intended purpose drives classification; classification drives everything else.

Two operational consequences. A vendor who writes "general business productivity" has transferred an unresolved question to you, because that phrase cannot tell you whether the tool may be pointed at candidate screening. And using a system outside its stated intended purpose puts you outside whatever assurance came with it.

Substantial modification

One of the three routes by which a deployer becomes a provider. It is worth pinning down in your own policy what your firm will treat as substantial, in advance, because the decision is otherwise made retrospectively by whoever is uncomfortable.

Serious incident

Article 3(49) defines it: death or serious harm to a person's health; serious and irreversible disruption of critical infrastructure management and operation; infringement of Union law obligations intended to protect fundamental rights; or serious harm to property or the environment.

Note the third limb. An infringement of fundamental rights protections is a serious incident in its own right, with no physical harm required. A discriminatory outcome in an Annex III system is therefore capable of starting a reporting clock, which is not how most incident taxonomies in financial services are currently drawn.

AI literacy

Article 3(56) defines it as skills, knowledge and understanding allowing informed deployment and awareness of opportunities, risks and possible harm. Article 4 carries the duty, and the Digital Omnibus softened its wording from ensuring a sufficient level to supporting its development. It remains an obligation of means rather than result: no certification is mandated, the level is proportionate to role, context and the people affected, and it applies to all AI systems rather than only high-risk ones. It reaches staff and others operating systems on your behalf, which the Commission reads to include contractors.

Words that decide nothing

"Algorithm" appears in every discussion and carries no legal consequence. "Machine learning" is a technique, not a category. "Automated decision-making" belongs to data protection rather than to the AI Act: it appears in GDPR Article 22 and in Article 21 of the revised Swiss Federal Act on Data Protection, and it binds independently of whether AI is involved at all. Keeping the two vocabularies separate matters, because a system can sit outside every AI Act tier and still be a solely automated decision with significant effect, which brings duties to inform, to allow the person to state their position, and to allow them to request review by a natural person.

Worked example: three sentences in a steering pack

An insurer's quarterly AI steering pack contains three sentences drafted by a product manager. "We have deployed the vendor's underwriting assistant across life new business." "We have tuned it on our own historical cases to improve accuracy." "The tool is branded in our customer portal as our own decision support."

Read as vocabulary rather than as prose, those three sentences say something different. Life insurance risk assessment and pricing is an Annex III high-risk category. Tuning on the firm's own cases raises substantial modification. Branding it as the firm's own is, on its face, the first of the Article 25 routes by which a deployer becomes a provider.

Nobody in the meeting intended to announce that the firm had acquired provider obligations for a high-risk system. That is what the pack says. The reason to drill this vocabulary into a governance function is not pedantry. It is that the words are load-bearing whether or not the person writing them knows it, and the pack is a dated internal record.

As of 24 September 2026.