For regulatory classification, assess the complete deployed application—not just its model. A system is generally machine-based, operates with some autonomy, pursues objectives, infers from inputs, and generates outputs that can affect physical or virtual environments. Inference is the key distinction: deriving predictions, content, recommendations, or decisions from inputs is more likely to qualify than following only explicit human-written rules. section 7.2.1
A fraud-detection module may qualify even when users never see it; model weights sitting idle do not constitute a deployed system. A retrieval-augmented generation pipeline may qualify because it uses a trained model to rank documents by meaning. A generated policy document, once static, is not itself an AI system. section 7.2.1
Table 7.2 When is a system an AI system?
|
Element
|
What it
means
|
Qualifies
as AI System
|
Does not
qualify
|
|
Machine
based
|
Runs on
software or hardware
|
A fraud
detection model running on a server
|
A human
reviewing documents manually
|
|
Operates
with some autonomy
|
The system
can produce an output without each step being manually controlled or exactly
specified by a human. Human supervision for some parts does not disqualify
autonomy
|
A chatbot
that generates its own replies
|
A chatbot
that answers with predetermined responses.
|
|
Designed for
objectives
|
The system
works toward a goal, whether the developer encoded it directly (minimize
prediction error, maximize relevance) or it emerges from training data.
Objectives are internal to the system; intended purpose (Art. 3(12))
describes the deployment context.
|
A claims
model trained to minimize payout prediction error (explicit goal). An LLM
whose response style emerges from training data without a single stated
optimization target (implicit goal).
|
Nothing.
Almost every piece of software has a goal, so this phrase doesn’t actually
exclude any system from the definition. Instead, it acts as a regulatory
“catch all”. The law intentionally covers both explicit goals (what you
specifically tell the system to do) and implicit goals (what the system
inherently learns to do). This prevents companies from dodging AI regulations
by claiming, “We never programmed a specific purpose into this model.”
|
|
Infers from
inputs
|
Derives
outputs by machine learning or by logic and knowledge-based reasoning, not by
following rules a person wrote out in advance. Code or documents a model
generated are not AI systems; the model that generated them is.
|
A system
that reads a customer message and determines intent through a learned model.
An LLM that generates a claim summary from unstructured documents.
|
Software
that routes messages containing a specific keyword to a queue.
|
|
Generates
outputs
|
Produces
predictions, content, recommendations, or decisions.
|
A risk score
flag (prediction), a generated claim summary (content). A suggested
settlement amount (recommendation). An automated denial (decision).
|
A dashboard
that displays raw data without deriving any conclusion. Purely descriptive
statistics tools that present data without recommending or deciding.
|
|
Can
influence environments
|
Outputs must
be capable of having real-world effects (directly on physical objects or
indirectly through digital spaces), data flows, or software ecosystems. The
system need not act autonomously; the influence test is about the potential
reach of its outputs.
|
A triage
recommendation a clinician acts on an automated denial letter sent to a
claimant. A fraud flag that triggers an account freeze. A risk score surfaced
on a dashboard that adjusters can see, even if no one currently acts on it.
(The test is whether the output is accessible to decision-makers, not whether
anyone currently uses it.)
|
An LLM
running in an isolated test environment, producing predictions that never
reach any user, system, or decision process
|
|
Adaptiveness
(optional)
|
Optional
self learning capability that allows behavior to change while the system is
in use. A system can still qualify as an AI system without this feature
|
A
recommendation model that continuously retrains on user feedback. A
fine-tuned LLM updated on new claim data.
|
A static
rules engine deployed once without further learning
|
Compare qualifying and non-qualifying examples across the definition’s criteria. section 7.2.1

