What E-ARI’s AI is, under the AI Act

E-ARI's own rule engine, the classifier its customers use. Anyone can run the facts below through it again and get the same result; the sealed copy records the exact version of the rules. The rules matched ai assistant (Art. 50(1)) and language model (Art. 50(2)) in the description. No Annex III area or Article 5 practice matched, so the tier needs no human confirmation.

Description

Software for organisations that govern their own AI systems under the EU AI Act and ISO/IEC 42001: a register of AI systems, rule-based risk classification with article citations, evidence management with sealed exports anchored in a public transparency log, obligation tracking, vendor questionnaires, AI policies and AI-literacy training for the customer's staff. A general-purpose language model drafts narrative text, compliance documents and interview summaries, and extracts clauses from uploaded documents. An AI assistant answers signed-in users in a chat, and a Discovery Agent interviews them about how their organisation uses AI.

Purpose

Help compliance, legal and IT teams record, classify and evidence their organisation's AI systems and obligations. Outputs are records and drafts that the customer reviews; the risk tier is decided by rules, not by a model.

Sector

Software for compliance and governance teams

Role

provider

People affected

Staff of customer organisations who use the platform: compliance, legal, risk and IT professionals.

The engine leaves 4 questions open for a system like this. E-ARI’s answers:

Article 50(1)

is it already obvious to users that they are dealing with an AI system?

E-ARI does not rely on that. Every chat says so (see Article 50(1) below).

Article 50(2)

does the system show people generated content, or only labels and scores?

Generated content: narrative text and drafts that people read.

Article 50(2)

are outputs marked in a machine-readable format?

Yes, with a signed label and paragraph fingerprints, but without a watermark (see Article 50(2) below).

Article 50(2) exception

does the system only assist editing, without substantially changing the input?

No. The model writes new text, so the exception does not apply.

These are the duties E-ARI’s own obligations catalogue lists for a provider of a limited-risk system, each with E-ARI’s account of where it stands.

Art. 50(1)

Tell people they are interacting with an AI system.

The chat is titled 'E-ARI AI Assistant' in all five languages. The Discovery Agent says 'You are talking to an AI system, not a person' above the conversation and in its first message. The Assessment Advisor labels each answer a model wrote 'AI Generated'.

Check: /discovery

Met

Art. 50(2)

Mark generated content in a machine-readable format, so it can be detected as artificially generated.

Text the model writes carries a signed, time-stamped label marking it as AI-generated, on the page and in exported Word and PDF files, and its paragraphs are fingerprinted so a free public checker can recognise them. There is no watermark in the wording itself, which the Commission's marking Code of Practice expects for free-form text over 200 tokens; only the model provider can add one.

Due by 2 December 2026 (Art. 111(4): a system already on the market before 2 August 2026.)

Check: /security#ai, /api/public/ai-content

Partly met

Art. 50(3) · Art. 50(4)

Disclose emotion recognition, biometric categorisation and deep fakes.

E-ARI does no emotion recognition or biometric categorisation and produces no deep fakes.

Does not apply

Art. 4

Take measures to support the development of AI literacy among the staff and others who operate or use the system on the provider's behalf.

E-ARI's own AI-literacy measures are not yet written down.

Not documented

Art. 5

None of the 10 practices Article 5 prohibits.

They are subliminal or manipulative techniques that cause significant harm; exploitation of vulnerabilities; social scoring; predicting criminal offences from profiling alone; facial recognition databases built by untargeted scraping; emotion recognition at work or in education; biometric categorisation by sensitive characteristics; real-time remote biometric identification for law enforcement; generating or manipulating intimate images of identifiable people without consent (from 2 December 2026); producing child sexual abuse material (from 2 December 2026). E-ARI does none of these, and its model writes text only.

Met

• The risk tier is decided by deterministic rules that cite the article they rely on. The model writes explanations and drafts; it never sets the tier.
• A tier that depends on unanswered questions is marked provisional, with the questions attached.
• Readiness scores use no model and no randomness: the same answers always give the same score.
• Legal citations the model produces that the rules did not apply are discarded before anyone sees them.
• Text a model writes is labelled and marked as AI-generated (Article 50 above).

• Sealed evidence exports (.see) can be checked with the open-source verifier e-ari-verify (npm, MIT licence), without an E-ARI account.
• Every seal is entered in a public transparency log at https://www.e-ari.com/log, and this statement is sealed the same way.
• Readiness scores in a sealed export can be recomputed offline with e-ari-verify replay.

No accuracy figures are published for the language-model features.

• E-ARI is not high-risk, so the Article 26 duties of deployers of high-risk systems do not apply to you.
• Article 4 applies to you as well: take measures to support the AI literacy of the people who use E-ARI on your behalf.
• Article 50(3) and 50(4) are not engaged by what E-ARI does.

A use that changes this. Using E-ARI's records or drafts to make decisions about individual employees, such as judging performance from training records, is outside its intended purpose. Annex III point 4 makes that use high-risk, and Article 25(1)(c) can then make you its provider.