One sentence can make your AI system high-risk.
Edit the sentence. Watch the rule signal and review record change.
Browser-only preview. No account or document upload. The simplified rules return possible legal signals, not an AI-generated report or a binding classification. No match is not clearance; real assessments require verified facts and human review. Learn about the method.
Real systems are not one sentence.
Paste a paragraph from a specification. See the highest-priority rule signal without hiding unanswered questions.
Each sentence is evaluated independently against illustrative rules. Negated or unclear statements do not become confirmed facts.
Change the description on the left. The inspection updates locally; no text is sent to an assessment service.
The model finds sentences. It never decides.
Four steps, four different responsibilities. The legal rule application must be repeatable; the facts and sign-off still belong to people.
Models extract, rules evaluate, humans approve
Finds candidate facts and keeps the exact source sentence. Never classifies.
Confirms each fact. An unconfirmed fact cannot support a classification.
Applies Article 5, Annex III, GPAI and national overlays to confirmed facts. Same input, same answer.
Sign the record. Open evidence gaps stay visible instead of disappearing.
Generic checks flag possibilities. GRC systems organise work. AZComply links reviewed facts, statutory rules and accountable sign-off.
What lands on your team's desk: a decision record, not a score.
Inspect the source, legal basis and open questions. This illustrative dossier is awaiting approval; a completed record requires a named reviewer's sign-off.
“Candidates are scored and ranked before human review.”system-description.pdf · p. 4 · FACT-017 · sample wording
Same evidence. Three languages. One rule.
A legal finding can be read by a multilingual review team without silently changing its EU AI Act classification.
Why this remains an open Belgian review question
For relevant Belgian employers, CAO 39 can require information and consultation when introducing technology with important collective consequences. Its scope includes a workforce threshold (normally 50 employees), plus defined impact conditions. A recruitment AI tool does not automatically trigger it. The question stays open in the dossier until checked. Belgian FPS Employment — official guidance
Use the right tool for the record you need.
Swipe horizontally to compare the four columns
| Feature | Basic LLM checkers | Enterprise GRC platforms | AZComply engine |
|---|---|---|---|
| Classification | Risk suggestions may depend on prompts and implementation | Policy management, questionnaires and review workflows; varies by platform | Deterministic statutory rules applied to reviewed facts |
| Evidence traceability | Source trace depends on the tool and configuration | Evidence links and provenance vary across platforms | Anchored source excerpts with open evidence gaps listed |
| Human review | Review and approval depend on implementation | Review workflows included; scope varies by platform | Fact confirmation and sign-off before the record is complete |
| Deliverable | Varies by tool: summaries, exports or structured results | Compliance registers, attestations and workflows | Versioned, source-linked dossier for named review |
Pay per decision.
Assessments start at 1 credit. Document assessments use 2; trilingual dossiers add 4. Credits are deducted only after the PDF is generated and stored.
Free risk preview
- Preliminary risk level and article signals
- Main reasons behind the preview
- No LLM report, no dossier
Workspace
- 1 active AI system, 8 credits a month
- Full dossier (PDF + evidence), versioned history
- Evidence Vault, audit trail, reassessment on change
- Collaboration for DPO and legal
Organization
- Several AI systems and team seats
- Advisor, API and white-label options
- Procurement and security review
Prohibited practices (Article 5) and Article 50 transparency duties.
2 December. High-risk systems in Annex III: employment, education, credit.
2 August. AI embedded in regulated products such as medical devices.
Source: Regulation (EU) 2024/1689, amended by Regulation (EU) 2026/1744 (Article 113). The dates shown concern specified provisions and system categories, not every AI Act obligation.
Questions people ask first.
What is AZComply and what does it do?
AZComply is a regulatory detection tool for EU 2024/1689 (EU AI Act). It classifies your AI system against the regulation's risk framework: Annex III high-risk categories, prohibited practices under Article 5, GPAI obligations, and national law overlays for Belgium, Germany, France, the Netherlands, Italy and Spain. It does not provide legal advice.
What is an Annex III high-risk AI system?
Annex III lists 8 categories of AI systems classified as high-risk: biometric identification, critical infrastructure safety components, education and vocational training, employment and worker management, access to essential services, law enforcement, migration and border control, and administration of justice. High-risk systems face the most extensive compliance requirements.
Do you use LLMs? I thought this was deterministic.
Both, in separate roles. Our multilingual extraction pipeline (sovereign EU infrastructure in europe-west4, Zero Data Retention) extracts structured facts from your plain-language description in any of 6 languages. A person confirms them. Our pure Python deterministic engine then classifies risk against Article 5, Annex III, GPAI and national overlays. The LLM never decides risk level. The engine is authoritative.
Which national laws does AZComply cover?
National routing overlays for 8 EU member states: Belgium (CAO 9/39/81/100), Netherlands (WOR Art. 27), Germany (BetrVG §87(1)(6), BNetzA), France (CSE, CNIL), Italy (Statuto Art. 4, Garante, INL), Spain (ET Art. 64.4(d), RDL 9/2021, AESIA), Finland (Traficom, FIN-FSA) and Ireland (AI Office of Ireland, WRC, DPC). Overlays add national obligations and authority routing. They never change the AI Act risk tier, which the deterministic engine owns.
How does AZComply handle data privacy and security?
All data is processed within the EU. The LLM pipeline runs on dedicated sovereign infrastructure in europe-west4 (Netherlands) with Zero Data Retention (subject to provider and deployment verification). Database and storage use Supabase in eu-central-1 (Frankfurt). Retention and file-processing policies depend on the deployed configuration; request the current DPA and technical controls before sharing sensitive materials.
What is a FRIA and do I need one?
Under Article 27, a Fundamental Rights Impact Assessment (FRIA) can be required for certain deployers of Annex III high-risk systems, including public bodies, private entities providing public services, and deployers of specified creditworthiness and insurance systems. The provision has scope limits and exceptions. A FRIA examines potential impacts on fundamental rights; applicability must be checked against the deployer and system profile.
What are the penalties for non-compliance?
Fines are tiered. Prohibited practices (Article 5): up to €35M or 7% of global annual turnover. High-risk violations: up to €15M or 3%. Incorrect information to authorities: up to €7.5M or 1%. SMEs and startups may receive proportionate penalties.
Is this legal advice?
No. AZComply is a detection tool: it identifies regulatory indicators and generates analysis reports. Its output is not a substitute for qualified legal counsel. The framing is “Article X requires…”, not “you are compliant”.
Start with one sentence from your own spec.
The free preview returns preliminary signals. A full assessment connects your evidence, statutory rules and reviewer approvals in a versioned decision record.