Governance: AI Act, NIS2, DORA, GDPR and assurance

Compliance evidence from the records you already keep.

EUVYRA derives evidence from runs, evaluations, releases, gateway traffic and incidents instead of asking people to tick boxes. Compliance leads see which controls hold and which are contradicted. Engineers keep working in the tools that produce the evidence.

The assurance overview for a medium-sized company in Germany: high-risk obligations under the AI Act, important entity under NIS2, not a financial entity under DORA, personal data processed under the GDPR, and an NIS2 notification due in 41 hours.

Know every AI system you build or buy.

Each AI system records its purpose, intended users, affected persons, deployment context, your role under the AI Act (provider, deployer, importer or distributor), its owners, whether it processes personal data, its human oversight measures and its DPIA reference. Bought systems, such as transcription tools or model APIs, sit in the same inventory with their suppliers, the business function they support and a continuity plan.

New use cases start as an intake with an AI Act pre-screen, so a proposal is classified before anyone trains a model. The inventory flags what is missing and exports as CSV.

Risk classification with four eyes

A questionnaire follows Articles 5, 6 and 50 and Annexes I and III of the AI Act and proposes prohibited, high, limited, general-purpose or minimal risk. The assessor may choose a different level with a written rationale. An assessment takes effect only once someone with the compliance or risk function, who is not the assessor, approves it.

An AI system page for a radiology triage assistant: its trust chain from build to prove, and how many of its controls are satisfied.

Controls that check themselves

Per AI system, 32 controls cover the AI Act (Articles 4, 5, 9 to 15, 17, 19, 26, 27, 50, 53, 72 and 73), the GDPR (Articles 5, 22, 25, 30, 32, 35 and 44 to 49, when personal data is processed) and ISO/IEC 42001 Annex A, referenced by number. Which apply depends on the risk level, your role and personal data. Another 27 organization controls cover NIS2, DORA and GDPR breach notification.

Each control names automatic checks that read your technical records. A control is satisfied when its evidence passes, and contradicted when someone marked it as implemented while the evidence fails.

Examples of automatic checks
CheckPasses when
Training lineage completeevery released model version links a training run, a dataset version, a git commit and an artifact digest
Bias evaluation presenta completed evaluation of the released version includes a bias evaluator
Release approvals recordedall approval steps are signed, by at least two different people
Logging active, six months' retentiontraces arrived in the last seven days and trace retention is at least 183 days
Audit chain intactthe organization's audit hash chain verifies
Incidents reported in timeserious incidents were reported within the AI Act Article 73 deadline
Supplier recordedbought systems link a supplier with a contract and data locations

From the same records EUVYRA drafts the technical documentation for a system along the structure of Annex IV of the AI Act, so the document starts from what was actually built and tested.

Approvals bound to people's functions

Every person has governance functions on their membership, such as engineering, privacy, compliance, risk, security or medical. Approval policies are ordered steps, each bound to a function, and only people holding it can sign that step. API keys cannot approve anything.

Default approval steps by risk level
Risk levelSteps, in order
HighEngineering, privacy, compliance
LimitedEngineering, privacy
General-purpose or unclassifiedEngineering, compliance
MinimalEngineering

You can override the steps per risk level and per system, and set release rules: whether production needs an approved risk assessment, which risk levels need a passing evaluation gate, and whether the creator of a release may approve it. Submitting a release runs pre-flight checks first, so nobody is asked to sign something that would fail.

Evidence an auditor can verify offline.

For any AI system, EUVYRA writes a versioned evidence package: the system record, risk assessments, controls, releases with their trust chains, evaluations, monitoring, incidents and the audit trail, with a readable report for people.

A manifest lists every file with its SHA-256 and is signed with the installation's Ed25519 key. The CLI checks hashes and signature without contacting anyone. With the installation's public key, it also proves which installation issued the package.

Every state change writes an audit event in the same database transaction, chained by hash to the previous one, and PostgreSQL rejects updates and deletes on that table.

euvyra evidence verify evidence-radiology-v3.zip \
  --public-key installation.pem

# manifest.json    every file with its SHA-256
# manifest.sig     Ed25519 signature
# system.json  controls.json  releases.json
# evaluations.json  incidents.json
# audit_trail.json  report.html

Assurance across the AI Act, NIS2, DORA and the GDPR

An admin enters the company profile: country, headcount, turnover, balance sheet, sectors, the DORA entity type and whether you supply ICT or AI services to NIS2 or financial entities. EUVYRA works out a status per regime, with the reasons.

  • NIS2: essential or important entity, covered by DORA instead, supplier to NIS2 entities, or out of scope, using the size thresholds of Recommendation 2003/361/EC.
  • DORA: one of the 20 financial entity types of Article 2(1), the simplified framework of Article 16, or an ICT supplier to financial entities.
  • AI Act: obligations per system from your roles and risk levels, with the dates as amended by the 2026 Digital Omnibus.
  • GDPR: in scope as soon as one AI system processes personal data.

A supplier register records contracts, data locations, audit rights and exit strategies and exports the fields the DORA register of information asks for. A crosswalk shows which controls give evidence for which provision in each framework, without counting anything twice.

Incident reporting clocks

Reporting deadlines EUVYRA tracks per incident
RegimeReports and deadlines
AI Act, Art. 73within 15 days of awareness; 10 days on a death; 2 days for a widespread infringement or serious disruption of critical infrastructure
NIS2, Art. 23early warning within 24 hours and notification within 72 hours of awareness; final report one month after the notification
DORA, Art. 19initial notification within 4 hours of classification as major and at the latest 24 hours after awareness; intermediate report after 72 hours; final report after one month
GDPR, Art. 33 and 34supervisory authority within 72 hours of awareness; affected people without undue delay when the risk is high

Someone answers the classification questions per regime, and the clocks follow. EUVYRA drafts the text of each report and records when it was submitted, to which authority and under which reference. It does not send anything itself.

The result is a signed assurance report covering scope, inventory, suppliers, controls and reporting history. Companies that supply NIS2 entities or banks can hand it to customers who must check their supply chain.

EUVYRA collects evidence and supports review. It does not certify conformity, its classification and scoping are documentation aids rather than legal advice, and national NIS2 laws can deviate from the directive. The AI Act guide explains the rules themselves, with sources.

Questions from compliance teams

Can AI assistants approve releases or classify incidents?

No. Approving releases, reviewing risk assessments, classifying incidents, recording notifications to authorities and erasing personal data are refused over MCP and stay with people in the console.

What happens when someone marks a control as done without evidence?

If the control has automatic checks and they fail, it shows as contradicted rather than satisfied, and it does not count towards readiness.

Does EUVYRA cover AI we buy rather than build?

Yes. Bought systems are inventoried with their suppliers, contracts and data locations. Checks that need logs EUVYRA cannot see give a warning instead of a failure, and the owner records where those logs live.

Can auditors work in EUVYRA directly?

The auditor role can read everything and export evidence and audit logs, but cannot change anything. Evidence packages can also be verified without any access to the installation.