Marco Combetto

AI & Digital Transformation — Public Sector · AI Act · Data Science

What is required for a Public Administration to be compliant when using an AI system?

For an Italian Public Administration (PA), the obligations are legal, not optional.

In practice:

An Italian PA must comply with:

  • GDPR
  • Codice dell’Amministrazione Digitale (CAD)
  • Linee guida AgID su AI e cloud
  • AI Act (now adopted at EU level)
  • Public procurement code (Codice degli Appalti)

This means no PA can freely use public AI tools like ChatGPT SaaS for institutional data.

2. What an Italian PA is allowed to do

An Italian PA may use AI only if all of the following are true:

  • ✅ Data is hosted in EU or Italy
  • ✅ No training on PA data
  • ✅ Contractual guarantees on confidentiality
  • ✅ DPIA (Data Protection Impact Assessment) completed
  • ✅ DPO approval
  • ✅ Cybersecurity validation (ACN / ISO 27001)
  • ✅ Procurement via public tender or framework agreement

Without these, use is illegal.

3. What an Italian PA typically does in practice

There are three compliant paths:

  1. Use an EU-approved GPAI via a controlled environment
    • Example: Microsoft + OpenAI via EU cloud with contractual safeguards
  2. Use an open-source LLM deployed internally
    • Full control, higher cost, maximum compliance
  3. Sector-specific AI platforms certified for PA use

4. Key difference vs. EU Institutions

  • EU Institutions → Approval by EDPS
  • Italian PA → Approval by:
    • DPO
    • AgID
    • ACN (cybersecurity)
    • Public procurement authority

Bottom line (no euphemisms)

An Italian Public Administration cannot legally use any AI system unless it is contractually governed, EU-hosted, DPIA-approved, cybersecurity-certified, and procured under public law.
“Trying tools informally” is a regulatory violation.

Below is the exact meaning and operational structure of
“EU-approved GPAI via a controlled environment” for Public Administrations.

1. What “EU-approved GPAI” actually means

It does not mean:

  • “Approved by the EU in general”
  • “Certified by Brussels”
  • “Compliant by design”

It means:

A commercial GPAI (e.g., GPT-4/4o, Claude, Gemini) that is contractually constrained, technically isolated, and legally validated for use by public bodies.

Approval is obtained through:

  • ✅ Formal contract (DPA + SCCs)
  • ✅ GDPR Art. 28 processor qualification
  • ✅ No data reuse for model training
  • ✅ EU-only data residency
  • ✅ Security certification
  • ✅ DPIA + DPO sign-off

Without those: not approved.

2. What “controlled environment” technically means

A controlled environment is not a website login. It is:

A. Infrastructure Control

One of the following:

  • Dedicated EU cloud tenant (Azure EU, AWS Europe Sovereign, Google EU)
  • National sovereign cloud (e.g., Polo Strategico Nazionale – Italy)
  • On-premise deployment (rare for closed models)

Requirements:

  • ✅ Physical data location in EU
  • ✅ EU legal jurisdiction only
  • ✅ No third-country admin access

B. Model Access Control

The PA does not access the public SaaS. It uses:

  • Private API endpoints
  • Network-isolated access
  • Role-based access control (RBAC)
  • Logging + audit trails

This prevents:

  • Shadow usage
  • Prompt leakage
  • Cross-tenant contamination

C. Data Processing Guarantees

Contractually enforced:

  • ✅ Prompts and outputs not stored for training
  • ✅ Data retention limited and configurable
  • ✅ Encryption at rest and in transit
  • ✅ Incident notification obligations

3. Real examples of this model

These are typical implementations, not endorsements:

  • Microsoft Azure OpenAI (EU tenant + DPA + no-training clause)
  • Google Vertex AI (EU regions + sovereign configuration)
  • EU-hosted Claude via enterprise contracts
  • National AI platforms integrated into PSN (Italy)

What makes them “approved” is:

The legal + technical perimeter, not the brand.

4. Additional constraint introduced by the EU AI Office (AI Act)

Under the AI Act, the EU AI Office will publish and maintain a:

List of GPAI models with “systemic risk” (systemic high-risk models)

These are frontier models exceeding capability/compute thresholds and subject to extra regulatory obligations at provider level (robustness, evaluations, incident reporting, cybersecurity, transparency).

Impact on an Italian PA:

Before adopting any GPAI, the PA must now verify:

  1. ✅ Whether the model appears on the EU AI Office systemic-risk list
  2. ✅ If yes, that the provider is fully compliant with the AI Act GPAI obligations
  3. ✅ That this status is explicitly reflected in:
    • DPIA
    • Security assessment
    • Contractual documentation

If the provider is not compliant, the model becomes legally unsafe to adopt, even if hosted in the EU.

This adds a mandatory model-level regulatory checkpoint on top of:

  • GDPR
  • Cybersecurity
  • Procurement
  • Sovereign cloud controls

5. Mandatory governance steps for an Italian PA

To activate such an environment, the PA must do all of the following:

  1. Define use case and data classification
  2. Perform DPIA
  3. Verify AI Act GPAI / systemic-risk status via EU AI Office
  4. Obtain DPO approval
  5. Cybersecurity risk assessment (ACN / ISO 27001 alignment)
  6. Procurement via public tender or framework
  7. Contract with:
    • Data Processing Agreement (DPA)
    • No-training clause
    • EU-only processing
    • Liability and breach clauses
  8. Operational monitoring + logging

Skipping even one step = non-compliant.

6. What this enables (and what it does not)

✅ Permitted:

  • Document drafting
  • Internal decision support
  • Citizen service automation
  • Code generation
  • Data summarization

❌ Still forbidden:

  • Feeding classified data without clearance
  • Automated legal decisions without human validation
  • Biometric or high-risk profiling without AI Act safeguards
  • Using public ChatGPT UI with institutional data

Bottom line

“EU-approved GPAI in a controlled environment” means using powerful commercial AI under full legal, infrastructural, cybersecurity, and AI-Act regulatory sovereignty.
It is not innovation freedom. It is regulated industrial deployment.