🎯 Intended purpose
Authorised tasks, prohibited tasks and limits on later changes of use.
Public procurement · complete contract
Public Procurement of Judicial AI: Essential Contract Clauses
A poor contract can lock the state into a system it does not understand and cannot leave.
Judicial AI is not an ordinary software purchase
The contract must govern a system that can evolve, depends on data, produces variable errors and may affect fundamental rights. The procurement documents must therefore define authorised uses, prohibited uses, acceptance criteria, human oversight, security, audit rights, continuity and exit arrangements before suppliers submit bids.
Ten essential clause families
🧪 Acceptance
Test datasets, thresholds, rejection criteria and retesting after updates.
🧾 Documentation
Models, versions, data sources, parameters, limitations and change logs.
👤 Human oversight
Competence, available time, stop authority and ability to reject outputs.
🔐 Security
Access control, encryption, logging, vulnerability management and subcontractors.
⚖️ Rights
Information, remedies, impact assessments and non-discrimination testing.
🔍 Audit
Access to evidence needed by the buyer, regulators and authorised independent auditors.
🚨 Incidents
Notification deadlines, minimum content, containment and cooperation duties.
↩️ Reversibility
Export formats, documentation, skills transfer, deletion and migration assistance.
🛟 Continuity
Degraded mode, fallback tools, recovery objectives and regularly tested plans.
Why these clauses must exist before the tender
French public-procurement law allows the buyer to define technical specifications and performance requirements linked to the subject matter of the contract. The French CCAG-TIC already provides a framework for reversibility and transferability. A judicial AI system nevertheless requires additional clauses covering model versions, datasets, error measurement, fundamental rights, audits and incidents.
What a supplier must demonstrate
| Area | Evidence |
|---|---|
| Performance | Reproducible results on agreed test datasets and defined operating conditions. |
| Data | Origin, quality, rights of use, transformations, retention and deletion. |
| Versions | Model history and the ability to return to an approved version. |
| Security | Architecture, tests, access rights and incident-response arrangements. |
| Reversibility | A real exit exercise completed before the contract reaches its final stage. |
Public-interest rule. The buyer must remain capable of understanding, auditing, suspending and leaving the system. A low purchase price cannot justify strategic or technical captivity.
Main official sources
consolidation audit
Category 11 financial audit — social support, transparency and anti-corruption →