Executive summary
Three new pilot projects supporting the use of generative AI in public administrations began on 1 July 2026: FLOODS & DROUGHTS, EUNOMIA.AI and EuropAI.
The European Commission states that the projects will pilot trustworthy European generative-AI solutions addressing concrete needs of public authorities and citizens. Participating administrations are expected to procure, test and deploy solutions adapted to their public-service needs.
The pilots create an opportunity to establish an evidence standard before success claims are made. Terms such as simplification, accessibility, efficiency and trustworthy AI are policy objectives. They become demonstrated results only when measured against a transparent baseline.
Core forensic question
What evidence would be required to show that generative AI improved a public service rather than merely introducing a new technical layer?
A credible evaluation must compare performance before and after deployment and distinguish technical capability from administrative outcome.
Minimum baseline
Before deployment, each participating administration should publish or record:
- the service or process being changed;
- current processing time;
- backlog and completion rate;
- error and correction rate;
- staff time per case;
- cost per transaction;
- number and type of complaints;
- accessibility barriers;
- language coverage;
- quality of reasons given to citizens;
- existing human-review arrangements.
Without a baseline, later claims of improvement cannot be independently tested.
Outcome indicators
Efficiency
- median and average processing time;
- backlog reduction;
- staff time saved;
- cost per completed case;
- system downtime and integration costs.
Accuracy and quality
- factual error rate;
- incorrect legal or procedural guidance;
- rate of human correction;
- consistency across comparable cases;
- quality and completeness of reasons.
Rights and safeguards
- proportion of outputs reviewed by a human;
- number of adverse decisions influenced by AI;
- availability and use of appeal or correction mechanisms;
- personal-data incidents;
- discriminatory error patterns;
- documentation of model limitations.
Accessibility
- performance by language;
- accessibility for persons with disabilities;
- success rates for users with low digital literacy;
- availability of non-digital alternatives;
- user comprehension.
Public trust
- user satisfaction separated from legal correctness;
- complaint patterns;
- understanding that AI was involved;
- confidence in human review;
- willingness to rely on the service after errors.
Quantitative-formalism risks
The following indicators may be accurate but insufficient:
- number of AI tools deployed;
- number of documents processed;
- number of users;
- percentage of automated interactions;
- total hours theoretically saved;
- number of participating administrations.
A large volume of AI-assisted activity does not prove that the service became more accurate, fair or accessible.
A central risk is:
Automation rate is reported as improvement without measuring correction, exclusion or remedy.
Procurement and vendor accountability
Because participating administrations will procure solutions, procurement records are part of the expected evidence chain.
The file should show:
- functional and rights-based requirements;
- model and data documentation;
- testing criteria;
- audit rights;
- security and incident obligations;
- rules on subcontractors;
- data retention and reuse;
- exit and portability arrangements;
- ownership of logs and evaluation data;
- responsibility for errors.
A pilot cannot be fully evaluated if the public authority lacks access to the evidence needed to audit the system.
Alternative explanations
A pilot may initially increase processing time because staff must learn the system, verify output and correct integration problems. Early errors do not necessarily show that the technology cannot improve.
Conversely, faster processing may result from procedural simplification or additional staffing rather than from AI. The evaluation should isolate, as far as possible, the contribution of the system.
Different use cases will require different standards. A tool summarising internal documents does not create the same risk as a tool communicating legal guidance or influencing eligibility decisions.
Preliminary finding
The Commission’s announcement identifies aims and project structures, not demonstrated outcomes. The strongest public-interest contribution would be a common evaluation protocol published before full deployment.
Trustworthy public-sector AI should be evidenced through accuracy, rights protection, accessibility, reviewability and real administrative outcomes—not through adoption volume.
Recommended Civic Forensics modules
- Analysis of AI-Assisted Public Administration
- Quantitative Claims Module
- Detector of Quantitative Formalism
- Expected Evidence Trace Plan
- Procurement and Contract Evidence Review
- Institutional Integrity Index
Source
European Commission, New GenAI pilots for public administrations , 22 July 2026: https://digital-strategy.ec.europa.eu/en/news/new-genai-pilots-public-administrations
Limitations
This is a prospective analysis based on the Commission’s launch announcement. It does not evaluate systems that have completed implementation and does not conclude that any pilot is ineffective or non-compliant.