Challenge

How can we help city councils quickly and reliably anonymize the minutes of municipal plenary sessions to protect personal data, facilitate their publication and guarantee compliance with regulations?

Context and problem

The AOC provides digital services for local administrations that often involve the publication of administrative documents with potentially sensitive content. In this context, the anonymization of personal data before publication is a key element to ensure compliance with data protection regulations and reduce risks to individuals' privacy.

One of the most representaThese are the minutes of municipal plenary sessions, which local entities must proactively publish on the portals of transparència, in accordance with Law 19/2014, of transparència, access to public information and good governance. Only in 2024 were published more than 8.000 plenary sessions to the portal of transparència of the AOC.

These minutes may contain personal data that is not subject to publication, such as names of individuals, ID cards, addresses, emails or bank details. Their review requires distinguishing what information should be kept public and what should be hidden to protect the rights of affected individualsThis decision depends not only on the type of data, but also on the context in which it appears within the document. It is a task that requires time, knowledge of the regulations and careful legal criteria.

For example, the name of an elected official in a plenary session does not require anonymization, while the name of a person affected in a sanctioning resolution does. Similarly, a date of birth is a personal data item that must usually be anonymized, while the date of an administrative action is part of the information that must be kept public.

When this process is done manually, increases the workload of the responsible teams and also the risk that some personal data is exposed involuntarily, with the consequences that this may entail in terms of compliance with the General Data Protection Regulation (GDPR). Furthermore, when it is detected that personal data has been improperly published, the document must be withdrawn, corrected and a new version published, a slow procedure that the AOC must manage practically every week.

In this context, the AOC considered in 2025 whether artificial intelligence technologies could support this process, helping municipal staff identify personal data that needs to be anonymized and always maintaining human supervision over the final decision.

Proposed solution

To respond to this problem, the AOC has launched a proof of concept (PoC) to validate whether generative AI can automate the anonymization of the minutes of municipal plenary sessions before their publication on the portal. transparència.

The initiative explores how to combine the ability of generative AI to interpret the context of documents with human supervision, with the aim of reducing the workload associated with this process, minimizing the risks of personal data disclosure and ensuring compliance with data protection regulations.

The objective is not to replace the legal criteria of the administrations, but offer a support tool that automatically detects personal data and proposes anonymization that is subsequently validated by a person before publishing the document. This approach allows you to take advantage of the advantages of AI while maintaining human control over the final decision.

The proof of concept has been designed as a minimal risk system: AI does not make automated decisions about people or affect their rights, but acts solely as a support tool to transform documents before publication.

What is the solution?

The proof of concept consists of a cloud web platform that uses generative AI models to identify and anonymize personal data contained in plenary sessions.

Unlike traditional systems based solely on predefined patterns, the platform interprets the context of the document to distinguish between information that should be kept public—such as that relating to elected officials or institutional activity—and that which should be protected because it corresponds to personal data. This ability is especially relevant because the decision to anonymize information often depends on the context in which it appears and not solely on the type of data.

The solution incorporates several mechanisms to guarantee the quality and reliability of the process:

  • Contextual detection of personal data, based on generative AI.
  • Whitelists, which allow certain terms to be excluded from anonymization to avoid over-anonymization.
  • Human supervision (“human-in-the-loop”), so the system only proposes the anonymizations and it is the user who reviews and validates them before publishing the document.
  • Traceability and explainability, recording the criteria applied in each anonymization so that the process is reviewable and auditable.

Com funciona?

The anonymization process is structured in four simple steps:

  1. The user uploads the document (PDF or DOCX) to the platform.
  2. Select the anonymization criteria you want to apply (names and surnames, ID, addresses, dates, telephone numbers, bank accounts, etc.) and start the process.
  3. The system analyzes the document, interprets the context of the information and generates a proposal for an anonymized document. To do this, it identifies the personal data that, according to the selected criteria, should be anonymized and hides them using masks.
  4. The user reviews the result, accepts or modifies the generated masks, can create new ones if they detect personal data that has not been anonymized and, finally, downloads the document prepared for publication.

Proof of concept results

The proof of concept has been validated with 52 documents provided by 9 municipalities, complemented by multiple internal tests carried out by the AOC during the development of the solution. The validation process has allowed to evaluate both the quality of the anonymization and the behavior of the platform in real use situations.

The results obtained confirm the technical and functional feasibility of the proposal, as well as the adequacy of the model based on the combination of generative artificial intelligence and human supervision. Specifically, the validation has allowed us to verify that:

  • The platform achieves an overall accuracy of 96%, that is, when the system identifies a piece of data to anonymize, it usually does so correctly, reducing the need to review false detections.
  • The overall F1 indicator, which combines the accuracy and coverage of personal data detection into a single quality metric, exceeds 90%. In the lasteres PoC iterations have achieved overall quality levels above 95%.
  • The combination of AI and human supervision It is an effective model to reduce manual effort while maintaining the necessary guarantees before publishing documents.

The proof of concept has also served to identify the main challenges that need to be addressed before an eventual deployment in production:

  • Increase detection coverage (Recall) to reduce the cases in which some personal data may remain unanonymized, especially in documents with complex or heterogeneous structures.
  • Evolve the solution architecture to improve its scalability, performance and efficiency, making a sustainable shared service possible for a large number of administrations.
  • Continue training and refining the models based on the results obtained and new validations with real documents, incorporating the knowledge generated during the user review.

In this sense, the PoC has also allowed us to define how the architecture of the solution should evolve. The current version solves the entire anonymization process using a single high-capacity generative AI model, an adequate approach to validate the viability of the solution but not very efficient for large-scale deployment. The expected evolution separates the identification of personal data from the decision on which data should be anonymized based on context, so that each phase can use the most appropriate and efficient AI models. This evolution will improve the scalability, performance and sustainability of the future shared service.

Overall, the proof of concept confirms that this approach is viable and provides a solid foundation to continue evolving the solution towards future production deployment.

Resources and collaboration

The proof of concept has been developed jointly between the AOC, responsible for the functional leadership of the project; the technology company OMNIOS, in charge of developing the platform and analyzing the results; and 9 city councils, which have provided real documents to validate the solution in an environment. representause of.

Since its design, the platform has been conceived in accordance with the security principles, transparència, traceability and human supervision specific to AI systems applied to the public sector, with the aim of exploring an anonymization model assisted by artificial intelligence that can evolve towards a future shared service for local administrations.

Value for local entities

The proof of concept highlights the potential of generative AI to support administrative processes that require a high burden of document review and careful legal judgment. If the solution evolves into a shared service, it could bring significant benefits to local authorities:

  • Reducing the risk of accidental disclosure of personal data, minimizing the likelihood of publishing information that should not be publicly accessible.
  • Saving time and improving efficiency, automating much of the data identification process personal and reducing manual review.
  • Greater homogeneity in the application of anonymization criteria, promoting more coherent treatment between administrations.
  • More guarantees of traceability and auditing, thanks to the registration of the anonymizations carried out and the criteria applied.
  • Ability to continuously improve the solution, incorporating user input and advances in increasingly powerful AI models.

Conclusion

The proof of concept has allowed the technical and functional feasibility of a tool based on generative artificial intelligence to support the anonymization of the minutes of municipal plenary sessions before their publication to be validated.

The results obtained show that the combination of automatic detection and human supervision is an appropriate approach to reduce the workload of local entities and reduce the risk of disclosure of personal data, while maintaining the necessary guarantees in a process of this nature.

The PoC has also allowed the identification of the main areas for improvement before a possible production deployment, especially in the detection of certain types of personal data in complex documents and in the integration of the solution with the AOC services. The PoC provides a solid basis to continue evolving the project and assess its transformation into a future shared service for local administrations.

Status of the project

PoC carried out and potential validated. The continuity of the initiative is in the process of being contracted with the realization of a pilot with a wider scope.

More information

AOC Blog:

Presentations and articles:

Videos: