Compliance

AI and data protection in Madagascar: what Law 2014-038 means

Controllers, processors, transfers abroad, automated decisions: what Madagascar's personal data law means for your AI projects.

7 min read ORIA, Antananarivo

An assistant that searches your customer files, a tool that reads your invoices, a model that flags unusual behaviour: most artificial intelligence projects handle personal data at some point. In Madagascar, the reference text is Law No. 2014-038 on the protection of personal data, adopted by the National Assembly on 16 December 2014 and promulgated on 9 January 2015.

The law was not written with AI in mind, but it applies fully to your projects. Here is what it means in practice, and the questions to ask before launching a pilot. Quotations below are our own translations of the official French text.

What the law covers

The law applies to processing of personal data, automated or not, carried out wholly or partly in Madagascar (Article 5), in particular where the controller is established there or uses processing means located there (Article 6).

Personal data means any information relating to an identified or identifiable natural person, directly or indirectly (Article 7): names in contracts, customer numbers, HR files, call recordings, scanned ID documents. And "processing" is defined very broadly (Article 8): indexing documents for an assistant or extracting fields from a form both qualify.

Controller or processor: who does what?

The controller (responsable du traitement) decides to create the processing and determines its purposes and means (Article 9). In an AI project run on your behalf, that is normally your organisation.

The processor (sous-traitant) processes data on the controller's behalf and according to its instructions (Article 10). That is the usual role of an AI integrator or hosting provider. Article 16 frames the relationship:

  • the processor acts only on the controller's instructions;
  • it provides sufficient guarantees of security and confidentiality;
  • the controller remains responsible for ensuring those measures are respected;
  • the contract sets out the processor's security and confidentiality obligations.

In practice, it is up to you to define the purpose, complete the formalities and inform the people concerned. The AI model provider, if there is one, also acts on your behalf and must appear in the contractual chain.

The principles that weigh most on AI projects

Purpose and reuse

Data must be collected for specified, explicit and legitimate purposes, and may not be reused for other purposes without the person's consent or a legal provision (Article 14). Data collected to manage contracts does not automatically become available to train a scoring model.

Minimisation

Data must be "adequate, relevant and not excessive" in relation to the purposes (Article 14). In practice: index only the documents the use case needs, send the model only the excerpts required, and pseudonymise where possible.

Accuracy and retention

Data must be accurate, kept up to date and retained no longer than necessary (Article 14). Indexes, logs, caches and chat histories are copies, so they need their own retention periods.

Legal basis and sensitive data

All processing requires consent or another condition listed in Article 17, such as a legal obligation, performance of a contract or legitimate interest. Sensitive data, such as health, biometric data or religious beliefs, may not be processed in principle, except under specific exceptions (Article 18); HR files and insurance claims often contain it.

Automated decisions

This is probably the provision that matters most for AI. Article 3 states:

"No administrative or private decision involving an assessment of human behaviour may be based solely on computer processing of data intended to define the profile of the person concerned or to assess certain aspects of their personality."

CV screening, customer scoring, fraud detection: AI can prepare, but the decision must remain human. Article 23 also gives individuals the right to obtain information allowing them to understand and challenge the logic of automated processing when a decision based on it has legal effects on them.

Security: precautions proportionate to the risks

The controller must take "all useful precautions, having regard to the nature of the data and the risks involved" to protect data against destruction, loss, alteration, disclosure or unauthorised access (Article 15). For an AI project, this typically means:

  • role-based access control, enforced all the way into the AI's answers;
  • encryption of data in transit and at rest;
  • logging of access and queries;
  • controlled deletion of data and indexes at the end of the contract;
  • an incident management procedure, even though the law does not set out a breach-notification duty comparable to the EU GDPR.

Using a model API hosted abroad

The most capable language models are often offered as APIs from servers outside Madagascar. If a request sent to the model contains personal data, that is a transfer to a foreign State, governed by Article 20:

  • it is allowed only if the destination State's legislation provides a similar level of protection, assessed case by case;
  • failing that, the CMIL may authorise it if the controller offers sufficient guarantees, notably appropriate contractual clauses;
  • exceptions exist, such as the unambiguous consent of a duly informed person, or performance of a contract with that person;
  • the recipient may not transfer the data onwards without the agreement of the original controller and of the CMIL.

People must also be informed of transfers (Article 27). You have three levers: reduce what leaves (minimised and, where possible, pseudonymised excerpts); constrain the provider (location, retention, a ban on training its models with your data); or remove the transfer by running open-source models on your own servers. Note that pseudonymised data remains personal data if the person can still be identified.

Formalities, data protection officer and regulator

In the private sector, processing must be declared to the CMIL before it starts (Articles 43 and 45); processing that presents particular risks to rights and freedoms requires prior authorisation (Article 46). The declaration covers, among other things, security measures, transfers to other States and the use of a processor (Article 49).

Every controller appoints a data protection officer, resident in Madagascar, who keeps the register of processing activities (Articles 51 and 53). That appointment exempts the controller from declarations, but not from authorisations (Article 52).

The regulator created by the law, the Commission Malagasy de l'Informatique et des Libertés (CMIL), was a long time in the making. According to the Malagasy press, its members were sworn in in August 2025 and the commission is still building up its services. Practical arrangements (forms, simplified standards, online declarations) may therefore change: check them directly with the CMIL.

The stakes are real. The CMIL can issue a warning, order processing to stop or impose a fine of up to 5% of pre-tax turnover for the last closed financial year (Articles 55 and 59). The law also provides for criminal penalties, for example for processing carried out without the required prior formalities, even through negligence (Article 62).

NDAs and DPAs: what your contracts should cover

A non-disclosure agreement (NDA) should be signed before any sensitive information is shared. The data processing agreement (DPA) then turns Article 16 into verifiable commitments:

  • purposes, categories of data, the controller's instructions;
  • security and confidentiality measures;
  • a list of sub-processors, model providers included, and hosting locations;
  • a ban on reusing the data for another client or to train third-party models;
  • retention periods, and return and deletion at the end of the contract;
  • assistance with data subject requests and incidents, and audit rights.

This is the framework ORIA works within: as a processor under Law No. 2014-038, with an NDA, a DPA and a processing register from the very first project, and a path toward ISO/IEC 27001 (information security) and ISO/IEC 42001 (AI management systems). Details are on our security and compliance page.

Governance checklist before an AI project

  1. Map the data in scope: personal? sensitive? who has access?
  2. Write down the purpose and check the legal basis.
  3. Appoint a data protection officer or declare the processing; check whether authorisation is required.
  4. Minimise: exclude, filter or pseudonymise whatever is not needed.
  5. Choose the architecture according to sensitivity, and analyse transfers abroad.
  6. Sign an NDA and a DPA; list every sub-processor.
  7. Put access rights, encryption and logging in place.
  8. Keep a human in the loop for any decision about an individual.
  9. Inform data subjects and organise how they exercise their rights.
  10. Set retention periods (indexes and logs included), document everything, and review it whenever the system changes.

An AI audit builds this data and risk mapping into the roadmap; each project, such as a document AI assistant, is then designed around these principles.

This article is for general information only and does not constitute legal advice. For a specific situation, consult a lawyer or the CMIL.

Planning an AI project that involves personal data? Let's review it together during a free 45-minute diagnostic call, or assess your position with our AI readiness self-assessment.

Frequently asked questions

Does using an AI API hosted abroad count as a data transfer?
Yes, as soon as personal data is sent to the model: Article 20 of Law No. 2014-038 applies. The destination country must offer a similar level of protection, or the CMIL must authorise the transfer based on sufficient safeguards, or one of the law's exceptions must apply.
Is my AI provider the data controller?
Usually not: it acts as a processor, on your behalf and according to your instructions (Article 10). You remain the controller, and the contract must set out the provider's security and confidentiality obligations (Article 16).
What penalties does Law 2014-038 provide for?
The CMIL can issue a warning, order processing to stop or impose a fine capped at 5% of pre-tax turnover for the last closed financial year. The law also provides for criminal penalties, notably for failing to complete prior formalities or to implement security measures.

Read next

Next step

45 minutes to identify your first win.

A structured conversation with an AI architect — no commitment, no jargon.

Book my free assessment