Your AI model still answers to Rwanda’s data protection law

Banks, insurers, lenders and fintech companies in Rwanda are testing artificial intelligence for credit scoring, fraud detection and customer service. Rwanda has had a National Artificial Intelligence Policy since 2023, but a policy is not a statute. There is no law in Rwanda written only for artificial intelligence. That does not leave a loophole. Any model that uses information about people is governed by Law n° 058/2021 of October 13, 2021 relating to the protection of personal data and privacy.
The law came into force on October 15, 2021. Existing businesses were given two years to comply, and that period ended on October 15, 2023. Compliance is now expected in full.
Registration comes before the model
Under Article 29, every data controller and data processor must register with the National Cyber Security Authority, which supervises the law through its data protection office. Article 30 lists what the application must contain, including the purposes of processing, the categories of data, the recipients and the countries to which data may be sent. Under Article 31 the Authority issues a certificate within 30 working days of a complete application, and under Article 32 changes must be reported within 15 working days.
A company that feeds customer data into a model without a registration certificate has a problem before the model produces a single result. The administrative fine for operating without registration is between RWF 2,000,000 and RWF 5,000,000, or 1% of the previous year’s global turnover for a company.
A lawful reason for every use
Article 46 sets out the grounds on which personal data may be processed. They include consent for a specified purpose, performance of a contract with the person, a legal obligation, and legitimate interests that are not overridden by the person’s rights.
Training a model is a use of data in its own right. Data collected to open a bank account was not necessarily collected to train a scoring model. A business should record which ground it relies on for training and which for live decisions. If the answer is consent, the consent must have covered that purpose.
People must be told when a machine decides
Article 42 requires a controller to tell a person when automated decision-making, including profiling, is used. The controller must explain the logic involved, the significance of the processing and the consequences the person can expect. Under Article 21, a person has the right not to be subject to a decision based solely on automated processing where it has legal or other significant consequences. The exceptions are explicit consent, necessity for a contract with that person, and authorisation by law with suitable safeguards.
A lender that refuses a loan because a model said so, and cannot explain why, is on weak ground. A system nobody inside the company can explain is not going to fly. Article 38 also requires a data protection impact assessment where processing is likely to create a high risk, and it names profiling that leads to significant decisions and the use of new technologies.
Sensitive data needs extra care
Health records, biometric data, genetic data and criminal records are sensitive personal data. Article 11 requires additional safeguards for them, such as separate storage, pseudonymisation or encryption.
Where the data sits
This is the point that catches most technology teams. Under Article 50, personal data may be stored outside Rwanda only if the registration certificate authorises it. Under Article 48, a transfer abroad is allowed only in listed cases, such as authorisation from the Authority, the person’s consent or the needs of a contract. Article 49 requires a written contract that sets out the roles of each party.
Most large artificial intelligence services run on servers abroad. Sending a prompt that contains a customer’s name, account number or medical history to such a service is a transfer of personal data. A team that connects a foreign model to live customer records without checking the certificate may be in hot water without knowing it.
Breaches and penalties
Under Article 43, a controller must notify the Authority within 48 hours of becoming aware of a personal data breach, and a processor must notify the controller within the same period. Under Article 44, a detailed report must follow within 72 hours. A model that leaks training data, or a chatbot that reveals one customer’s details to another, is a breach like any other.
Criminal provisions apply to unlawful access, use, sharing and sale of personal data. Under Article 62, a company convicted of an offence may be fined 5% of its previous year’s annual turnover.
A short checklist
Confirm that the registration certificate is valid and covers the purposes and countries involved. List every model in use and the personal data each one touches. Record the legal ground for training and for live use. Write a plain explanation of how each automated decision is reached. Keep a person in the process for decisions with serious effects. Complete the impact assessment before launch. Appoint a data protection officer where Article 40 requires one.
None of this is rocket science. Companies that apply ordinary compliance to this new tool will stay on top of things. Those that treat artificial intelligence as an exception will be the first to hear from the regulator.