Rwanda’s AI partnerships and a lesson global digital health has already learned

Within three months, Rwanda’s health system took on three new partners. In December the government signed a bilateral health MOU with the United States. In January the Gates Foundation and OpenAI launched Horizon1000 in Rwanda, starting with AI tools for health workers in more than 50 clinics. In February Anthropic signed a three-year MOU to support the Ministry of Health’s work on cervical cancer, malaria and maternal mortality, and to give government developer teams Claude and Claude Code.

In Tech Policy Press, Javaid Iqbal Sofi asked who is accountable when a private AI company becomes “load-bearing infrastructure” for a government. More agreements have followed since he asked. In May, Anthropic and the Gates Foundation announced a partnership that includes work with health ministries on health-intelligence data. On August 28, Rwanda and the US signed the agreement that puts their MOU into operation.

I’m not arguing that Rwanda should have said no. Rwanda has built its digital health system more deliberately than most governments, and it has good reasons to want AI support in clinics that are short of health workers. Most of the commentary I’ve read weighs these agreements one at a time, asking whether each company can be trusted and whether each deal is fair. Rwanda is the clearest case so far, and the same funders and companies are preparing similar work with other health ministries. Global digital health learned more than a decade ago that a health ministry has to manage its partners together, as parts of one health system, and that lesson applies to all of them.

Global digital health learned the cost of uncoordinated partners in the pilot era

By 2008, dozens of NGOs were testing mobile health programs in Uganda at the same time. Each pilot could be defended on its own terms. Together they overwhelmed the ministry that was meant to coordinate them. In 2012 the Ugandan government halted new eHealth pilots to put in place “stronger eHealth stewardship and regulatory frameworks, to ensure that public and donor funds are used for maximum benefit to the general population.”

Ministries and their partners responded by making the health system the unit of planning. Countries wrote national digital health strategies, set interoperability standards, and asked every partner to fit a national architecture. Rwanda’s own 2018–2023 digital health strategy found that its health information systems had frequently been built on different platforms, which duplicated effort and made data hard to consolidate.

AI partnerships recreate the coordination problem at the level of the model

A pilot brought an app and a database. An AI partnership brings a model that health workers, clinics and government developers build their daily work around, and that changes what a ministry has to coordinate. Anthropic’s documentation commits to at least 60 days’ notice before retiring a publicly released model, and other providers publish similar schedules. A decision-support tool checked against one model version needs checking again against the next, and a three-year MOU is likely to outlast several versions.

Rwanda’s health system now works with two frontier model providers under separate agreements, alongside a US agreement that covers disease surveillance. None of the three announcements I read says how its work relates to the other two. The text of the US MOU was not released when it was signed, and I couldn’t find the text of the Anthropic MOU published anywhere. Health workers, patients and Rwanda’s parliament, who would live with the consequences, cannot read the terms their health system has agreed to.

What health ministries can ask of every AI partner

Instruments for holding AI partners to account already exist. WHO’s 2024 guidance on large multimodal models recommends that governments assign a regulator to assess these models for use in health care, and require independent audits and impact assessments after release. Rwanda’s cabinet approved a National AI Agency in June, and its Ministry of Health set up a digital health technical working group to coordinate partners under its 2018 strategy. The pilot era adds one principle. The ministry sets the terms once, and every partner signs up to them.

First, publish the agreements, as the UK government already does for its AI memorandums on gov.uk.

Second, write continuity into them, covering what happens to a clinical workflow when a model version is retired, when prices change, and when the partnership ends. The evaluation data, prompts and integrations the ministry builds should stay with the ministry and move with it to another provider.

Third, evaluate health effects independently of the vendor, for each model version, and publish the results. That includes performance in Kinyarwanda. Audace Niyonkuru of Digital Umuganda has warned that “deploying AI technologies that do not operate in Kinyarwanda would pose a serious barrier to effective care.”

None of this has to wait for a continental framework. The Gates Foundation, OpenAI and Anthropic can make these terms standard in every health agreement they sign, starting with Rwanda.

Rwanda is not the only government with several AI partners. The UK has signed separate memorandums with Anthropic, OpenAI and Cohere, among others, and Horizon1000 plans to reach clinics in other African countries by 2028. Global digital health spent a decade learning to judge a project by what it leaves the health system able to do. That test applies to AI partners in London as much as in Kigali, and Rwanda, which has applied it to its own digital health system for years, is well placed to show other ministries how.