What African governments and their partners can do to use AI for better public service delivery

AI is already changing every industry and impacting the lives of billions, and African governments are under growing pressure to adopt AI tools. How can governments use AI and the data they already have to improve public services now, while building the governance foundations to do this responsibly at scale?

That question kept coming up at this year’s UN General Assembly, where people from around the world pressed on sovereignty, AI’s place in government systems, and how AI can close the inequality gap. IDinsight and the UK Foreign, Commonwealth & Development Office co-hosted a side event at UNGA 81 on AI and public service delivery in Africa, which covered a lot of these friction points.

Becca Gong Sharp, IDinsight’s CEO, and Helen King, UK Ambassador to the UN Economic and Social Council, opened the discussion. Meg Battle then moderated a panel with Seyni Ndiaye Fall (Senegal’s Ministry of National Education), Dr. Divya Nair (Last Mile Health), Clio Dintilhac (Gates Foundation), and Charlene Migwe (IDinsight).

The last couple of years have shown that AI could potentially change how our sector improves lives. Nothing guarantees it will, though, and there’s a real risk of serious harm along the way, even when no one intends it.

“Right now, AI governance and policy is wet clay. It can still be shaped. But clay dries, and the choices leaders make about AI today could shape public service delivery for generations.”

 

– Becca Gong Sharp, IDinsight

Here are some key takeaways from the event:

1. Governments need to be in the driver’s seat

AI tools that scale are built with governments from the start.

Seyni Ndiaye Fall of Senegal’s Ministry of National Education described a scene many government decision-makers will know. A partner walks into a government office with a finished product and funding already agreed with a donor, and asks the Ministry to sign. “When the money goes directly to the company, the Ministry is not part of the contract,” he said. “We now ask to be a part of it.”

Senegal now sets its own conditions before an AI tool reaches its schools. The first is sovereignty, but what Seyni called “realistic sovereignty”:

“Today, governments cannot protect all data in the same way. But if we wait until everything is perfect, we will never start. So, we identify our most sensitive data—student data, teacher data, and our national student ID. We protect this data and move forward with the rest.”

The second is the Ministry’s own capacity to judge whether the tool works. Before any expansion, Ministry staff review the tool’s answers themselves. “No one outside Senegal should decide what is correct in Pulaar,” he said. “Pulaar is a national language.” Seyni acknowledged that this took time. The Ministry also kept the tool narrow on purpose: “It supports teachers; it does not make decisions for them.”

“What almost stopped us was not the technology. It was the fear of losing control. Our answer was simple: set clear limits, check the results, and move forward.”

 

– Seyni Ndiaye Fall, Senegal Ministry of Education

Dr. Divya Nair of Last Mile Health described a similar approach in Ethiopia’s health system. HEP Assist is a decision-support tool LMH and IDinsight built with Ethiopia’s Ministry of Health. It connects community health workers to clinical supervisors, who use the AI tool, trained on Ministry protocols, to talk them through decisions. Eventually, Ethiopia’s 40,000 community health workers will have direct access to the AI tool. Working with the government allowed LMH to test this at a large scale, provided agility to be able to iterate and pivot, and helped government partners balance a key tension: the pressure is to be good and fast simultaneously.

“As a result of all of this, I believe this tool is scaling responsibly, with all of us working together, towards a shared goal. Our advantage was never the tool but that we are building with and for the government system.”

 

– Dr. Divya Nair, Last Mile Health

2. Evaluation as a tool for sovereignty

Charlene Migwe gave the sovereignty debate a new angle. “Data control matters,” she said, “but real sovereignty also means a government can sit in a room and say, ‘No, that model won’t work for our citizens,’ with evidence that shows why.” That takes investment in people who can test models and products, and ways to evaluate a policy or intervention at every stage of a rollout. Without it, governments are left taking vendors at their word.

Charlene pointed to one practical answer: the AI Evaluation Playbook, which IDinsight developed with the Center for Global Development and the Agency Fund. It breaks evaluation into four levels: the model, the product, the user, and the impact. 

“A year ago we were hoping people would care about that framework. Now governments are asking for it. They’re tired of the hype and want practical, step-by-step ways to measure what these tools do.”

 

– Charlene Migwe, IDinsight 

3. Language is a lever to bridge access gaps

Language came up from nearly every angle. For Seyni Ndiaye Fall, it’s the heart of the project: the leading models work in English and French, but not well enough yet in Wolof, Pulaar, or Seereer. Dr. Nair’s team hit the same wall in Ethiopia. They tested a range of open-weight and closed models, and some returned gibberish in the languages community health workers actually speak. 

This is emerging as a clear priority across the sector. The day before the panel, Gates Foundation convened 60 organizations, including frontier AI labs, Senegal’s Ministry of Telecommunications and Digital Affairs, the Ethiopia AI Institute, and FCDO, behind a joint commitment. Within five years, the estimated 3.4 billion people whose languages are underrepresented in today’s AI models should be able to use AI tools in their own language and voice. Clio Dintilhac pointed to the Gates Foundation’s pledge of at least $1 billion over two years for AI equity, with 40% going to education.

“AI can be a great equalizer or push in exactly the opposite direction, and what we build in the next 12 to 18 months will decide which.”

 

– Clio Dintilhac, Gates Foundation

4. AI investments must start and end with people

To invest in AI, governments and partners need to invest in the people and communities it’s meant to serve. People need both the skills and the access to shape how AI works for them.

Many people in LMICs that AI is meant to serve have little or no access to digital tools. The most disenfranchised among these populations face many layers of inequality. For AI tools to be able to bridge the gap, the impact must reach people’s lives offline. Divya noted that most of the community health workers Last Mile Health supports use feature phones rather than smartphones, so the next version of HEP Assist has to work in low-resource settings. Clio described how Senegal chose to focus its AI use cases on three languages spoken by much of the bottom half of students, with results broken down by gender.

“Success will not be measured by the sophistication of the technology but by whether it improves outcomes for real people.”

 

– Helen King, UK FCDO

Another level is the people who build and run these tools. “You need people from the country, from the community, to design the solutions that work for them,” Clio said. AI is still new, and few people have experience deploying it in government. That makes it urgent to build skills inside government. Ministries need their own staff who can test, adapt, and run these tools, so ownership stays with them after partners leave.

5. Balancing ambition for scale with speed and thoughtfulness

The pressure to move fast is coming from several directions at once. A new wave of funding looking for places to land, and governments eager to lead in their regions. Health workers, Divya noted, want to move faster than the technology currently allows. The panelists agreed that fast and safe can coexist, but it takes deliberate work.

Divya’s call to action was aimed at funders: invest in the trust. “It’s tough, and it takes time.” She pointed to how helpful it was when the Gates Foundation encouraged her team to slow down while they pivoted HEP Assist. And when some countries move slower on AI, she said, there’s usually a reason. They’ll ask for it when they’re ready.

Charlene Migwe argued for depth over breadth, since there’s still so much to learn about how these tools work in specific places and sectors. She also pointed to shared tools that help everyone scale, like the knowledge graphs IDinsight’s data science team is building to give AI tools the context of national education curricula. Any government or innovator can plug into them.

Clio Dintilhac advised starting with the problem communities face, then work back to the technology. If there’s one thing to carry out of the room, it’s that. Move as fast as trust and evidence allow, and no faster.

None of the panelists claimed to have a finished roadmap, but the same picture kept coming through across health and education, and from governments, funders, and implementers alike. Governments should set the terms, evidence checks the claims, and partners should move at the speed of trust. That conversation doesn’t end at UNGA, and we plan to keep it going with governments, funders, and partners.

Want to continue the conversation? Reach out to Meg Battle at meg.battle@idinsight.org or Charlene Migwe at charlene.migwe@idinsight.org.