In a remote clinic in Nigeria, a nurse speaks into a tablet, and an AI system instantly transcribes the patient’s symptoms — no internet required. Meanwhile, in a Kenyan hospital, a digital assistant combs through an electronic medical record via the cloud and analyses a patient’s history within seconds.

We’ve reached a pivotal challenge in the age of AI: how to bring cutting-edge technology to all settings without leaving the least-connected behind.

Palladium’s experience implementing digital health projects across Africa shows that bridging this digital divide is possible when design and deployment are guided by conscious choices that ensure AI elevates everyone, from rural medical dispensaries to urban health centres.

The Digital Divide in the AI Era

The digital divide has shifted in the age of AI. While it once described gaps in access to technology, today’s divide is increasingly about connectivity, especially with the rise of large language models (LLMs). Advanced AI tools like LLMs and cloud-based analytics require stable internet and capable devices.

In addition to connectivity challenges, offline-capable solutions are increasingly important for protecting local data and supporting national data sovereignty. Minimising the need to transmit sensitive information to external servers or large language models helps ensure that data remains within country borders and under local control. But still the challenge remains: in clinics and health posts with limited connectivity, even the most basic LLMs are out of reach, regardless of local data or hardware.

“For LLM-based applications, the real divide is connected versus not connected, less about data,” explains Jonathan Friedman, Palladium’s Data Science Director. “In well-connected environments, these technologies can be transformative. If not connected, then essentially you are cut off, even from small, open-source models.”

Unlike traditional machine learning, which can often run offline on smaller models, LLMs demand infrastructure that many places in the world still lack. This means bridging the digital divide now depends just as much on expanding connectivity as on building vast datasets.

Rawlsian Lens: Maximising the Minimum

Friedman and his teams navigate these challenges to bring cutting-edge analytics to sites facing connectivity constraints. One way they approach this work is through an ethical lens inspired by philosopher John Rawls. His “veil of ignorance” thought experiment asks: How would we design solutions if we didn’t know which clinics have connectivity?

Rawls’ maximin principle guides us to prioritise the least-resourced sites, delivering value to underserved clinics while enabling advanced sites to benefit from innovation. In context, that means measuring our AI initiatives by how well they uplift the least-resourced clinics first, not just by the excitement they generate at the best hospitals.

This balanced strategy ensures equity and progress, avoiding extremes that either limit potential or leave communities behind.

AI Solutions Tailored to Context

We spoke to Friedman about how he and teams across Palladium apply the Rawlsian lens, and he shared several practical examples.

1. Voice-to-Text in Nigeria: Offline AI for All

In Nigeria, many clinics operate without reliable internet, making cloud-based AI impractical. Palladium addressed this by creating an offline voice-to-text tool that transcribes patient notes locally in English and regional languages as part of the Data for Implementation project.

Optimised for low-power devices and trained on local accents, it improves documentation accuracy and saves time for health workers. Designing for offline environments allows even the most remote clinics to benefit from AI, while laying the groundwork for future enhancements as infrastructure develops

2. Agentic AI in Kenya: Cloud-Enhanced Innovation

Palladium’s work in Kenya builds on years of deploying offline machine learning models at scale.

Now, AI agents integrated into national health systems review outpatient encounters during outpatient visits, suggesting pathways for investigating potential cases of priority public health diseases under Integrated Disease Surveillance and Response. These agents use natural language queries and cloud-based analytics, enabling clinicians to access insights quickly.

While connectivity limits adoption to better-resourced sites, this approach complements offline tools and demonstrates the value of advanced AI. “We don't just design for the lowest common denominator and we don't withhold capabilities from better-resourced sites – we try to develop for both,” Friedman adds.

Piloting these tools in connected facilities not only showcases the potential of advanced AI but also encourages investment in broader infrastructure, illustrating how innovation can coexist with inclusivity.

3. Multi-Tiered AI Architecture

In practice, Palladium often develops both offline and cloud versions of a solution. The voice-to-text rollout is a good example: an offline model has been deployed, and a cloud-based version using a more powerful engine via AWS is being explored for connected sites. Similarly, machine learning models for predicting patient no-shows or medicine stock-outs run locally but can be enhanced centrally when data pipelines exist.

A tiered approach not only provides offline clinics with context-appropriate solutions that can be upgraded as infrastructure develops, but also allows sensitive health data to remain local, reducing reliance on external platforms and supporting compliance with national data protection laws.The guiding philosophy is flexibility: AI solutions must bend to fit the environment, instead of demanding the environment’s conformity. By designing for both constraints and abundance, these solutions deliver immediate value while preparing for future upgrades.

Responsible Innovation and Inclusive Design

Bridging the digital divide with AI will require deliberate, context-aware choices. Instead of adopting a one-size-fits-all strategy, solutions must be flexible, working in both offline and connected environments.

This approach considers real-world constraints like unreliable power, low-end devices, and limited training, while introducing advanced capabilities where infrastructure allows. Each solution is assessed for its ability to benefit underserved clinics and create pathways for future adoption. As Friedman observes, the work is “an ongoing process of learning and adaptation, making sure our solutions remain relevant as both technology and local needs evolve.”

Prioritising equity and adaptability enables AI to strengthen health systems and accelerate progress toward a more inclusive digital future. The flexibility to deploy solutions both offline and online extends the benefits of AI to a wider audience, ensuring that innovation reaches clinics, health workers, and patients no matter where they are.