Governance-Bound Artificial Intelligence for Evidence-Linked Medicines Reliance Review: A Design Science Research Study

Every year, national health agencies face a mountain of paperwork. When a company wants to sell a new medicine, it must submit a massive dossier containing thousands of pages of data about how the drug was made, how it works in the body, and whether it is safe. In many countries, especially those with limited staff, reviewing these documents takes a long time, delaying access to life-saving treatments. To speed things up without cutting corners, regulators have adopted a strategy called "reliance." Instead of starting from scratch, a local agency can look at the work already done by a trusted authority in another country. If that trusted authority has already approved the medicine, the local team can use that assessment as a foundation. However, this process is still incredibly difficult. The local team must still read through the original documents, check that the product is exactly the same, find any small differences, and write a new report that explains their decision. This requires navigating complex structures, spotting tiny errors hidden in text, and ensuring that every conclusion is backed by specific evidence. If a mistake is made, or if the reasoning cannot be traced back to a source document, the entire decision could be challenged, putting public health at risk.

Researchers at the University of the Western Cape in South Africa have tackled this problem by designing a new kind of computer tool, not to replace the human experts, but to act as a highly organized assistant. They call this tool RAiA. The team did not simply build a program to summarize text, because a summary might miss the tiny details that matter in medicine. Instead, they used a method called design science research, which focuses on building a specific solution to a real-world problem and defining exactly how it should work before it is ever tested on real data. Their goal was to create a system that could handle the heavy lifting of reading and organizing the documents while keeping the human expert firmly in charge of the final decision. The result is a prototype system that acts as a bridge between the chaotic pile of documents and the clear, defensible report a regulator needs to write.

The core of this design is a set of strict rules that govern how the computer behaves. The researchers realized that for a tool to be useful in a high-stakes environment like medicine, it must be "governance-bound." This means the system is built with safety rails that prevent it from making its own decisions or hiding its work. For example, the system is designed so that it cannot approve or reject a medicine on its own; it can only highlight potential issues and suggest connections between different parts of the paperwork. Every time the computer points out a fact, it must also show exactly where that fact came from, linking the finding to a specific sentence in the original document. This ensures that if a human reviewer questions the computer's work, they can immediately see the source and verify it. The system also keeps a detailed log of every step it takes, creating a paper trail that proves the work was done correctly and securely.

To build this, the team broke the massive task of reviewing a medicine dossier into smaller, manageable pieces. They designed the system to treat different types of information separately. One part of the system looks only at the chemical and manufacturing details, another focuses on clinical trial results, and a third checks the labeling and instructions. This is similar to how a team of specialists might work, where one person checks the engine of a car while another checks the brakes, rather than having one person try to do everything at once. The system gathers information from these different areas and brings them together, looking for contradictions or missing pieces. If the manufacturing report says the drug was made in one factory, but the clinical trial report implies it was made in another, the system flags this discrepancy for the human to investigate. Crucially, the system is programmed to be conservative; it is better to flag a potential problem that turns out to be nothing than to miss a real danger.

The researchers also paid close attention to how the human expert interacts with the machine. The interface is designed so that the human can easily accept, edit, or reject what the computer suggests. If the computer makes a mistake, the human can correct it, and the system records why the correction was made. This feedback is not used to instantly change the computer's brain, which could lead to unpredictable behavior, but is instead saved for a later, careful review by the team. This separation ensures that the tool remains stable and reliable. The system also protects sensitive information, such as trade secrets or private data, by encrypting it and ensuring that only authorized people can see it. The computer is not allowed to use this private data to teach itself, which prevents the risk of that information leaking out later.

The team built a working version of this system, known as a prototype, to show how the process would flow from start to finish. They tested it using made-up documents that looked like real medicine applications, but they did not use any actual confidential data from real companies. The demonstration showed that the system could take a large file, sort it into the correct sections, find relevant evidence, and draft a report that the human expert could then review and sign off on. The human expert remained the final authority at every step, deciding what the evidence meant and whether the medicine was safe. The prototype proved that the design was possible and that the flow of work made sense, but the researchers are careful to state that this is not yet a finished product ready for real-world use. They have not yet measured how much faster it makes the process or how accurately it finds errors in real documents.

The main contribution of this work is not a new type of computer algorithm, but a blueprint for how to build trustworthy tools for medicine regulation. The researchers have shown that it is possible to design an artificial intelligence system that respects the rules of governance, protects private data, and keeps humans in control. They have provided a set of ten guiding principles that other developers can follow if they want to build similar tools. These principles emphasize that the tool must always link its findings to evidence, must never make a decision on its own, and must be designed so that its work can be audited and checked. By focusing on these rules from the very beginning, the team has created a model that could help regulators in Africa and around the world manage the growing complexity of medicine approvals without compromising safety.

This study represents a significant step forward in thinking about how technology can support human experts in critical fields. It moves away from the idea that artificial intelligence should replace human judgment and instead focuses on how it can make human judgment more informed and efficient. The researchers have demonstrated that with the right design, a computer can handle the tedious work of sorting and checking documents, freeing up the human experts to focus on the complex reasoning and final decisions that protect public health. While the tool is still in the prototype stage and needs further testing, the design itself offers a clear path forward for creating safe, reliable, and accountable systems for the future of medicine regulation.