Procurement as the Invisible Bottleneck in Global Health

Discussions around AI in global health have largely focused on diagnostics and drug discovery. Important progress is now also being made in clinical decision-making, and the recent launch of the Horizon 1000 initiative by the Gates Foundation and OpenAI marks a meaningful step toward unlocking the use of AI and large language models in everyday care. The next major challenge to address is procurement.

In low and middle-income countries, procurement is the primary vector by which access, affordability, and continuity of care are shaped. Fragmentation, opacity, and manual coordination are not minor inefficiencies. They are structural fault lines along which patients downstream fracture and suffer.

Thousands of procurement units, hospitals, and pharmacies continue to operate in isolation, each navigating suppliers, logistics providers, and funders through bespoke and largely manual processes.
 
When these systems cannot scale or coordinate, investment elsewhere in the health system delivers diminishing returns. 

Procurement remains invisible precisely because its failures are distributed, yet it ultimately determines what medicines are available, when they arrive, and at what cost.

Solving this requires a fundamental shift in how procurement systems process information and make decisions, one that AI capabilities are now positioned to enable. 

Digital foundations already change outcomes
Before advanced AI enters the picture, basic digitization already creates meaningful impact. Automating core workflows, standardizing processes, and introducing real-time transparency to amplify human capacity. Procurement professionals can oversee more volume and complexity without sacrificing quality. This shift alone reduces fragmentation and lowers the cost of participation for suppliers, as we have already demonstrated in just a few months of marketplace operations at Axmed. A similar success story is found at the Global Family Planning Visibility and Analytics Network, where aggregation of supply chain data from multiple sources has provided real-time insights and accelerated decision-making.

These digital foundations are not optional. They represent the first transition from reactive procurement to controlled execution and create conditions under which intelligence can later emerge.

AI as connective tissue, not procedural replacement
AI delivers its greatest value not by replacing existing systems, but by connecting them. In environments shaped by heterogeneous tools and limited technical capacity, forcing procedural redesign or deep integrations often fails ([citation needed]). This is where agentic systems bring the maximum return and act as translation layers across emails, spreadsheets, ERPs, and procurement platforms. We have proven this with Haystack, our proprietary AI-matching model, which scanned more than 20,000 individual medicine requests since its launch in 2025.

By reconciling fragmented inputs, escalating exceptions, and surfacing decision-ready signals, AI reduces coordination cost and cognitive load. This matters because change management, not model performance, is the real constraint. AI succeeds when it adapts to existing ways of working and preserves human judgment while absorbing friction.

From operational scale to operational intelligence
As procurement systems scale digitally, they generate structured operational data that can be reused. Over time, this enables agents to identify bottlenecks, detect anomalies, and prioritize intervention points continuously rather than episodically. Scale without intelligence creates fragility. Scale with intelligence compounds resilience.

The transition from operational scale to operational intelligence requires a deliberate division of labor between two tightly coupled classes of agents. The first, now being prototyped at Axmed, is observational and operational in nature. These agents sit alongside procurement workflows and monitor key execution signals such as demand volumes, historical pricing, supplier responsiveness, pipeline bottlenecks, and processing delays. Their immediate value is acceleration: reducing latency, surfacing exceptions earlier, and supporting human decision-makers by escalating key insights without altering existing workflows. The second class is anticipatory and can only emerge once this operational layer has accumulated sufficient historical signal. These agents introduce predictive intelligence into procurement, inferring future demand from consumption patterns, anticipating supplier response times and supply chain delays, and enabling intervention before disruptions become visible. Together, these layers form the connective fabric of an effective procurement system: one that executes reliably in the present while progressively learning to act earlier and with greater precision.

The unresolved constraint: data gravity
Much upstream demand and consumption data remains manual or non-existent. AI cannot predict what has never been captured. Closing this gap will require last-mile and ad hoc digitization solutions that accelerate data capture where infrastructure is weakest. Efforts in this area are numerous, but success has yet to be demonstrated at scale. One example offers a useful reference: vaccine delivery pilots in rural Uganda highlighted persistent stock-outs and unreliable distribution, and showed how procurement-driven, coordinated effort resulted in a cost-effective strategy that improved last-mile vaccine delivery.

Technology’s role here is pragmatic: reduce friction, surface early signals, and enable real-time intervention.  

Large language models are beginning to lower the cost of digitizing and interpreting unstructured data, but their impact will depend on whether they are embedded into procurement systems rather than deployed in isolation.

Disciplined optimism
AI will not fix procurement overnight. But with upfront investment in digital foundations and data infrastructure, it can fundamentally reshape how procurement systems coordinate, learn, and scale. 

What is required now is collaborative courage. The data needed to generate procurement intelligence already exists, fragmented across manufacturers, distributors, procurers, logistics providers, and funders, locked behind institutional silos and rigid contractual frameworks. The cost of opening access to this data is modest relative to the budgets deployed across global health, yet progress remains slow because collaboration is treated as a risk to be minimized rather than a capability to be built.

Too many data-sharing efforts are encumbered by ironclad agreements that prioritize control over speed, delaying insight until it is no longer actionable. This inertia undermines the very resilience these systems are meant to create. Intelligence cannot emerge if data remains static.

The real opportunity lies in making procurement data open by default, shared across actors, and governed pragmatically with strong safeguards where patient data is involved. Only then can information flow through agentic systems, compound over time, and translate into anticipatory action.

The opportunity is real, but it must be built deliberately.