The AI story is not the model — it's the millions of interactions

There's a narrative in health tech that goes something like this: We have a machine learning model. It's trained on historical data. It makes predictions. That's AI. That's the value.

This narrative is wrong, and it's particularly wrong for health systems.

In reality, the value of machine learning in health supply chains isn't in the sophistication of the model. It's in the loops you build around how people actually use the system, and how the system learns from those interactions.

The prophecy of historical data

When we first built ForLab+'s forecasting engine, we did what the textbooks say to do. We trained models on years of consumption data from Ethiopian facilities. We cross-validated. We optimized metrics. We got reasonable accuracy.

And then we deployed it, and reality happened.

Real facilities don't follow the distribution that historical data predicts. They have emergencies, policy changes, stock-outs that create their own momentum, NGO campaigns that shift demand. You can't model all of that in advance. You can only model it once it's happening.

The first version of the system was wrong about exactly the things that mattered most. Not by a huge amount, but enough that planners couldn't trust it. They'd override the forecast, make decisions on intuition, and we'd learn nothing from the interactions because we weren't close enough to be useful.

Learning from how people actually work

This is when we changed our approach. Instead of training a model once and deploying it, we built a system that learns from every planning cycle.

Every time a planner sees a forecast, they either act on it or override it. Both contain information. If the forecast says "order 5,000 units" and the planner says "actually, order 3,000," that's data about local context, uncertainty, and risk tolerance that the model doesn't have.

We now capture that signal. The system learns from overrides. It learns which kinds of anomalies planners trust and which they second-guess. It learns how to calibrate confidence intervals based on how much planners are willing to act.

After a year of operations in a specific region, the system gets noticeably better because it's learned from thousands of interactions. After three years, it's learned enough about that region's patterns that it catches anomalies faster than any human could.

Interaction > algorithm

The temptation is to credit the algorithm. It's not. The algorithm is fine. It's a standard time-series model with seasonal adjustments. The value is in the interaction loop.

Specifically: - **Visibility.** The planner sees the forecast in time to verify it. If it's wrong, they know immediately, not six months later when consumption doesn't match prediction. - **Agency.** The planner can override the forecast and explain why. That explanation is captured and shapes future forecasts. - **Feedback.** The system shows the planner what they predicted and what actually happened. They see if they're getting better at estimating.

This is not novel machine learning. It's basic system design. But it's what actually creates intelligence in a health system: users getting better at their jobs through slightly smarter tools, and tools getting better through how users actually work.

Why this matters for clinical systems

We're applying the same philosophy to Link. Link has 200+ diagnostic algorithms, but the value isn't in the algorithms. It's in building loops where:

- Clinicians see what the system recommended. - They make a decision (which might match the recommendation or override it). - They record the outcome. - The system learns from the pattern of recommendations and outcomes. - Next time a similar case comes in, the system is slightly more calibrated to local disease patterns and this clinician's reasoning.

This requires that the system isn't just reporting predictions; it's asking for feedback. It requires that clinicians trust it enough to use it as a thinking tool. It requires that we capture enough data about reasoning, not just outcomes, to actually learn.

The honesty

Here's the honest version: if you measure our model's accuracy on a test set from historical data, we're maybe 10-15% better than a simple baseline. That's okay, not amazing. In papers, that would not be an impressive result.

But in practice, in a facility that's been using the system for two years, the forecast that incorporates local learning is vastly more useful than the baseline because it's learned the region's actual patterns. The system that learns from clinician reasoning is far more trusted than a black-box predictor.

That's the story. Not "we built a sophisticated model." It's "we built a system where intelligence comes from interaction."

Last reviewed: January 2026