When Not to Use AI
Not every problem in the field calls for the most capable model available. A great deal of AI development defaults to more of everything, more data, more compute, more capability, without first asking whether the environment the tool has to run in can support any of it. In many of the places we deploy it cannot, and building as though it can does not produce a better tool, only one that fails in use.
A large, highly capable model generally needs real compute, a stable connection and consistent power, and a rural health post or a smallholder farm often has none of those reliably. Shipping the most sophisticated version of a diagnostic model into that environment means building something that fails exactly when it is needed most, not because the model was wrong but because the conditions around it were never able to support it.
The design question is therefore not how good a model can be made, but how much model a specific environment can support and what the simplest version is that still does the job. Sometimes the answer is a much smaller on-device model rather than a large cloud-based one. Sometimes it is a rules-based fallback rather than a learned model at all, because a fixed set of clinical guidelines does not need machine learning to be applied correctly, and putting a model on top of something already well defined adds a point of failure without adding value.
Referral and reporting show this clearly. The instinct is to build a single AI-optimised channel for moving a patient record from a clinic to a hospital, but if that channel depends on a stable internet connection it is the wrong design for a place where the connection is not guaranteed. The better answer there has nothing to do with AI and everything to do with redundancy, several routes for the same record, so that one bad connection does not mean the record never arrives.
None of this is an argument against AI. It is an argument against reaching for the most capable version of something by default and expecting the environment to adjust around it, because the environment will not adjust and the tool has to.