Health
Reading scans, sorting queues, spotting outbreaks
Image models flag diabetic retinopathy from a retina photograph, tuberculosis from a chest X-ray, and suspicious tissue on a pathology slide. In places with a handful of radiologists per million people, screening that a nurse can run with a phone camera changes who gets diagnosed.
- Retinal screening in clinics
- TB triage from chest X-rays
- Sepsis early-warning scores
Where it breaks: A model trained on scans from one hospital often degrades on another's machines. It reports findings, not diagnoses, and it does not know what it has never seen.
Farming
Pests, rain and yield
A photograph of a leaf can be classified for fall armyworm or cassava mosaic disease on a phone with no signal. Satellite imagery plus weather data estimates yields district by district before harvest, which is what crop insurance and food planning need.
- Leaf-photo pest diagnosis
- Satellite yield estimates
- Index-based crop insurance
Where it breaks: Advice is only as good as the local training photographs. Models built on North American maize misread East African varieties, and a wrong pest call costs a season.
Money
Fraud, credit and collections
Card networks score every transaction in milliseconds against your own history. Mobile-money lenders score borrowers with no formal credit file using airtime top-ups, transaction rhythm and repayment behaviour, which is how millions got a first loan.
- Real-time card fraud scoring
- Mobile-money credit scoring
- Anti-money-laundering triage
Where it breaks: Alternative scoring is opaque. If the model quietly learns that a neighbourhood repays badly, it penalises everyone there, and the applicant is rarely told why.
Language
Translation, transcription and drafting
Speech recognition turns a clinic consultation into notes. Translation now covers Swahili, Amharic, Yoruba and hundreds of other languages that commercial systems ignored for decades. Drafting tools take a first pass at reports, emails and code.
- Clinical note dictation
- Swahili and Sheng translation
- Code completion
Where it breaks: Low-resource languages get thinner training data and worse output, and models state wrong facts in fluent sentences. Fluency is not accuracy.
Transport
Routing, maintenance and driving
Journey estimates come from live traffic models rather than distance. Fleet operators predict which bus will break down from vibration and engine data. Driver-assist systems hold a lane and brake for obstacles.
- Live traffic routing
- Predictive fleet maintenance
- Lane keeping and automatic braking
Where it breaks: Full self-driving has been eighteen months away for over a decade. Rain, hand signals, potholes and unmapped roads remain genuinely hard.
Learning
Practice, marking and tutoring
Adaptive practice systems track which sub-skill a student is missing and serve the next question accordingly. Tutoring assistants explain a step again, differently, without getting bored — which matters in a class of eighty.
- Adaptive maths practice
- Essay feedback drafts
- Language conversation partners
Where it breaks: A confident wrong explanation is worse than no explanation. These tools work when a teacher reviews the output, not when they replace one.