AI Across Industries

You have spent this track learning how AI systems are built: how they learn from data, how they represent the world as numbers, how training shapes a model's behavior. Now we zoom out. AI is not a lab experiment — it is embedded in the systems that run hospitals, banks, farms, factories, and roads. This final module of AI Foundations maps that landscape honestly: where AI works well, where it struggles, and what it means for you as someone who will live and work in an AI-shaped world.

Healthcare: Diagnosis, Discovery, and Risk

Medicine has become one of the most consequential domains for AI deployment. The applications span three broad areas. Diagnostic imaging is the most mature. Convolutional neural networks trained on tens of thousands of labeled scans can detect diabetic retinopathy from eye photographs, flag potential lung nodules in CT scans, and identify skin lesions that warrant biopsy. A 2019 study published in Nature Medicine showed a deep-learning system matching or exceeding the diagnostic accuracy of six radiologists on detecting lung cancer from chest CT. The model was not replacing radiologists — it was a second set of eyes, catching cases that fatigue or time pressure might cause a human to miss. Drug discovery is slower to see deployed results but structurally just as important. Training a model to predict whether a candidate molecule will bind to a protein target or cause off-target toxicity can narrow a search space of billions of possibilities before any physical synthesis occurs. DeepMind's AlphaFold2, released in 2021, predicted the three-dimensional structure of virtually every known protein with accuracy rivaling expensive experimental techniques — a result that researchers described as solving a 50-year-old grand challenge in biology. Clinical risk scoring applies machine learning to electronic health records to flag patients at elevated risk of sepsis, hospital readmission, or deterioration. Epic, the dominant U.S. electronic health record vendor, has deployed such models widely — and they have also surfaced concerns about how training data can bake in historical disparities. A much-cited 2019 study in Science found that a commercial risk-scoring algorithm was less likely to flag Black patients as high-risk than white patients with equivalent actual health burden, because it used past healthcare spending as a proxy for health need — and Black patients historically spent less due to systemic barriers to access.

Finance, Transportation, and Agriculture

Finance was an early adopter of machine learning. Fraud detection systems analyze millions of transactions per day, scoring each for anomalous patterns — an unusual location, an atypical purchase category, a velocity of transactions that doesn't match your history. Credit underwriting uses gradient-boosted models to assess loan default risk, though this has drawn regulatory scrutiny because the features a model finds predictive (zip code, for instance) can serve as proxies for protected characteristics. High-frequency trading firms have used ML models to predict short-term price movements for over a decade. These systems operate in microseconds and are not 'understanding' markets in any human sense — they are pattern-matching on historical order book data with a time horizon of milliseconds. Transportation: the self-driving vehicle is the most publicized application and the one whose technical challenges have most consistently exceeded early predictions. A fully autonomous vehicle must solve perception (identifying objects from cameras, lidar, and radar), prediction (modeling how pedestrians and other drivers will move), planning (choosing a path), and control (executing it) — all in real time, in an open world with infinite edge cases. Waymo operates commercial robotaxi services in Phoenix and San Francisco, but only in geofenced areas with extensive prior mapping. Level 5 autonomy — safe, unrestricted operation anywhere — remains an unsolved problem. Agriculture has seen quieter but genuine deployment. Computer vision systems on tractors identify weeds at centimeter resolution, allowing herbicide to be applied precisely to each plant rather than broadcast across a field — a product company John Deere acquired when it bought Blue River Technology in 2017. Satellite and drone imagery combined with ML models estimate crop yield, detect disease outbreaks early, and help allocate irrigation. These tools are most accessible to large commercial operations; access in smallholder farming contexts, which dominate global food production, remains limited.