Lesson 5 — Where AI is really used
Demonstrations are easy and deployments are hard. This lesson looks at where AI genuinely earns its cost in 2026, and is equally explicit about where it disappoints, because knowing the failures is what stops you funding one.
Health
What works. Medical imaging is the clearest success. Diabetic retinopathy screening from retinal photographs performs at specialist level and is deployed in national programmes, precisely because it targets a task with abundant labelled data and a shortage of specialists. Radiology assistance follows the same pattern: the model flags suspicious regions and a radiologist decides. Protein structure prediction shifted from a problem requiring years of laboratory work to one solved in minutes, which changed how drug discovery starts.
What disappoints. Systems intended to recommend treatment plans have repeatedly failed on contact with reality. The reasons are instructive: recommending a treatment requires reasoning over a patient's full history, comorbidities, preferences and local practice, and that information is scattered across incompatible records, much of it as free text written under time pressure. The model was never the bottleneck.
The pattern. AI succeeds in medicine on narrow perceptual tasks with plentiful labels, and struggles on broad judgement over messy records.
Finance
What works. Fraud detection is arguably the most mature commercial application in existence. Every card payment is scored in milliseconds against a model trained on billions of transactions, and it works because the conditions are close to ideal: enormous volume, fast and reliable labels — a customer disputes a charge and you learn the truth — and a business case where a small percentage improvement is worth a great deal.
Credit scoring, algorithmic trading and anti-money-laundering triage are all established. Documentation processing has quietly become routine: extracting fields from invoices, contracts and identity documents.
What is constrained. Regulation, not capability, is the binding limit. A refused loan usually requires an explanation, which rules out models that cannot produce one. This is why finance leans on gradient-boosted trees with explainability tooling rather than deep networks, even where the network scores marginally better. Lesson 3's rule shows up here as a legal requirement.
Industry and logistics
Predictive maintenance is the flagship. Vibration, temperature and current sensors on a machine feed a model that learns the signature preceding a failure. Replacing a bearing during planned downtime instead of after an unplanned stoppage is worth more than any accuracy metric suggests.
Visual quality inspection is now commodity technology. A camera on the line, a model trained on defect images, and a reject rate that outperforms a tired human at the end of a shift.
Demand forecasting drives stock levels, staffing and routing. It is also where the limits of extrapolation are most visible: any model trained before 2020 forecast 2020 catastrophically, because nothing in its history resembled what happened. A model interpolates within its experience and cannot extrapolate beyond it.
Products you already use
Most people interact with a dozen models a day without noticing.
- Recommendation on streaming and commerce platforms, driving a large share of what gets watched and bought.
- Search ranking, which has been a machine learning problem far longer than it has been an AI story.
- Spam and abuse filtering, still one of the highest-value classifiers ever deployed.
- Voice assistants and dictation, where transcription accuracy crossed the usability threshold around 2017.
- Machine translation, transformed by the same transformer architecture that later produced language models.
- Computational photography: your phone's night mode, portrait blur and face detection are all learned models.
The applications that generate the most value are unglamorous and invisible. Spam filtering, fraud scoring and search ranking each create more measurable economic value than most of what gets announced at conferences.
Generative AI, specifically
Since 2022 a distinct category has become commercially significant: models that produce content rather than a label.
Where it holds up. Assisted code writing, where a developer reviews every suggestion. Drafting and reformulating text where a human validates the output. Summarising long documents. Question answering over a company's own documents, using the retrieval technique covered in the RAG premium course. Concept imagery for design iteration.
Where it does not. Anything where an invented fact is expensive and nobody checks. Lesson 2 explained why: a language model generates plausible text, and plausibility is not truth. Cases of professionals submitting documents containing confidently fabricated references are now numerous enough to be a recognised category of incident.
The reliable pattern is assistance with review, not autonomy without review.
The three questions that decide a project
Before any technical discussion, three questions settle whether a project is viable. In practice they eliminate most ideas, which saves a great deal of money.
1. Do I have the data? Not "does data exist in the world" — do you have, in-house and legally usable, enough labelled examples of the thing you want to predict. If the answer requires a labelling project first, that is the project, and it should be budgeted as such.
2. Is an imperfect answer useful? A model returns probabilities, never certainties. Ranking, flagging and prioritising tolerate that comfortably. Anything demanding a guarantee does not.
3. Is the outcome measurable? If you cannot say what success looks like in numbers before starting, you will not be able to tell whether the deployed system works. "Improve the customer experience" is not measurable. "Reduce average handling time by 15% without lowering the satisfaction score" is.
In three sentences
AI earns its keep on narrow, high-volume tasks with reliable labels — fraud scoring, visual inspection, medical screening, recommendation — and disappoints on broad judgement over messy records. Generative AI is reliable as assistance with human review and unreliable as autonomy without it. Whether a project is worth starting is decided by data availability, tolerance for probabilistic answers and a measurable target, all of which precede any choice of algorithm.
Next — Lesson 6: recap and FAQ →