In 08-01 we built a method and applied it to the returns predictor. A method, however, is only truly understood when you see it work (or fail) in real projects, with budgets, pressures and consequences that no exercise reproduces. This lesson walks through six publicly documented case studies, chosen because each illustrates a phase of the life cycle or a technique from the course: recommender systems (from collaborative filtering to deep learning, with the Netflix Prize as the story), medical image diagnosis with convolutional networks (promises and generalisation problems), AlphaGo and AlphaFold (search, networks and reinforcement with scientific impact), the recruitment and scoring systems that discriminated (a hiring tool scrapped for gender bias and the COMPAS debate), conversational assistants built on large language models (and the case of an airline forced to honour a policy its chatbot invented) and an instructive failure, Watson for Oncology, together with police facial recognition. We tell them with caution: what is stated is what is publicly documented, and we point out where we simplify. Each case ends with the NovaMarket parallel: what Marta learns for her recommender, her assistant, her predictor or her photo classifier. It matters because the most expensive lessons of AI have already been paid for by others; the junior professional who knows them avoids repeating them.

1. How to read a case study
2. Case 1: recommender systems, from collaborative filtering to deep learning (Netflix Prize)
3. Case 2: medical image diagnosis with CNNs (diabetic retinopathy, skin cancer)
4. Case 3: AlphaGo and AlphaFold, from games to science
5. Case 4: recruitment and scoring with bias (the scrapped tool and COMPAS)
6. Case 5: conversational assistants and LLMs in customer service (the airline chatbot)
7. Case 6: an instructive failure, Watson for Oncology (and police facial recognition)
8. A small runnable example: the fairness lesson in five lines
9. Summary table: case → technique → course lesson → lesson learned
10. Common Mistakes and Tips
11. Exercises
12. Conclusion

AI successes and failures are often told as headlines ("AI beats doctors", "the racist algorithm"). To learn from them you have to read them with the template of 08-01: context (who, what for, under what pressure), problem (the question and the decision that changes), technique (which model and why; we link it to the course lesson where it was covered), result (what was measured, and what was not), lessons (which phase of the cycle failed or worked) and, for us, the NovaMarket parallel. And with two cautions: many famous figures come from the companies' own press releases or from articles that were later qualified, and in several cases there was litigation or conflicting versions. When we say "according to public reports" it is because there is no single, undisputed source.

Context. In the early 2000s, e-commerce and content platforms discovered that most of what they sold did not fit on a front page: a different front page was needed for each customer. Amazon published in 2003 its item-to-item collaborative filtering method ("customers who bought this also bought that"), the same principle we programmed in 01-03 for NovaMarket. In 2006, Netflix, then a DVD-by-post rental service, launched the Netflix Prize: one million dollars for the first team to improve by 10 % the error (RMSE, 04-05) of its Cinematch system in predicting one-to-five-star ratings, on a dataset of about a hundred million anonymised ratings.

Problem. Predict what rating a user will give a film they have not seen (regression, 04-02) in order to rank the catalogue. The decision that changes: which titles are shown first.

Technique. The competition, which ran until 2009, popularised matrix factorisation: representing each user and each film with a vector of latent factors (the idea behind the embeddings of 05-04) whose dot product predicts the rating; and it demonstrated the power of model ensembles (04-04): the winning team combined hundreds of models. With the arrival of streaming, platforms moved from explicit ratings to implicit signals (what is played, for how long, what is abandoned) and, in the 2010s, to deep neural networks combining user, content and context embeddings (05-04), as YouTube described in a widely cited 2016 paper.

Result. The prize was awarded in 2009. But Netflix explained publicly years later that the full winning solution was never deployed as such: the engineering cost of maintaining hundreds of models in production did not justify the improvement, and the business had changed (from DVDs and stars to streaming and implicit signals). Moreover, researchers showed that the "anonymised" dataset made it possible to re-identify some users by cross-referencing it with public ratings on another site, which led to a lawsuit and to the cancellation of a second edition of the competition.

Lessons.

• The competition metric was not the business metric: improving RMSE by 10 % is not the same as people watching more or cancelling less. It is mistake 3 of 08-01 (wrong metric) on a million-dollar scale.
• Complexity has a deployment and maintenance cost (phases 6-7): a model 1 % better that demands ten times more engineering may be the worse decision.
• Anonymisation is not trivial (02-03, 02-04): removing names is not enough when the data is rich.
• Simple model first (collaborative filtering, factorisation) is still the baseline to beat before a deep network.

The NovaMarket parallel. For the recommender (use case 1), Marta starts with the item-to-item collaborative filtering of 01-03 on purchase and browsing data as the baseline; she defines the business metric in the definition document (revenue and clicks of the "You may also like" block, measured with A/B, exercise 3 of 08-01), not the error of a rating; and she treats purchase histories as personal data, neither publishing them nor sharing them "anonymised" lightly. If a more complex model improves clicks by 1 % but requires a new service, the decision is a business one, not the metric's.

Context. Diabetic retinopathy is a major cause of avoidable blindness; screening for it requires a specialist to examine photographs of the back of the eye, and there are far more patients than specialists. In 2016 a Google team published in a medical journal a study in which a convolutional network (05-04), trained on more than a hundred thousand photographs labelled by ophthalmologists, reached sensitivity and specificity comparable to specialists on two validation sets. In 2017 another group (Stanford) published in Nature a skin-lesion classifier trained with a CNN which, in the study's tests, performed at dermatologist level in distinguishing malignant from benign lesions.

Problem. Image classification (04-02): from a photo, are there signs of disease that justify referral to a specialist? The decision that changes: who is referred and with what priority.

Technique. Pretrained and fine-tuned convolutional networks (transfer learning, 05-04), exactly what we did with the CNN on NovaMarket's incident photos (98.7 % in our fictional case), with many more images and labels from several specialists per image to reduce labelling noise.

Result. The laboratory results were solid and opened up an entire field. But the move to the clinic taught the limitations the whole sector knows today: when the retinopathy system was tested in real clinics in Thailand (study published in 2020), a considerable share of the images taken in real conditions (lighting, cameras, nurses with little specific training) did not meet the quality the model required and was rejected, generating extra work and frustration; and the workflow (internet connection, waiting times) mattered as much as the AUC. In dermatology it was documented that the datasets had little representation of dark skin, and that some models learned shortcuts: surgical marker pen marks or measuring rulers present mostly in the photos of malignant lesions, which the model used as a clue. In chest X-rays it was shown that a model could recognise the hospital of origin of the image (from the equipment or embedded labels) and exploit it, so its performance dropped in new hospitals.

Lessons.

• Generalisation across centres (04-06): a model validated on data from the same hospitals that trained it is not validated for others. Validation must be done in different centres and in real conditions.
• Leakage and shortcuts (04-03): rulers and marker pens are the medical version of return_reason.
• Sampling bias (02-03, 02-04): if a group is barely in the data, performance on it is unknown, not "the same".
• Deployment is a workflow problem (08-01, phase 6): image quality, response time and the training of whoever uses the system decide as much as the model. And always with human review proportional to the risk: it is high risk under the AI Act (02-04).

The NovaMarket parallel. The incident-photo classifier (use case 9, 05-04) was trained on photos taken by the Zaragoza team in good light; before generalising it, Marta tests with customer photos taken on a phone at home (worse light, framing, resolution) and with the Getafe ones, measures per warehouse and per device type, and adds an image quality check that asks for another photo instead of misclassifying. And she looks for shortcuts: does the network learn "there is NovaMarket packaging in the photo" instead of the defect?

Context. Go was AI's great outstanding game: its tree of moves is so large that search with a hand-written evaluation function, which was enough for Deep Blue in chess (03-03), did not work. In 2016 AlphaGo, from DeepMind, beat Lee Sedol, one of the best players in the world, 4-1 (01-01). Its successor AlphaGo Zero learned without human games, only by playing against itself, and surpassed the original. In 2020, the same company presented AlphaFold 2 at the CASP protein-structure prediction competition, with an accuracy that for many proteins rivalled experimental methods; it later published a database with predictions for hundreds of millions of proteins, and in 2024 its main authors shared the Nobel Prize in Chemistry with David Baker.

Problem. In Go: deciding the best move (adversarial search, 03-03) in an unfathomable space. In proteins: predicting the three-dimensional shape of a protein from its amino-acid sequence, a problem open for decades and key to understanding diseases and designing drugs.

Technique. AlphaGo combined Monte Carlo tree search (a search that samples games instead of traversing the whole tree, in the family of 03-02/03-03) with two deep networks, one that proposes moves and one that evaluates positions (05-04), trained first on human games and then by reinforcement learning playing against itself (04-02). AlphaFold 2 is a system of networks with attention (05-05) over evolutionarily related sequences and over the geometric relationships between amino acids, trained on the known experimental structures and, in part, on its own high-confidence predictions. We simplify a great deal: both systems are large-scale research engineering.

Result. Beyond the headline, what matters is the scientific impact: structure prediction stopped being the bottleneck of many biology projects, and the techniques (search + networks + reinforcement, attention over structured data) have spread to materials chemistry, weather forecasting and the control of physical systems.

Lessons.

• Combining techniques wins: symbolic search (module 3) plus networks (module 5) plus reinforcement, not a single tool. It is the neurosymbolic argument of 06-04 seen from another angle.
• The problem was well defined and measurable: winning games; an accuracy metric on known structures in a blind competition. AI's great successes have a clear performance measure (02-01).
• The data existed: decades of experimental structures deposited in a public database. Without them there is no AlphaFold.
• Beware of extrapolation: beating a Go champion does not mean "AI is smarter than people"; it is narrow AI (02-02) in a fully observable environment with fixed rules.

The NovaMarket parallel. It is not going to train an AlphaGo, but the lesson of "combining search and models" it already applies: the van routes (03-04) improve if the estimate of time per leg comes from a learned model instead of a constant; and reinforcement learning is a candidate, in the medium term, for stock replenishment policies, always with a prior simulation (never learning by making mistakes with real customers). And the data lesson: if NovaMarket wants good models in three years, it has to start today to store properly what happens (dates, decisions, outcomes), as the biologists did with the structures.

Context (a). As the press reported in 2018, a large technology group had developed internally a tool to score CVs and thus shortlist candidates, trained on the CVs received over some ten years. The team discovered that the system penalised CVs containing the word "women's" (for example, "captain of the women's chess club") and those of graduates of certain women's colleges, because the hiring history for technical positions was dominated by men. Attempts were made to fix it by neutralising those terms, but there was no guarantee the model would not find other clues, and the project was abandoned.

Context (b). COMPAS is a commercial tool used in several US states to estimate the risk of reoffending of accused or convicted people, with influence on judicial decisions. In 2016 a journalistic investigation analysed thousands of cases and concluded that, among people who did not reoffend, Black people had received high-risk scores far more often than white people (more false positives), and the reverse for false negatives. The company replied that the tool was equally calibrated for both groups (a score of 7 meant the same probability of reoffending whatever the race). Researchers later showed that, when base rates differ between groups, it is mathematically impossible to satisfy calibration and equal error rates at the same time: you have to choose which notion of fairness to prioritise, and that choice is not technical.

Problem. Supervised classification (04-02) about people: is this a good candidate? will they reoffend? The decision that changes: who goes through to interview; which precautionary measure is applied.

Technique. Supervised models trained on past human decisions as the label. That is the root: the model learns to imitate the past, with its biases, and encodes them in proxies (words, postcodes, history) even if the protected feature is removed (02-04, section 3).

Result. The recruitment tool was scrapped before being used to decide; COMPAS remains in use in some places amid an open debate about fairness, transparency (the model is proprietary and opaque) and accountability.

Lessons.

• The label can be biased: "hired" or "reoffended" (in reality, "arrested again") are not the neutral truth they appear to be (02-03).
• Removing the protected column does not remove the bias if there are proxies; you have to measure per group (02-04, section 10) and decide with which metric.
• The definition of fairness must be chosen before training, with lawyers and affected people, and documented (impact assessment).
• High risk: employment and justice are high-risk domains under the AI Act; opacity plus impact on people is the combination that demands effective human review and the right to an explanation (GDPR).

The NovaMarket parallel. The returns predictor does not decide about employment or justice, but it does decide about customers: the lesson of 02-04 (flagging 90 % of one zone and 20 % of another with the same actual rate) is this case in miniature. That is why the definition document of 08-01 sets an impact ratio ≥ 0.8 between zones, the model card says that postcode_zone does not decide without review, and the monitoring calendar measures parity every month. And if one day HR asks for "a model to rank applications", Marta knows the first question is not which algorithm, but what the history says and who is accountable.

Context. Since 2023 many companies have added assistants based on large language models (05-05) to their customer service, to answer frequent questions, look up the status of an order or handle simple procedures. The results published by the companies themselves are usually positive (more queries resolved without an agent, shorter waiting times), although they are press-release figures and must be read as such. And there are public cases of the opposite. The most cited: a customer of a Canadian airline asked the website chatbot about bereavement fares; the chatbot told him he could buy the ticket and request the partial refund after the trip, something the company's actual policy did not allow. When the airline refused the refund, the customer went to a small-claims tribunal. The company argued that the chatbot was a "separate entity" responsible for its own words; the tribunal rejected that in 2024 and ordered it to honour what the chatbot had said: the information on its website, whether given by a page or by a chatbot, is the company's responsibility. Other public cases include delivery-company chatbots that, provoked by the user, insulted the company itself or composed poems against it, and systems that "confirmed" non-existent discounts.

Problem. Answering in natural language questions about policies, orders and products, with the decision that changes often being a promise to the customer.

Technique. LLM with instructions (prompting), ideally with RAG (05-05): retrieve the fragments of the actual policy and answer only from them; and with validation rules on the output (06-04, exercise 3: if the answer contains a sum of money, a deadline or a promise, check it against the source or hand over to an agent).

Result. Well-bounded assistants work and save effort; those deployed without limiting the scope or verifying the answers hallucinate policies and commit the company. The legal lesson is clear-cut: the company answers for what its chatbot says.

Lessons.

• Hallucination is not a rare fault, it is a property of generative models (05-05, section 11): the design must assume it will happen.
• RAG + cited sources + rule-based validation (neurosymbolic, 06-04) reduce the risk; and for anything that commits money, hand over to a person.
• Logging and continuous evaluation of conversations (08-01, phase 7): weekly sampling of answers reviewed by the customer-service team, and adversarial testing before deployment.
• Accountability (02-04): the chatbot's answer is a communication from the company; legal must review the scope.

The NovaMarket parallel. The assistant (use case 7) we designed with RAG in 05-05 answers only from fragments of the returns policy and the product sheets, cites the fragment, and the rules of 06-04 block any answer with unverified amounts, deadlines or promises and hand it over to a person; it never confirms a refund: it creates the request and leaves it in Diego's queue. In the assistant's model card, the "out-of-scope use" includes "committing to commercial terms". And Marta adds to the monitoring calendar the weekly sampling of 50 conversations reviewed by customer service.

Context. After winning the Jeopardy! quiz show in 2011 (01-01), IBM presented Watson as a platform for many sectors, and in particular Watson for Oncology, a system for recommending cancer treatments, developed with a prestigious US cancer centre and marketed to hospitals in several countries. According to public reports in the specialised press and a university audit (2017-2018), the project suffered several problems: part of the training was done with hypothetical cases prepared by the centre's specialists, not only with real records; the recommendations reflected that centre's practices and did not always fit the guidelines or resources of hospitals in other countries; internal documents leaked to the press described "unsafe and incorrect" recommendations in tests; and a large Texas hospital cancelled its multi-year project after spending that the audit put at tens of millions of dollars, without it ever being used with patients. IBM's health division was sold in 2022. We simplify a long story with conflicting versions; what is not disputed is the distance between the expectation created and the result.

Problem. Recommending personalised treatments from the clinical history and the medical literature; the decision that changes, an oncologist's about a patient. A problem with scarce and debatable labels, heterogeneous data and maximum risk.

Technique. Language-processing and information-retrieval systems of the time (before the transformers of 05-05), combined with knowledge curated by specialists: at bottom, an expert system (06-02) fed with data and literature, whose "knowledge base" turned out to be biased towards one centre.

Result. Progressive withdrawal of the product and a lesson for the whole sector about over-expectation (the winters of 01-01 started like this) and about the difference between winning a television quiz and helping in a consulting room.

Police facial recognition (brief). A parallel case: in 2018 an academic study (Gender Shades) showed that several commercial face-classification systems had much higher error rates on dark-skinned women than on light-skinned men, because of unrepresentative training sets; later, wrongful arrests of Black people identified by police facial-recognition systems were documented in the United States. Several cities restricted or banned its police use, some companies stopped selling it to the police and the European AI Act severely restricts remote biometric identification in public spaces (02-04).

Lessons.

• Data that does not represent the reality of use (hypothetical cases from one centre; faces from a single group): the model learns a different problem (phase 2 of 08-01).
• External validation before selling: no deployment in a new country without evaluating on that country's data and guidelines (04-06, generalisation).
• Over-expectation kills projects (01-01): promising to "cure cancer" with a recommendation system creates a debt impossible to repay.
• In high risk, human review must be real, not an "accept" button pressed out of trust in the machine (automation bias, 02-04).
• Measuring per group is mandatory when the error has serious consequences; and for certain uses, the right answer is not to deploy.

The NovaMarket parallel. The incident diagnosis (use case 9: the Bayesian network of 06-03 plus the photo CNN) is NovaMarket's small-scale "Watson": its knowledge comes from the Zaragoza technicians. Before using it at Getafe, Marta validates with Getafe incidents, does not promise to "solve all incidents" but to "propose a diagnosis that the technician confirms or corrects" and records every correction to improve the probability tables. And on facial recognition, the lesson for the future: if someone proposes "identifying customers through the physical shop's camera" or "detecting fraud from the ID-card photo", the answer starts with the impact assessment and with legal, not with the model.

Case 4 can be reproduced in miniature with fictional data, applying the rate parity of 02-04. We simulate a history of applications in which the past decision depended on experience and on a bias against one group, and in which a word in the CV acts as a proxy for the group. Then we do what the team in the real case did: remove the group column and train on the rest.

Output (run in the course environment):

Explanation: the model has never seen the gender column and yet it recommends 35 % of the men and 19 % of the women (ratio 0.56, well below 0.8). It manages this through the proxy: the coefficient of proxy_word is −1.1, that is, the presence of the word strongly reduces the score, because in the history that word went with applications that were rejected for another reason. It is exactly what was reported in the real case, and it is the reason why the per-group measurement of 02-04 is mandatory even though the model "does not use" the protected feature. Genuinely fixing it requires acting on the label (which history is acceptable as truth?), on the features (which words go in?) and on the decision (human review, interview quotas, or not automating).

• Reading the cases as headlines. "AI beats doctors" and "the racist algorithm" hide specific phases that failed or worked; always look for the metric, the data and the deployment.
• Taking press-release figures as results. Distinguish what was published with peer review or an audit from what the interested company announced.
• Believing that bias is removed by dropping a column. The example in section 8 disproves it in five lines; always measure per group.
• Thinking that failures belong to "others" or "another era". Watson and the airline chatbot are stories of over-expectation and poorly bounded scope; they repeat every year with new technology.
• Copying the technique without copying the context. AlphaFold worked because there were decades of public data and a blind metric; without that, the same architecture is useless.
• Tip: when you are offered a project, look for the closest public case and read how it ended; then write the definition document of 08-01 with that story in front of you.

Choose a public AI case not covered in this lesson (for example, a bank fraud detection system, an autonomous vehicle, a welfare-benefits system that generated complaints, or a demand forecasting model in a retail chain) and fill in the template of section 1: context, problem, technique (with the course lesson), result, lessons and NovaMarket parallel. State explicitly which parts of your profile are public knowledge and which are your own assumptions.

A customer asks NovaMarket's assistant: "Can I return the NovaBrew coffee maker after 45 days if I don't like it?". The actual policy is 30 days. Describe how the assistant should be designed (RAG, validation rules, hand-over) so that it cannot answer "yes", what is logged from that conversation and which item of the 08-01 monitoring calendar would detect a failure if it happened. Relate it to case 5.

With the code of section 8, try two "repairs" and measure the impact ratio in each: (a) train without proxy_word (with experience only); (b) keep both features but set the recommendation threshold per group so that the rates are equalised. Discuss what problem each solution has and why neither replaces the question about the label.

Solution 1. There is no single answer; the criterion is that the profile separates facts from assumptions and links each element to the course. Example with a welfare-benefits system that generated complaints: context: a public administration automated the detection of fraud in benefits; problem: classifying applications as suspicious (04-02); technique: supervised model with history and demographic data, in which nationality or postcode acted as proxies (02-04); result: thousands of families wrongly flagged, political resignations and sanctions (public knowledge in the best-known case, although the technical details of the model are only partially known); lessons: biased label, high risk without real human review, no right to an explanation; parallel: the returns predictor never denies or penalises automatically and measures parity between zones.

Solution 2. Design: the assistant answers with RAG over the returns policy (05-05); the retrieved fragment says "30 days", so the generated answer should be "no, the period is 30 days", citing the fragment. Since generation can fail, a validation rule (06-04) detects that the answer contains a period ("45 days") and checks that it appears in the retrieved fragment; if it does not, it replaces the answer with the literal policy or hands over to a person. In addition, any answer that implies an exception or a promise ("yes you can", "we will refund you") is blocked and creates a ticket for customer service. What is logged: the question, the retrieved fragments, the generated answer, the validation applied and the final answer, with a conversation identifier. It would be detected by the weekly sampling of conversations reviewed by customer service (the row added to the calendar in section 6), and even earlier by the adversarial tests before deployment. It is case 5 in miniature: if the assistant had said "yes", NovaMarket would have to honour it.

Solution 3. (a) Without proxy_word, the model uses experience only. In this fictional dataset experience is distributed equally between groups, so the recommendation rates come closer and the impact ratio rises clearly (in the course environment: 30.7 % of men and 36.4 % of women recommended, ratio 0.84, above 0.8; the small imbalance in favour of the other group is chance in the generation). Problem: in a real case there are always other proxies (name of the college, gaps in the CV, hobbies) and they cannot all be removed; moreover, if in the history women with the same experience were hired less, the label is still biased even if the features are neutral. (b) Per-group thresholds: for each group, the probability threshold that produces the same recommendation rate is chosen (for example, that of the most favoured group); the impact ratio becomes 1 by construction. Problem: the protected feature is being used explicitly to decide (something that in many domains requires legal justification and in others is prohibited), and one metric (rate parity) is equalised at the expense of others (calibration, error rates), the COMPAS dilemma. Neither of the two touches the underlying question: if hired_history reflects biased decisions, the model is learning to reproduce them; the real repair starts by deciding which label is acceptable as truth (for example, later performance of those hired, or blind assessments), which definition of fairness is adopted and what role human review plays.

Six cases, six lessons we already knew in theory and that here have a name and an invoice. From the Netflix Prize and the recommenders: the competition metric is not the business one, complexity is paid for in production and anonymising is hard. From image diagnosis: models shine in the laboratory and stumble when the hospital, the camera or the skin tone changes, and the workflow decides as much as the AUC. From AlphaGo and AlphaFold: combining search, networks and reinforcement on measurable problems with data accumulated over decades produces real scientific impact, and it is still narrow AI. From biased recruitment and COMPAS: the label inherits the past, removing the column is not enough, you have to measure per group and choose the definition of fairness (we reproduced it in five lines: impact ratio 0.56 without seeing gender). From the airline chatbot: hallucination is a property, the company answers for what its assistant says, and RAG with validation and human hand-over is the minimum design. From Watson for Oncology and facial recognition: data that does not represent the reality of use, over-expectation and the absence of external validation sink projects, and sometimes the right decision is not to deploy. And for NovaMarket, each case has left a concrete decision in the recommender, the photo classifier, the routes, the predictor, the assistant and the incident diagnosis.

With the method (08-01) and other people's experience (08-02), what remains is to look ahead: what is changing in AI right now, what is solid and what is speculation, and what a junior professional and a company like NovaMarket should prepare for. That is the last lesson of the module, 08-03, Future Trends in AI.