In the two previous lessons we saw where artificial intelligence comes from and what its areas are. Now it is time to answer the question Marta and Diego are asking at NovaMarket: "what is this actually being used for?". We will go through the main sectors (health, transport and logistics, finance, e-commerce and marketing, industry, customer service, education, home and leisure) and, for each application, identify which technique lies underneath (connecting with the map of areas from lesson 01-02) and what benefits and risks it brings. At the end we will bring everything down to NovaMarket: we will fix its list of candidate use cases, which will be the map for the rest of the course, and we will build a first trivial recommender in Python so you can see that the distance between the idea and a prototype can be surprisingly short.

1. How to read an AI application: problem, technique, benefit, risk
2. Health
3. Transport and logistics
4. Finance
5. E-commerce and marketing
6. Industry
7. Customer service
8. Education
9. Home and leisure
10. Summary table: sector, problem, technique, example
11. NovaMarket's candidate use cases
12. Example in Python: a trivial recommender

Before going through the sectors, it helps to fix a reading scheme. Every time you see an AI application, ask yourself four questions:

1. What problem does it solve? It is almost always one of these four things: predicting something (demand, failure, risk), classifying something (image, text, transaction), optimising something (route, timetable, price) or interacting with people (answering, conversing, generating).
2. What technique lies underneath? Use the map from 01-02: machine learning, NLP, vision, robotics, expert systems, search and planning, agents, generative AI. Many applications combine several.
3. What benefit does it bring? Cost savings, more revenue, more speed, more safety, a better experience.
4. What risk does it introduce? Errors with consequences, biases, opacity, dependency, privacy, impact on employment. These risks are covered in depth in lesson 02-04; here we only point them out.

This scheme is the one Marta will use to evaluate proposals at NovaMarket, and it is the one we will apply to each sector.

• Image-based diagnosis: detecting tumours in mammograms, diabetic retinopathy in retinal photographs, fractures in X-rays. Technique: computer vision (convolutional networks). Benefit: quick second opinion, mass screening. Risk: false negatives, models trained on populations that do not resemble the local one.
• Diagnosis support and triage: systems that, from symptoms and test results, suggest diagnoses or prioritise patients in A&E. Technique: expert systems historically (remember MYCIN), today combined with ML and probabilistic reasoning. Risk: legal liability, over-reliance by the doctor.
• Risk prediction: probability of hospital readmission, of sepsis, of complications. Technique: ML on clinical records. Risk: historical biases in the data (for example, populations with less access to healthcare appear "healthier").
• Drug discovery and protein folding: AlphaFold (2020) predicted the 3D structure of almost all known proteins. Technique: deep learning.
• Assistants for professionals: transcribing consultations, drafting reports. Technique: NLP and generative AI.

• Navigation and route computation: the GPS that picks the fastest way in real time. Technique: graph search (A* and variants) over the street map, with traffic prediction by ML. Benefit: time and fuel.
• Delivery route optimisation: how to deliver 200 parcels with 8 vans while minimising kilometres and meeting time slots. Technique: optimisation algorithms (local search, heuristics, mathematical programming). Benefit: fewer vehicles, fewer emissions. This is exactly one of Diego's problems.
• Transport demand forecasting: how many trips there will be in each area of the city in the next hour (ride-hailing, bike sharing). Technique: time-series ML.
• Autonomous driving and driver assistance: emergency braking, lane keeping, driverless cars in restricted zones. Technique: vision, sensors (lidar, radar), planning, reinforcement learning. Risk: accidents in situations not seen in training; liability.
• Predictive maintenance of fleets and aircraft: anticipating breakdowns from sensors. Technique: ML on sensor time series.

• Card and payment fraud detection: flagging a suspicious transaction in milliseconds (unusual amount, different country, strange purchase pattern). Technique: classification and anomaly-detection ML. Benefit: losses avoided. Risk: blocking legitimate payments (false positives) and annoying the customer. NovaMarket suffers exactly this problem in its store.
• Credit scoring: deciding whether to grant a loan and at what rate. Technique: ML (often interpretable models because of regulatory requirements) and rules engines. Risk: indirect discrimination (the model learns historical biases), legal obligation to explain the decision.
• Algorithmic trading: systems that buy and sell in microseconds. Technique: ML, optimisation. Risk: market instability (flash crashes).
• Regulatory compliance and money laundering: detecting patterns of suspicious movements. Technique: graph analysis, ML, rules.
• Assistants and document analysis: reading contracts, extracting clauses, answering queries. Technique: NLP and generative AI.

This is the sector closest to NovaMarket, so we will linger a little longer:

• Recommender systems: "customers who bought this also bought...", "recommended for you", the ordering of the home page. On large commerce and video platforms, a very significant share of sales or views comes from recommendations. Technique: ML (collaborative filtering, content-based models, neural networks). Benefit: more sales per customer, catalogue discovery. Risk: recommendation bubbles (always the same thing), recommending inappropriate products.
• Demand forecasting and stock management: how many units to buy and in which warehouse to keep them. Technique: time-series ML. Benefit: fewer stock-outs and less capital tied up. Risk: failures on unprecedented events (a pandemic, a viral campaign).
• Dynamic pricing: adjusting prices according to demand, competition and stock. Technique: ML and optimisation. Risk: perceived unfairness by the customer, price wars.
• Catalogue search: understanding that "charger for my daughter's phone" is looking for USB-C chargers. Technique: NLP.
• Opinion analysis (reviews, social media): knowing which products generate complaints about shipping and which about quality. Technique: NLP (sentiment and topic classification).
• Customer segmentation and churn prediction: grouping customers by behaviour and anticipating who will stop buying. Technique: unsupervised and supervised ML.
• Content generation: product descriptions, campaign images, personalised emails. Technique: generative AI. Risk: errors in descriptions, generic content.
• Programmatic advertising: deciding in milliseconds which ad to show and how much to bid. Technique: click-prediction ML.

• Visual quality control: cameras that detect defective parts on the production line. Technique: computer vision. Benefit: inspection of 100% of parts instead of sampling.
• Predictive maintenance: predicting when a machine will fail from vibration, temperature and power consumption. Technique: time-series ML and anomaly detection. Benefit: fewer unplanned stoppages.
• Industrial and collaborative robotics: arms that assemble, mobile robots that move material in warehouses (picking). Technique: robotics, vision, planning. Large logistics operators use thousands of robots in their centres; NovaMarket has it on its "medium-term" list.
• Production and energy optimisation: scheduling shifts and machines, cutting the energy consumption of a plant or a data centre. Technique: optimisation, reinforcement learning.
• Digital twins: simulations of a factory fed with real data to test changes before applying them.

• Conversational assistants (chatbots and voicebots): they answer frequently asked questions, report the status of an order, handle simple procedures. Technique: NLP and generative AI, integrated as agents that query internal systems. Benefit: 24/7 service, lighter load on the human team. Risk: made-up answers ("hallucinations"), frustration if the bot does not know how to hand over to a person, exposure of personal data.
• Ticket classification and routing: reading a complaint email and sending it to the right department with the appropriate priority. Technique: classification NLP.
• Assistance for the human agent: suggesting replies, summarising the conversation, searching the knowledge base. Technique: generative AI.
• Voice-of-the-customer analysis: detecting the emotional tone of calls and chats in order to step in in time. Technique: NLP and audio processing.
• Rule systems for policies: automatically applying the conditions for returns, warranties or compensation. Technique: expert systems / rules engines. Benefit: consistency and traceability.

• Intelligent tutors and adaptive learning: platforms that adjust the difficulty of exercises to each student's pace. Technique: ML, student models, sometimes expert systems.
• Automatic marking and feedback: assessing open answers, code or essays. Technique: NLP and generative AI. Risk: unfair or gameable assessments.
• Drop-out detection: identifying at-risk students from their activity on the platform. Technique: ML.
• Generating materials: summaries, quizzes, alternative explanations. Technique: generative AI. Risk: content errors, student dependency.
• Accessibility: automatic subtitling, read-aloud, translation. Technique: speech recognition, NLP.

• Voice assistants (smart speakers, phones): they recognise the voice, understand the request and act. Technique: speech recognition (deep learning), NLP, agents.
• Home automation: thermostats that learn schedules, robot vacuums that map the house. Technique: simple ML, robotics, planning.
• Content recommendation: series, music, short videos, news. Technique: recommendation ML. Risk: addiction, filter bubbles.
• Video games: AI-controlled characters (search, state machines, minimax in board games), level generation, difficulty adjustment. Technique: adversarial search, planning, ML.
• Photography and creativity: automatic photo enhancement, background removal, image and music generation. Technique: vision and generative AI.
• Translation and communication: instant translation of text and speech. Technique: transformers.

Two observations about the table:

• The same technique appears in many sectors. Anomaly detection serves for bank fraud, industrial breakdowns and suspicious returns. Learning the technique equips you for many domains.
• The risk is almost never "the AI rebelling". The real risks are mundane: false positives, biases in the data, unforeseen situations, opaque decisions. We will come back to them in 02-04.

Marta and Diego have made the same journey you have and have come out with a list of NovaMarket problems that AI could tackle. This list matters because it is the map for the rest of the course: each use case will be worked on in one or more modules, always with the company's fictional data (customers.csv, orders.csv, products.csv, reviews.csv, incidents.csv).

In addition, transversally:

• Module 2 will use these cases to talk about agents, types of AI, data and ethics (what happens if the fraud detector systematically blocks customers from one neighbourhood?).
• Module 7 will present the tools they are built with (Python, NumPy, pandas, ML libraries).
• Module 8 will develop the complete life cycle of a project, taking one of these cases as an example.
• Module 9 contains the labs, and its capstone project combines several cases.

Not everything gets done at once. Marta and Diego apply three criteria, which will serve you for any company:

1. Value: how much money, time or satisfaction is at stake? (Returns cost NovaMarket around 4% of turnover; delivery, about the same.)
2. Feasibility: do we have the data and are they of good quality? Is the technique mature? (Recommendation and demand forecasting: abundant data and mature technique. Vision for damage in photos: few examples so far.)
3. Risk: what happens if the system gets it wrong? (Recommending the wrong product is harmless; blocking an order as false fraud angers the customer; an assistant that promises a return the policy does not allow creates a legal problem.)

With these criteria, NovaMarket decides to start with recommendation (high value, abundant data, low risk) and demand forecasting, and to leave the conversational assistant for when the team has more experience. It is a typical and sensible decision: first whatever teaches without doing harm.

To close the module, we will build the simplest possible version of use case 1: a recommender of the "customers who bought X also bought Y" kind. No libraries, just Python dictionaries and lists. The goal is not for it to be good, but for you to see the mechanics and check that the idea of "learning from data" we introduced in 01-02 applies here too.

We simplify orders.csv to a list of orders, each with the products it contained:

The idea is to count, for each pair of products, how many times they appear together in the same order. We will store the counts in a dictionary of dictionaries: co_purchases["laptop"]["mouse"] will be the number of orders in which both appear.

Line by line:

1. co_purchases starts empty. We walk through each order and, inside it, each product.
2. If it is the first time we see product, we create an empty inner dictionary for it.
3. For each other product in the same order (skipping the product itself), we increment the counter co_purchases[product][other]. If the counter did not exist, we initialise it to 0 first.
4. When done, co_purchases["laptop"] tells us the laptop has appeared 3 times with the mouse, 2 with the sleeve, 1 with the headphones and 1 with the mouse pad.

This step is the "training" of our recommender: we have turned the data into a model (the dictionary of counts). There is no hand-written rule about what goes with what; everything comes out of the orders.

Recommending is straightforward: given a product, we sort its co-purchases from highest to lowest and return the first few.

Explanation:

• neighbours.items() returns (other, times) pairs; sorted(..., key=lambda pair: pair[1], reverse=True) sorts them by the second element (the times), highest first.
• We keep the first n and return only the names.
• If the product does not appear in the model (nobody has ever bought it), we return an empty list. This is the famous cold-start problem of recommenders: you cannot learn from what has no data.

A NovaMarket customer rarely buys a single product. We can add up the counts of all the products in the cart and exclude what is already in it:

For the cart ["laptop", "mouse"], the sleeve adds up to 2 (laptop) + 1 (mouse) = 3 and the mouse pad adds up to 1 (laptop) + 2 (mouse) = 3; the headphones add up to 1. The tie is resolved by the dictionary's insertion order, which in this case puts the sleeve first. In a real system you would need to define tie-breaking criteria (for example, product margin or novelty).

This recommender works, and in fact it is the basic idea of item-based collaborative filtering that Amazon popularised in the early 2000s. But it lacks things that are covered in later modules:

What matters is that you have now seen the complete mechanics of an AI application: data, "training" (counting), model (dictionary) and prediction (recommending). The rest of the course refines each of these pieces.

• Confusing the application with the technique. "A chatbot" is not a technique; it may be a rule system (like ELIZA), an intent classifier plus templates, or a generative model with access to data. Always ask what lies underneath, because it changes the cost, the reliability and the risks.
• Choosing the flashiest use case instead of the most feasible one. Many companies start with the conversational assistant because it impresses, and fail through lack of data, integration or realistic expectations. NovaMarket starts with recommendation for a reason.
• Ignoring the data. In the use-case table, each row lists its files. If a case has no data (or the data are poor quality), it is not an AI project yet: it is a data-collection project (lesson 02-03).
• Measuring success by "whether it works in the demo". A recommender is evaluated by whether it increases sales or satisfaction; a fraud detector by how many frauds it catches and how many legitimate customers it annoys.
• Forgetting the risk. Before deploying, ask yourself what happens when the system gets it wrong and who suffers. That reflection, formalised, is AI ethics (02-04).
• Tip: keep the NovaMarket use-case table. When you reach each module, come back to it and check which row you are solving: it will help you keep the thread between theory and application.

For each of these applications, state the sector, the type of problem (predict, classify, optimise, interact), the main technique and one risk:

a) An app that, using the phone's camera, identifies a plant and says whether it is diseased. b) A system that decides in which order to prepare orders in the Getafe warehouse so that the delivery drivers do not have to wait. c) A system that reads NovaMarket's complaint emails and sends them to the right department. d) A system that estimates how many coffee makers will be sold at Christmas. e) A speaker that plays music when you ask it to.

Diego proposes tackling use case 9 (automatic incident diagnosis) first because it takes up a lot of the team's time. Marta points out that incidents.csv has only 400 records, half of them with the "cause" field empty, and that the causes change every time the courier company changes. Apply the three prioritisation criteria (value, feasibility, risk) and reason in a few lines whether it makes sense to start there and, if not, what would be needed to make it feasible.

Extend the recommender from section 12:

1. Add a function popular_products(orders, n) that returns the n best-selling products overall. Use it as the default recommendation when recommend has no data for the product (cold start).
2. Add two new orders to orders, ["tv", "wall_mount"] and ["tv", "soundbar", "wall_mount"], retrain and check what it recommends for "tv".

Solution 1

Solution 2 (indicative)

Value: high (it saves the team time and improves the response to the customer). Feasibility: low today: 400 records are few, half of them lack the label we want to predict, and the phenomenon changes over time (each change of courier invalidates part of the history). Risk: medium: a wrong diagnosis delays the resolution of an incident. Conclusion: it is not advisable to start there. To make it feasible you would first need to make filling in the "cause" field mandatory and structured (a closed list of causes instead of free text), accumulate a few months of data and, in the meantime, approach the case as an expert system based on the rules Diego's agents already apply in their heads (module 6), which does not need much data and helps collect it properly. Later, with data, it can evolve towards a learned model.

Solution 3

Explanation: popular_products counts how many orders each product appears in and returns the most frequent; with the original data these are the laptop (4 orders) and the mouse (4 orders). When the TV has no history, recommend_v2 falls back on that list. As soon as we add two orders with the TV and retrain (call count_co_purchases again), the model already knows that the TV goes with the wall mount (2 times) and with the soundbar (1 time). Notice the pattern once more: we have not written a single rule about TVs; giving it data was enough.

In this lesson we have gone through the real applications of AI by sector (health, transport and logistics, finance, e-commerce and marketing, industry, customer service, education, home and leisure), reading each with the same scheme: what problem it solves, what technique lies underneath, what benefit it brings and what risk it introduces. We have confirmed that a handful of techniques (classification, prediction, optimisation, search, NLP, generation) recur across every sector, and we have fixed the list of nine candidate use cases for NovaMarket (recommendation, demand forecasting, fraud and returns, reviews, routes, order assignment, assistant, return rules and incident diagnosis) together with the module of the course in which each will be worked on. Finally, we have built a trivial recommender in Python that already contains, in miniature, all the pieces of an AI application: data, training, model and prediction.

With this we close module 1. You now know where AI comes from, what it is and what it is used for. In module 2, Basic Principles of AI, we will start building the conceptual foundations: what an intelligent agent and its environment are, what it means to act rationally, what types of AI exist, why data are the raw material of everything we have seen today, and what ethical considerations NovaMarket must take into account before putting any of these use cases into production.