IS5542 Generative Artificial Intelligence for Business

How IS5542 is assessed

The University's catalogue entry sets continuous assessment at 50% and a two-hour examination at 50%, with minimum passing requirements of 30% for coursework and 20% for the examination; students must pass both to pass the course. The coursework split comes from the first lecture: 20% for class participation and assignments and 30% for the group project, made up of a 20% team presentation and 10% peer evaluation. The catalogue calls its assessment line indicative, and the course materials describe the paper twice: first as multiple choice 20%, short answers 30% and essays 50%, later as multiple choice 20% and essays 80%, four questions from five at 20 points each. Check the course Canvas site for your own offering's structure and dates.

What IS5542 covers

Fourteen chapters follow the four course modules, from how generative models work and how to compare them, through ideation, building and measurement, to industry cases, the law of AI and responsible AI.

The published aim is that you learn to use generative techniques to build conversational agents, automate customer interactions, improve decision-making and create value across business domains, combining theory, hands-on exercises and case studies.

The teaching runs in four modules: an introduction to generative AI, its strategic implementation, industry use cases presented largely by guest practitioners, and responsible AI and future trends.

The final examination rewards one skill above all, which is applying a concept to a business case and naming the trade-off that decides it.

How Generative AI Works: From Rules to Transformers

The first chapter separates classification from regression with a weather example, then shows why language is harder: variable length, long-range context and the need to turn words into numbers.

Tokens, embeddings and attention answer those problems, and the context window becomes a business variable. A short history runs from pattern-matching chatbots through the deep learning turn to the Transformer and today's multimodal models.

Benchmarks and Leaderboards for Comparing LLMs

Vendors publish scores on standard suites, and the course teaches you to read them as evidence with limits.

Each suite probes one kind of ability; top models saturate old tests; leaked questions contaminate results.

Blind head-to-head voting ranks models by user preference instead, which is useful for a shortlist and not for a final decision about your own task.

Choosing a Model: Six Vectors and Model Families

Modality, latency, performance, context window, throughput and cost pull against each other, so the right model depends on the workload.

A real-time voice agent and a complex financial analysis are worked through every vector, and compound systems that route simple traffic to cheap models are presented as the way businesses manage the trade-offs.

Framing the Business Problem with Cynefin Domains

Before choosing technology, the course sorts the problem by how well cause and effect are known.

Routing simple inquiries, debugging a large codebase, launching a personalised learning platform and countering a disinformation cascade each call for a different order of sensing, analysing, probing and acting.

Prompt and Context Engineering for Business

The ideation workshop opens with prompting: personas, audiences, constraints, the task and the output format for business requests, five-part formulas for images and video, and the technique of asking a model to improve a weak prompt.

Its summary is that context is the input you control.

Jobs to Be Done and How Might We Statements

Opportunities are found as job stories, when a situation arises, what someone wants to do and what outcome they expect, then turned into How Might We problem statements whose purpose clause names the job.

Statements are sorted on a matrix of value against complexity and three are chosen to take forward.

The Ten-Step GenAI Build and Operations Plan

A ten-step process takes an application from a business domain and persona, through a tiger team of business person, prompt engineer and operation lead, to interface design, wider testing and an operation plan with human oversight, and finally to new use cases.

Measuring Value: KPIs, Evaluation and ROI

Key performance indicators are organised in three tiers: model performance, system quality and business impact.

Human evaluation and automated metrics complement each other, and the group presentation asks for results against a baseline and an estimate of return on investment.

GenAI Use Cases in Marketing and Retail

Practitioners present a maturity framework for AI-powered marketing, reported results from named retail cases, the step from chatbots to tool-using agents, and the idea of agentic commerce in which agents discover and buy on a customer's behalf.

GenAI in Hong Kong Banking and Insurance

The financial-services session moves from a vision of the bank of tomorrow to service use cases sold on specific KPIs, an agent that investigates unusual market movements, and the constraints of data residency and AI-enabled fraud.

AI Law: Copyright, Ownership and Licensing

A solicitor's guest lecture treats law as the environment every model runs in.

The chapter covers sources of law, the gap between ethics and regulation, the training-data copyright dispute and fair use, and how ownership of generated output differs between jurisdictions.

AI Liability, Data Privacy and Algorithmic Bias

Negligence and products liability, the chain from model developer to affected person, disparate impact and proxy discrimination, and privacy rules on lawful basis, erasure and automated decisions explain how an AI deployment becomes a legal claim.

Regulating AI in Mainland China and Hong Kong

The Mainland regulates through layered administrative measures, including dedicated rules for generative AI, while Hong Kong applies its common law and privacy ordinance with non-binding guidance.

Businesses in Hong Kong often meet both.

Responsible AI and Safety Mitigation

The last module presents responsible AI as good business and as engineering: published principles, product review that identifies, assesses and mitigates risk, mitigation plans across technical, policy and documentation families, and safety evaluation that moves further from the builders at each layer.

Worked example · free

Choose between two models for a customer-service rollout

Key terms

Large Language Model

A deep neural network trained on very large text collections to predict the next token, general enough to serve many tasks after pre-training and fine-tuning.

Context Window

The amount of input a model can attend to in one request. Beyond the effective limit, extra text adds noise rather than accuracy.

Hallucination

Fluent, grammatical output that is not grounded in source data, such as an invented citation or a policy the business never offered.

Groundedness

A model performance KPI measuring how far an answer uses only the information supplied in the prompt or retrieved context.

Six Vectors

The characteristics the course uses to fit a model to a use case: modality, latency, performance, context window, throughput and cost.

Cynefin Framework

A way of sorting problems into clear, complicated, complex and chaotic domains, each with its own order of sensing, analysing, probing and acting.

Job Story

An opportunity written from the customer's side: the situation they are in, what they want to do, and the result they hope for.

Tiger Team

The small cross-functional team that builds a GenAI application, combining business knowledge, prompt design and production operations.

Operation Plan

The plan for running an application in production, covering infrastructure, monitoring, quality evaluation, audits, updates, security, human oversight and incident response.

Agentic Commerce

Shopping in which an AI agent anticipates a need, compares options and, with permission, completes the purchase.

Data Residency

Commitments about where customer data is stored at rest and, separately, where machine learning processing of it takes place.

IS5542 FAQ

Questions students ask in each IS5542 chapter

Why does a language model sound confident even when it is wrong?

Because fluency comes from predicting likely word sequences, not from checking facts. A sentence can be highly probable given the model's training and still be false for your situation, which is why grounding answers in supplied documents matters more than choosing a model that writes well. Chapter 01 · How Generative AI Works: From Rules to Transformers

Which benchmark should I cite for reasoning in mathematics?

It depends on the level of the claim. Grade-school word problems needing several steps are covered by one suite, while competition-level problems needing heuristics and proof are covered by another; citing the easier one for a claim about hard mathematics overstates what the evidence shows. Chapter 02 · Benchmarks and Leaderboards for Comparing LLMs

Should a business always choose the model at the top of the leaderboard?

No. The top model is usually strongest on general performance, which may not be the vector your application depends on. A voice service needs low latency and native audio, a batch job needs throughput and low cost, and paying for top reasoning where it is not needed raises cost without improving results. Chapter 03 · Choosing a Model: Six Vectors and Model Families

Why does the order of sense, probe and act matter so much?

Because it decides what the first version of a system should do. In a routine process the first step is reading and sorting inputs; in an uncertain one it is running cheap experiments; in a crisis it is taking immediate action. A system designed around the wrong first step will be either too slow or too careless for its situation. Chapter 04 · Framing the Business Problem with Cynefin Domains

Why does the same prompt give different results each time?

Generation samples from probabilities rather than retrieving a stored answer, so each run can choose different words or images. A precise prompt narrows the range of likely outputs but does not remove variation, which is why teams generate several options and review them rather than trusting a single run. Chapter 05 · Prompt and Context Engineering for Business

Why should the purpose of a How Might We statement be the customer's job?

Because a statement aimed at the company's goal, such as more revenue, invites any idea that sells more, including ones customers do not want. Naming the job keeps the solution anchored to a real need, and the revenue follows if the job is done better than by the alternatives. Chapter 06 · Jobs to Be Done and How Might We Statements

Which job roles are the best candidates for GenAI support?

Routine roles that firms struggle to fill and to keep, with little room for promotion. Automating their routine work gives the biggest productivity gain and frees people for more strategic tasks, whereas roles built on physical work or long-term judgement benefit far less. Chapter 07 · The Ten-Step GenAI Build and Operations Plan

Why is accuracy alone not enough to justify a GenAI rollout?

Accuracy is only the model tier. A system can be accurate and still fail because it is slow, unavailable or expensive, and a well-running system can still fail the business if nobody adopts it. The course measures all three tiers so that a rollout decision rests on the whole chain from model to money. Chapter 08 · Measuring Value: KPIs, Evaluation and ROI

How should I use the percentages quoted in the industry lectures?

As attributed illustrations. They come from vendor case studies, press reports or surveys, each with its own purpose and method, so in an essay name the source type, say what the number measures, and avoid treating one company's result as a general rate for the industry. Chapter 09 · GenAI Use Cases in Marketing and Retail

Why does the banking lecture keep attaching KPIs to each use case?

Because financial institutions approve projects on measurable outcomes. Tying a conversational agent to resolution rate, handling time or acquisition cost lets the bank compare it with the current process and decide whether to scale it, and it gives the regulator and auditors something concrete to review. Chapter 10 · GenAI in Hong Kong Banking and Insurance

Can I own the copyright in an image I generated with AI?

It depends on where and how. The US requires human authorship and has refused works credited to an AI system, while a Beijing court recognised copyright where the user's choice of prompts and settings was original, and UK law protects computer-generated works for fifty years. Documenting your creative input strengthens any claim. Chapter 11 · AI Law: Copyright, Ownership and Licensing

Who is liable when an AI system gives harmful advice?

Potentially several parties along the chain. The developer may face products liability or negligence claims, the organisation that deployed the system owes context-specific duties, and a professional who relied on the output without checking it remains personally accountable. Contract terms shift some risk between them but cannot remove it. Chapter 12 · AI Liability, Data Privacy and Algorithmic Bias

Does Hong Kong have an AI law?

Not a dedicated one. AI in Hong Kong is governed by existing common law and statutes, above all the 1995 privacy ordinance, together with the privacy commissioner's non-binding 2024 framework and sector guidance from bodies such as the banking regulator. The lecture notes that reform towards a GDPR-style regime is under consultation. Chapter 13 · Regulating AI in Mainland China and Hong Kong

Why does the course call responsible AI successful AI?

Because the controls that make AI responsible also make it commercially durable. The lecture cites benefits including brand differentiation, customer trust and loyalty, long-term profitability and readiness for regulation, so responsible practice is presented as part of the product's value rather than a cost added after launch. Chapter 14 · Responsible AI and Safety Mitigation

How to study for the exam

Pick one organisation and follow it through the whole course: choose its model by the vectors its workload cannot compromise, frame its problem in a Cynefin domain, write its job stories, plan its build and tiger team, set its KPIs, and ask which laws it would meet in Hong Kong and the Mainland. Then practise the essay move the review models: concept, business application, trade-off, in that order, for each point.

Your AI Information Technology tutor for IS5542

Stuck on a hard IS5542 question? Sia is AskSia’s AI Information Technology tutor — ask any IS5542 Generative Artificial Intelligence for Business question and get a clear, step-by-step explanation grounded in how the course is actually taught and assessed. Read this whole study guide free, then take your hardest questions to Sia.