PMGM7023 AI-Powered Management Analytics

How PMGM7023 is assessed

The four published components total 100%, and the syllabus states them as points as well as percentages: 200, 150, 100 and 50 out of 500. The grade table is headed as tentative, so confirm the current position on Moodle before relying on a weight for planning. Hurdle requirements are not stated in the course materials available here, so none is claimed in either direction. The team project carries a separate published deduction table, applied cumulatively after the three rubric ratings are summed, and a late-submission schedule of 10% up to one day, 20% beyond that but before the project debriefing, and 100% after the debriefing.

Current PMGM7023 dates

Dates are as published in Dates are taken from the course Moodle pages and the published schedule for this offering.. Confirm exact deadlines and submission settings in the live LMS.

What PMGM7023 covers

Twelve source-led chapters move from directing analytical work, through the vocabulary of the system you are directing, to collecting evidence you can defend and the assessed deliverable.

Its own syllabus is explicit that it is not designed for technicians but for future managers developing data literacy, and that no prior programming or analytics knowledge is required.

Ten sessions pair about two hours of lecture with about one hour of hands-on tutorial work in GitHub Copilot CLI, and the course states that for consistency and fairness all coursework must be completed with that tool.

The skill being built is direction rather than execution.

You learn to state a question precisely enough that an analysis can answer it, to acquire data that genuinely bears on it, to recognise when an output is confident and wrong, and to say what should happen next.

The opening session makes the point with a monitoring report whose arithmetic is perfect and whose headline is wrong, because several recorded contacts about one event were counted as several events.

How the sessions are organised

The first session covers managing an AI analyst: the progression from records to judgment, the difference between a conversation that returns advice and a supervised agent that returns files you can inspect, the six-decision analytics workflow, and the four question types that decide which analysis is relevant at all.

The second builds a working vocabulary in five layers, from what the system receives through the engine that generates the response, the evidence that supports it, the interfaces that let it act, and the controls that limit it.

The third session turns a management question into evidence: a six-line collection contract, eight families of data source, the discipline of aggregating every source to one analysis unit before joining, and the checks that separate genuine data from simulated data.

Its tutorial covers web scraping, including three permission checks that must all pass before any collection begins.

The fourth session is a clinic on data traps, covering sampling and nonresponse, measurement validity, descriptive summaries and denominators, and missing values and outliers, followed by a tutorial that cleans a deliberately messy dataset and reports what changed.

Later sessions listed in the published schedule move into survey analytics, hypothesis testing, measurement theory, regression and complex data, closing with an integrated framework, the project debriefing and the examination.

How it is assessed

Four graded components and nothing else.

The final examination is worth 40% and is an open-book multiple-choice paper held in the final session, in the usual lecture slot, with no make-up session and no electronic devices permitted. The team analytic project is worth 30% and is one written report with no presentation, judged on three dimensions of fifty points each.

Markdown files from the tutorials are worth 20% and are graded Plus, Check or No Record, with late submissions automatically receiving No Record. Class participation is worth 10% and is recorded session by session.

The project carries its own arithmetic.

Its body must be 3,500 words or fewer, its cover-page executive summary 250 words or fewer, and it must use at least two academic citations placed inside the body and attached to specific arguments.

Nine published deduction lines apply cumulatively: six formatting requirements at five points each, the citation requirement at ten, and twenty points each for data that are not genuine and for findings that cannot be reproduced from the submitted data and workflow.

A submission one day late or less costs 10%, later than that but still before the debriefing session costs 20%, and once the debriefing has passed the work receives nothing.

Where marks are most often lost

Three places, none of which is analytical skill. The first is the unit of analysis: a technique applied to the wrong row definition gives a precise answer to a question nobody asked.

The second is the denominator: a change is only large or small against a baseline, and a rate without its count can describe one person.

The third is format compliance on the project, where a team can lose a quarter of a performance band to word counts and file types without a single weakness in the argument.

How to use this guide

Each chapter takes one part of a session, names the decision it teaches, works one case through to a finished answer rather than describing what an answer would contain, and gives two prompts to write yourself.

Worked cases here are ours, built on new settings and new numbers, so that they train the method instead of the recognition of a familiar example. The closing chapters are a mixed practice set in the shape the examination uses and a pre-examination checklist. Confirm every operational detail, including deadlines and any current requirement, on the course Moodle page.

Worked example · free

Audit a unit before trusting a total

Key terms

Row Grain

Row grain is the statement of what one row of a dataset represents, such as one contact or one completed event. It has to be fixed before any count, rate or comparison can be interpreted.

Information Boundary

Information boundary is the set of facts available to whoever or whatever is calculating. Moving it, by supplying context an export omitted, changes the answer without changing the arithmetic.

Descriptive Analytics

Descriptive analytics summarises what happened, covering size, timing, distribution, concentration and segments. It establishes the pattern but cannot on its own establish a cause.

Diagnostic Analytics

Diagnostic analytics tests rival explanations for an observed pattern. Each rival usually needs a different comparison, so the data required follows from which explanation is being tested.

Prescriptive Analytics

Prescriptive analytics chooses an action under constraints, weighing feasibility, cost and consequence. It requires information about options and ownership that no dataset contains.

Retrieval-Augmented Generation

Retrieval-augmented generation is a question-time process that searches an approved source collection, adds the relevant passages to the model's context, and asks it to answer from those passages.

Hallucination

Hallucination is fluent output stating something without adequate support from the input, the evidence supplied or reliable knowledge. It is detected by asking where the support is, not by reading for confidence.

Guardrail

Guardrail is a rule, filter, permission, validation or halt that is actually enforced on an AI system. An instruction to be careful is not one, because nothing enforces it.

Context Window

Context window is the finite amount of information a model can use for the current response. Anything falling outside it cannot be used for that answer, whatever its importance earlier.

Survivorship Bias

Survivorship bias is the distortion that arises when only cases that persisted can be observed. It is repaired by naming the cases that failed, withdrew or vanished and never reached the file.

Measurement Validity

Measurement validity is the question of whether a recorded number captures the concept it is meant to capture. It fails through proxy, outcome and definition mismatches.

Winsorization

Winsorization is capping verified extreme values at a stated limit rather than removing the cases. It requires the limits to be set once on the original data and the number of changed values to be reported.

PMGM7023 FAQ

Questions students ask in each PMGM7023 chapter

What is the difference between AI chat and an AI agent?

A chat returns an explanation of what could be done, and the person carries out the work; the conversation is the whole interface. An agent, once granted permission, inspects files, writes and runs code, checks results, revises and produces deliverables. The lecture frames the difference as advice against artifacts: one leaves you with a suggestion, the other with cleaned data, logs and an owned conclusion. Chapter 01 · Managing an AI Analyst

How do I check whether a total in a business file is meaningful?

Write one sentence stating what a single row records, then ask how many rows one real-world event can produce. If a single event can generate several rows, the total counts contacts or attempts rather than events, and it has to be grouped by an event identifier before it can be compared with anything. Report both figures and label which one the decision uses. Chapter 02 · What One Row Represents

What are the four types of data analytics in this course?

Descriptive work establishes what happened through size, timing, distribution and concentration. Diagnostic work tests why it happened by comparing rival explanations. Predictive work estimates what may happen and states the uncertainty. Prescriptive work chooses an action under constraints. One situation usually contains all four, and the type determines what evidence the answer is obliged to carry. Chapter 03 · The Analytics Workflow and Its Four Questions

What should a manager specify before delegating an analysis?

Four things. The objective, naming the outcome, the unit and the comparison. The boundary, naming which files may be read and changed and where output goes. A checkpoint before the expensive step, so a wrong assumption surfaces while it is still cheap. And an acceptance test describing how the finished result will be confirmed, which is what stops a plausible answer being treated as a verified one. Chapter 04 · Human Judgment and Agent Execution

Where should I put a preference that applies to every task?

In the standing instruction rather than in each request. A preference repeated message by message consumes space in a finite working context and disappears as soon as earlier turns fall outside it. Written once as a standing instruction, it is read at the start of a session and applies across tasks, which is both cheaper and more reliable than restating it. Chapter 05 · Prompts, Context and the Model Engine

How can I tell whether an AI answer is actually supported?

Ask it to point at the line that carries the support. A system that has read or searched something can point at the passage; one that has produced a plausible continuation will restate the claim in different words. Polished wording is not proof, and a specific-looking reference with a journal and an identifier is one of the most common shapes an unsupported claim takes. Chapter 06 · Evidence, Tools and Control in AI Systems

Why start Copilot CLI in a specific folder?

Because the folder you start in is the boundary of what can be read or changed without a further approval. Launching from a default location usually means starting in your home directory, which grants the widest possible reach for no benefit. Starting in the course folder keeps the materials, the data and the output together and makes any request to go outside it visible before it happens. Chapter 07 · Working Inside GitHub Copilot CLI

What should be decided before collecting any data?

Six things, written down: the decision the data will support, what one row will represent, the fields needed to answer the question, the population and time window, one candidate source per field group, and the checks that will show the file is complete. The unit line and the validation line carry most of the value, because the first determines which fields are even expressible and the second turns completeness into arithmetic. Chapter 08 · Specifying and Sourcing the Data You Need

Can I scrape a website just because it is public?

No. Three checks must all pass first. The content must be reachable without logging in or paying; the site's terms must not prohibit automated access; and the site's crawler instructions must not disallow the specific path you want. A failure at any one gate ends the collection for that path, and collecting more slowly does not change permission. Record all three verdicts with the date checked. Chapter 09 · Web Scraping Within the Rules

Why can a very large sample still give the wrong answer?

Because size does not repair selection. If the method of reaching people leaves part of the population unreachable, or if those who respond differ systematically from those who do not, then collecting more responses through the same channel reproduces the same bias at greater volume. The question to ask is how the sample was chosen rather than how many observations it holds. Chapter 10 · Data Traps and Defensible Descriptive Claims

Why describe a messy file before cleaning it?

Because the description is the only evidence that the cleaning was correct. A first summary that has already been tidied cannot show what changed, and the defects you are looking for, such as one category written three ways or a negative placeholder in an age column, are visible only while the raw values are intact. The same summary is produced again afterwards so the two can be compared. Chapter 11 · Cleaning a Real Dataset With a Supervised Agent

How many points can formatting mistakes cost on the project?

Six published lines cost five points each and one costs ten, so a project with a sound argument can lose twenty-five points to word counts, a missing cover-page element, the wrong file format and uncited references. Two further lines cost twenty points each, for data that are not genuine and for findings that cannot be reproduced, and neither of those can be repaired after submission. Chapter 12 · The Team Analytic Project

How to study for the exam

Work one session at a time and finish each with a one-page summary carrying the decision that session teaches, the two or three distinctions it turns on, and the trap it warns about. Practise routing unseen requests into the four question types and diagnosing described failures by system layer, since those are the judgements an open-book paper can still test.

Keep every tutorial file complete and submitted on time, and run the project deduction list as a checklist a week before the deadline rather than from memory.

Your AI Management tutor for PMGM7023

Stuck on a hard PMGM7023 question? Sia is AskSia’s AI Management tutor — ask any PMGM7023 AI-Powered Management Analytics 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.