AI, ML and generative AI

Each term is a smaller circle inside the last

Artificial intelligencemachines doing tasks that need intelligenceMachine learninglearns patterns from data, not written rulesGenerative AIcreates new text, images, audio and codefrom what it learnedAgentic AIuses generative models toplan and act toward a goalEach term is a smaller circle inside the last. Agentic AI is what generative AI does when it is given tools.All machine learning is AI, but not all AI is machine learning. Generative AI is one kind of machine learning.An exam option that says "AI" is never wrong for a generative task, only less exact.

AI contains machine learning, which contains generative AI. Agentic AI puts generative models to work.

AI, ML and generative AI · Going deeper

Analytics, BI, AI, ML and generative AI

Five terms the guide asks you to tell apartTermWhat it doesThe question it answersData analyticsExamines data to find patterns and causesWhat happened, and why?Business intelligencePresents analysed data to the business in reportsHow are we doing?Artificial intelligencePerforms tasks that normally need human judgementCan a machine do this?Machine learningLearns from examples to predict or classifyWhat is likely to happen?Generative AIProduces new content from a promptCan you draft, summarise or create this?Analytics and business intelligence describe the past. Machine learning predicts. Generative AI creates.A dashboard of last quarter's sales is business intelligence, even when the vendor calls it AI.

Analytics and BI describe the past, ML predicts, generative AI creates.

What ML is for

A model learns the pattern that nobody could write as rules

RULES WRITTEN BY PEOPLEif the amount is over 5,000: checkif the country is new: checkif 3 tries in a minute: blockFraud changes. The rules do not.A MODEL LEARNED FROM EXAMPLESmillions of past payments,each marked fraud or genuineModela new paymentfraud risk: 93%Retrain on new examples and it keeps up.Machine learning replaces rules that people can no longer write or keep up to date.Nobody can list every sign of fraud, or every way a customer phrases a complaint. A model finds the patternfrom examples, in tables or in text and images, and then makes the same decision millions of times a day.

ML replaces rule-based systems, reads structured and unstructured data, and scales decisions.

What ML is for · Going deeper

Three kinds of value

Three kinds of value from machine learningWhat it meansExampleReplacing rule-based systemsPatterns learned, not hand-writtenfraud detection, spam filteringInsight from all kinds of dataTables, and also text, images and audioreading claims forms and photosScaling decisionsThe same judgement, millions of times, at oncea price for every product, dailyMachine learning is the right tool when there is a pattern, there are examples, and rules would be endless.It is the wrong tool when a simple rule does the job, or when there is no data to learn from.

Replace rules, use all kinds of data, and scale the same decision.

Check yourself

three years of sales,with each day's weathera patternis learneda forecast fornext week

Which term names what the drawing shows?

Agentic AI

An agent is given a goal, where a model is given a prompt

Generative AIa promptmodelan answerAgentic AIa goalplanact with toolstools: search, databases, email, other agentschecknot there yet: plan againdonea person approveswhat mattersGenerative AI answers a prompt. An agent is given a goal, and works until it is met.The agent decides its own steps, uses tools to act in other systems, and checks its results. That is why itchanges work and not only writing: it can complete a task, such as resolving a refund, from start to end.

An agent takes a goal, plans, acts with tools, checks, and repeats until it is done.

Agentic AI · Going deeper

Where agents change the work

Six areas the guide says agentic AI changesAn agent canWorkforce productivityPrepare a meeting brief from mail, documents and the calendarCustomer supportResolve a request from start to end, and hand hard cases to a personSalesResearch an account and draft the proposalProduct innovationTest many design options and summarise what users saidOperationsWatch a supply chain and reorder before stock runs outResearchRead thousands of papers and report what is knownIn each area the agent does the routine steps, and people keep the judgement.Exam answers that have an agent acting with no human oversight at all are usually the wrong ones.

Productivity, customer support, sales, product innovation, operations and research.

Data quality

Six tests find the faults a model would learn

customeremailcountryjoinedupdatedCompletenessa value is missingUniquenessthe same customer twiceValidityno such dateConsistencySpain written two waysTimelinessnot updated since 2019Accuracyright form, wrong countryAna Ruizana@example.comSpain2021-03-142026-10-02Ben Okafor—Nigeria2022-07-012026-09-18Ana Ruizana@example.comSpain2021-03-142026-10-02Chi Tranchi@example.comVietnam2023-02-312026-08-30Dev Shahdev@example.comES2020-11-092026-10-11Eva Lindeva@example.comSweden2019-05-202019-05-20Fay Wongfay@example.comCanada2024-01-082026-10-20A model learns whatever is in its data, faults included. Six tests find the faults.Five of the six can be seen by looking at the table. Accuracy cannot: the last record is well formed, andwrong. Only checking against the real world finds it. Poor quality is the usual reason an AI project fails.

Completeness, uniqueness, validity, consistency, timeliness and accuracy.

Data quality · Going deeper

The six dimensions

The six dimensions of data qualityDimensionThe questionA failure looks likeCompletenessIs everything that should be there, there?empty fields, missing recordsUniquenessIs each thing recorded only once?the same customer under two idsTimelinessIs it recent enough for the use?last year's prices in today's modelValidityDoes it follow the required format and range?a date of 31 FebruaryAccuracyDoes it match the real world?the wrong address, well formattedConsistencyDoes it agree across systems?"ES" in one place, "Spain" in anotherValidity is about form. Accuracy is about truth. The exam likes to offer one for the other.A value can be valid and inaccurate, as a real-looking but wrong phone number is.

Validity is about form. Accuracy is about truth.

Check yourself

idcustomeremail2291Omar Haddadomar@example.com2292Lena Kochlena@example.com2317Omar Haddadomar@example.com

Which dimension of data quality do these records fail?

Explainable and responsible AI

A decision people cannot question is a business risk

OpaqueapplicationmodeldeclinedNo reason is given, so nobody cancheck the decision or challenge it.Explainabledeclined, and this is how much each reason counted:debts high against income45%short credit history35%a recent missed payment20%Responsible AI asks more than "does it work?"fairaccountableprivate and securesafetransparentExplainable AI can say why it decided. Responsible AI is built and used so that people can trust it.The business case is plain: regulators and customers demand reasons, an unfair model is a legal andreputational risk, and staff will not act on advice they cannot question.

Explainable AI shows why it decided. Responsible AI is fair, accountable, secure, safe and transparent.

Explainable and responsible AI · Going deeper · 1 of 2

Why Google Cloud for AI

Five benefits the guide gives for Google Cloud AIBest infrastructure for AIThe AI Hypercomputer: chips and systems designed for training and running modelsAI-ready data cloudData already in BigQuery and Cloud Storage can be used by models where it isFirst-party modelsGoogle builds its own models, such as Gemini, and offers others beside themAll-in-one developer platformOne platform to choose, tune, deploy and govern models and agentsPre-built agents and applicationsReady-made agents for common work, usable without building anythingThe five run from the chips at the bottom to finished agents at the top.They are the AI half of the six differentiators from lesson 2. The products are the subject of the next lesson.

Infrastructure, an AI-ready data cloud, first-party models, one platform, and pre-built agents.

Explainable and responsible AI · Going deeper · 2 of 2

Clue words for AI concepts

When the question saysThe answer isBecause"predict", "forecast", "classify"Machine learninglearns from examples"draft", "summarise", "create an image"Generative AIproduces new content"plans and completes a task itself"Agentic AIgoal, tools, action"report on last quarter", "dashboard"Business intelligencedescribes what happened"too many rules to maintain"Machine learningreplaces rule-based systems"missing values"Completenesssomething is absent"duplicate records"Uniquenessrecorded more than once"out of date"Timelinesstoo old for the use"wrong format"Validitybreaks the required form"does not match reality"Accuracywell formed, but untrue"systems disagree"Consistencydiffers between sources"why did the model decide that?"Explainable AIreasons can be shownHalf of these are data quality. Expect at least one question on a dimension.

Know the four AI terms and the six quality dimensions by their clue words.

Check yourself

goal: rebookevery passenger?plans the steps, reads thebooking system, sends emails,checks each resultdone; hardcases to staff

An airline wants the box marked "?" to deal with a cancelled flight. What is it?

Lesson complete

You finished “AI and machine learning concepts”

Answer the questions along the way to see your score here.

Open in a new tab