The Exam Room · AIF-C01
AI Practitioner
The foundational track: what the services do and when each one fits. Preparing for the AI Practitioner certification.
The domain-by-domain cheat sheet sorts the whole track into the exam's scored domains.
Exam Room · AI Practitioner
Telling AI, ML, Deep Learning, and Agentic AI Apart
Five systems, one governance review, five labels to assign. A rules engine, a churn model, an image checker, a summariser and a tool-calling agent are not the same kind of thing, and the label decides who reviews each one, what data it needs, and what it can explain about itself.
Read articleExam Room · AI Practitioner
Cheat Sheet: ML Fundamentals and the SageMaker Suite
A condensed revision sheet for the classic machine-learning ground the AI Practitioner exam assumes: the learning-problem taxonomy, overfitting and drift, and the SageMaker services that get confused for one another.
Read articleExam Room · AI Practitioner
Turning a Business Question Into an ML Problem
Six departments in a produce-delivery business arrive in the same fortnight with six asks, and every one of them gets introduced with the words 'can AI do this'. The route from a business question to a technique runs through two smaller questions: what does the answer look like, and has anyone written down the right answer before. Settle those and the AWS service mostly picks itself.
Read articleExam Room · AI Practitioner
Mapping an AI/ML Pipeline Onto AWS Services
A churn model that works in a notebook and a Bedrock summariser that works in a demo both have to become something a business can run. Drawing that on one page means naming the stages of an AI/ML pipeline without collapsing four data stages into one, and putting an AWS service against each. The two pipelines share a shape and differ in the middle.
Read articleExam Room · AI Practitioner
Traditional Model or Foundation Model
A lender has two features on the same backlog: a credit decision that has to be explained to the applicant it refused, and an email triage feature with no labelled history and a fortnight to ship. One of them wants a model trained on the lender's own data. The other wants a model somebody else already trained. The split falls out of what has to be explained and what has already been labelled.
Read articleExam Room · AI Practitioner
What Your Data Decides Before You Pick a Model
Four datasets land on one team in the same week: a subscriptions table in Amazon Redshift, three years of daily delivery counts, a folder of scanned delivery notes, and 400,000 support emails nobody has ever sorted. Only two of them can become a model this quarter, and the algorithm has nothing to do with it. What the data already is, and how many of its examples carry an answer, settles which learning method is available at all.
Read articleExam Room · AI Practitioner
When a Prediction Is the Wrong Answer
Four proposals land in the same budget round at a box-delivery business, and all four arrive described as AI projects. Two of them want a model. The other two want a refund calculation and a statutory eligibility check to come out the same way every time, which is a rule, not a prediction, however good the accuracy figure on the slide looks.
Read articleExam Room · AI Practitioner
How Much MLOps a Model Actually Needs
A churn model has been scoring subscribers every night for eight months, is getting quietly worse, and cannot be rebuilt because nobody knows which folder of training data produced it. How much operational machinery to put around a model is a real decision with three sensible answers, and the cheapest one is right more often than engineers like to admit.
Read articleExam Room · AI Practitioner
Measuring Whether a Model Earned Its Keep
A recommendation model arrives at the quarterly review with an F1 score of 0.81, up from 0.74, and the finance lead asks the one thing nobody prepared for: is it worth what we are paying for it. Model performance metrics say whether the model is right. Business metrics say whether being right returned more than it cost, and the two can move in opposite directions.
Read articleExam Room · AI Practitioner
Choosing How a Model Serves Its Predictions
One trained fraud model, four teams asking to call it, and four completely different jobs behind the same request. A checkout page needs a score in 200ms, an overnight job needs 2.4 million scores with nobody waiting, an analyst screen runs sixty times a day, and a document pipeline needs several minutes on a 400MB upload. Real-time, serverless, asynchronous and batch are four shapes of job rather than four sizes of the same one.
Read articleExam Room · AI Practitioner
Which Stage of the Foundation Model Lifecycle Is Yours
A retailer with 9,000 product pages and no descriptions has three proposals on the table: prompt a model on Amazon Bedrock, fine-tune one on four years of house copy, or train one on the catalogue from nothing. They are three entry points into the same seven-stage sequence, and what separates them is which stages the team signs up to own.
Read articleExam Room · AI Practitioner
What Belongs in the Context Window
A support assistant on Amazon Bedrock gets answers wrong in two ways: sometimes it does not know the refund policy, and sometimes it forgets what the customer said four turns ago. Both are the same shortage. Everything a model sees on one call arrives through one finite budget of tokens, and deciding what fills that budget is context engineering.
Read articleExam Room · AI Practitioner
When an AI Application Becomes Agentic
An internal travel assistant answers policy questions well and fails at every booking, because one call to a model can only write about acting. Giving it tools turns that call into a loop that decides what to do next, and the loop changes what the system costs, how predictable it is, and whether anyone can reconstruct afterwards what it did.
Read articleExam Room · AI Practitioner
What a Token Costs and What Changes the Bill
A claims team put a Summarise button on twelve-page documents and got a bill four times the estimate. Tokens are what a foundation model reads and writes, and tokens are what Amazon Bedrock meters, so the length of the prompt, the length of the answer, and the number of times the same document gets sent again are the three things moving the number.
Read articleExam Room · AI Practitioner
Where Generative AI Helps and Where It Hurts
An insurer has three features on one backlog: a claims-summary drafter, a policy-eligibility decision, and a customer chat assistant. The properties that make a generative model adaptable and conversational are the same ones that make it nondeterministic and hard to interpret, and they do not come apart. What separates the three features is whether a wrong answer is embarrassing or actionable.
Read articleExam Room · AI Practitioner
The AWS Building Blocks for a GenAI Application
Four teams ask for generative AI in the same week and each of them means something different by it. Amazon Bedrock, SageMaker AI, JumpStart, Amazon Quick, Kiro, Strands Agents and Bedrock AgentCore are seven building blocks, and two questions sort them: who the output is for, and how much of the stack you want to run yourself.
Read articleExam Room · AI Practitioner
Proving a GenAI Feature Paid for Itself
Six weeks after a retailer shipped a shopping assistant on Amazon Bedrock, the finance review wants to know whether to keep paying for it, and the only numbers anyone has prepared are model-quality scores. Those say the model works. Whether the business is better off is a different set of numbers, measured against a baseline somebody had to capture before launch.
Read articleExam Room · AI Practitioner
What to Weigh When You Pick a Foundation Model
A clinic-software company is adding three generative features at once, and the instinct is to wire the strongest model in the catalogue into all three. Model type, performance requirements, capabilities, constraints, compliance, cost, latency and model complexity are eight separate filters, and they rarely agree on one answer. Run them per job and the shortlist stops being a leaderboard.
Read articleExam Room · AI Practitioner
Answering Questions From Your Own Documents
A regional insurer wants its staff to ask plain questions of four thousand policy documents and get back an answer with a source they can open. A foundation model on its own cannot do that: the documents were never in its training data, and they do not fit in a prompt. Retrieval Augmented Generation closes the gap, and the decision that follows is where the vectors live.
Read articleExam Room · AI Practitioner
Deciding How Far to Customise a Foundation Model
A veterinary practice assistant reads clinical abbreviations wrong, quotes last year's medicine schedule, and answers in a tone the practices dislike. Three teams have proposed three different fixes, and a fourth voice wants to train a veterinary model from scratch. Six ways to customise a foundation model sit on a cost ladder, and which failure you actually have decides how far up it you need to climb.
Read articleExam Room · AI Practitioner
Picking a Prompting Technique for the Task
Four prompts in one logistics product fail in four different ways: one rambles, one invents categories that do not exist, one gets multi-step arithmetic wrong, and one keeps drawing text into the picture. Adding more examples fixes exactly one of them. Choosing a prompting technique starts with naming which part of the prompt is missing.
Read articleExam Room · AI Practitioner
Choosing a Foundation Model for the Job
A travel booking company ships three model-backed features in one release: a chat assistant answering in three languages, an overnight job that reads two-hundred-page contracts, and alt text for four hundred thousand hotel photographs. Modality, latency, multi-lingual coverage, input/output length and cost are separate filters, and they disagree with each other. Run them per feature, then set the inference parameters on purpose.
Read articleExam Room · AI Practitioner
Cheat Sheet: Generative AI Foundations
A condensed revision sheet for the generative-AI groundwork the AI Practitioner exam tests before any Bedrock service enters the picture: the architectures behind the models, the vocabulary around them, their known limitations, and the prompting techniques that predate any engineering track.
Read articleExam Room · AI Practitioner
Judging Whether a Foundation Model Is Good Enough
A news site wants to move its article summaries onto a cheaper model, and nobody can say whether quality would drop. Three different questions are hiding inside "is it good enough": does the output match a reference answer, would a person call it good, and does it clear the bar the job needs. Each one has its own measurement, and reaching for the wrong one is how a summariser gets scored with a translation metric.
Read articleExam Room · AI Practitioner
Which Kind of Training a Foundation Model Needs
A marine insurer has exhausted the cheap fixes and accepted that the model itself has to change. Four training jobs are on the table, and they eat completely different data: raw text, labelled pairs, a bigger model's answers, or human rankings. Which job to run is settled by which data the company actually holds and can afford to prepare.
Read articleExam Room · AI Practitioner
When an AI Agent Earns Its Place
An equipment-hire company wants three jobs automated, and one proposal has been made for all three: build an agent. Only one of the three has steps that nobody can write down before the request arrives, and that is the one an agent suits. The other two come out cheaper, faster and far easier to explain without one.
Read articleExam Room · AI Practitioner
Cheat Sheet: AWS AI Services
A condensed revision sheet for the AWS AI service list: one line each for the purpose-built services, the generative build surface and the named Bedrock features, plus the pairs that get mixed up under pressure.
Read articleExam Room · AI Practitioner
Keeping Prompts Safe and Versioned
Three incidents in one month at a retail bank: the assistant printed its own system prompt into a customer chat, a pasted support macro talked it out of its refusal rules, and a Friday wording change had it quoting fee amounts nobody could source. The prompt was a string literal in a Lambda, so there was nothing to roll back to. Naming the four risks correctly is where the repair starts.
Read articleExam Room · AI Practitioner
Measuring Whether an AI Feature Is Working
Three months after a telco shipped its support assistant, the model scores are healthy, deflection is up, and the head of support wants the thing switched off. Scores that grade an answer against a reference say nothing about whether the feature is worth running. Those are two different measurements, and only one of them renews the budget.
Read articleExam Room · AI Practitioner
Deciding Whether a Dataset Is Fit to Train On
Forty thousand support tickets, four years of them, filed into twelve queues by three people who never once compared notes. Training a triage model on that export costs a fortnight and a five-figure bill, and nobody has checked whether the data can carry a model at all. Two unrelated defects hide in a set like this, and each needs a different fix.
Read articleExam Room · AI Practitioner
Explaining an AI Decision to the Person It Affects
A storm-damage claim is declined and three people ask why: the claimant, the compliance lead, and the developer who spotted a flood clause in a letter about a patio roof. Same word, three different artefacts, and handing any of them the wrong one produces something nobody can act on.
Read articleExam Room · AI Practitioner
When the Decision Has to Be Explainable
A lender has to decline some applications and tell each applicant why. That obligation reaches back into the model choice: the more accurate model cannot say why it declined anyone, and the model that can say why is two points worse. This is the trade AWS asks you to name, and it is settled with numbers rather than preference.
Read articleExam Room · AI Practitioner
Reducing the Legal Exposure of a Generative Feature
A retailer wants generated product descriptions and campaign images across 60,000 lines, and legal has refused to sign it off. Counsel's memo lists five worries, and they close in three different places: a contract, a pipeline, and a person. Sorting which is which is most of the work.
Read articleExam Room · AI Practitioner
Keeping Watch on Bias After Launch
A loan-triage model passed its fairness check six months ago and nobody has looked since. Overall accuracy has not moved all year, which is why nobody noticed that one cohort's approval rate fell sixteen points. Building the ongoing watch means deciding which instrument covers which part of the job, and how often each one has to run.
Read articleExam Room · AI Practitioner
Diagnosing a Model That Works for Most People
A delivery-window model hits 94 per cent overall and misses badly for forty outer postcodes and for everyone over seventy. Three things cause that, and their repairs pull in opposite directions, so guessing deepens the failure. The diagnosis comes out of one comparison run twice: training against validation error, overall and then slice by slice.
Read articleExam Room · AI Practitioner
Picking a Model When Sustainability Is on the Scorecard
A building-services contractor has a published emissions target and a new summarisation feature to build, and procurement wants the energy cost of the model choice written on the same page as the accuracy and the bill. Model size, tokens processed and idle capacity are the three things that move that number, and all three are settled before a single request is served.
Read articleExam Room · AI Practitioner
Matching an AI Security Worry to an AWS Control
A small insurer's security review of its Bedrock assistant produces six worries and no agreement on what fixes what. An identity problem, a network path, an uncatalogued bucket, an encryption question, a model that talks about competitors, and an agent about to call the refunds API. Each has one control that answers it, and several that sound like they might.
Read articleExam Room · AI Practitioner
Sorting AI Security Risks Into the Layer That Owns Them
Eight incident reports from one quarter of running a retail assistant, and every review meeting stalls in the same place: somebody proposes a fix before anyone has said what kind of problem it is. A four-year-old library with a CVE, an insult in a reply, an order number belonging to a different customer. Sorting each one into the layer that owns it decides the fix.
Read articleExam Room · AI Practitioner
Cheat Sheet: Responsible AI, Security, and Governance
A condensed revision sheet for the two smallest-weighted but most-confused domains: the eight AWS responsible-AI dimensions with their canonical scope, and the security and governance vocabulary that sits underneath every AI workload, not just a generative one.
Read articleExam Room · AI Practitioner
Choosing the Service That Produces the Evidence
An assessor asks a healthcare startup five questions about its Bedrock summarisation feature and gives it a fortnight. Five different AWS services answer them, and the team keeps reaching for the wrong one because all five sound like 'the compliance service'. What separates them is the object each one inspects and who wrote the evidence down.
Read articleExam Room · AI Practitioner
Keeping an Assistant From Making Things Up
A council assistant has invented a bin-collection rule, quoted a fee that stopped applying two years ago, and cited a real policy document for something that document does not say. Three complaints in a fortnight, three different faults, and a fix for one of them that does nothing at all for the other two.
Read articleExam Room · AI Practitioner
Four Things a Dataset Has to Be Before a Model Sees It
Eight years of job applications, and four people in one meeting each want something different to be true of them before anyone fine-tunes anything. Is the data right, is it safe to hold, who may read it, and is it still what it was are four questions, not one. Sorting a requirement into the right one decides which AWS service answers it.
Read articleExam Room · AI Practitioner
Deciding Where AI Data Lives and How Long It Stays
Legal sends three questions about a Bedrock study assistant and the team cannot answer any of them: where does student data physically sit, how long do the conversation logs stay, and how would anyone notice if either answer stopped being true. A generative AI feature leaves a fresh copy of the data at every stage, and each copy has its own home, its own clock, and its own way of outliving the policy.
Read articleExam Room · AI Practitioner
Setting Up Governance Before the First AI Feature
Four AI uses land on a 200-person logistics company in the same month, and one operations manager is told to have governance in place before any of them ships. She does not need four programmes. She needs a register, a one-page policy, and a rule that decides which of the four has to pass which review, and how often it comes back.
Read articleExam Room · AI Practitioner
Showing Where an AI System's Data Came From
A regulator disputes one sentence a research assistant produced, and four people in the same company ask where did this come from in four differently shaped ways. One wants the passage, one wants the pipeline run, one wants the list of datasets and their owners, and one wants the model's build history. Four questions, four artefacts, and reaching for the wrong one wastes a fortnight.
Read articleExam Room · AI Practitioner
A Domain-by-Domain Checklist for the AI Practitioner Exam
The AI Practitioner certification scores five domains, weighted from 28% down to 14%. This is the whole AI Practitioner track on one page, sorted into those five domains: the cheat sheet that anchors each one, the decisions to work through, and the quizzes and cards to drill. Tick your way down, and wherever a line still feels shaky, that is where the next hour goes.
Read articleFlash cards
One subject a card: the five things that matter, when to pick it over the alternative, and why it is the wrong answer when it is.
Exam Room · AI Practitioner
Flash Card: Choosing a Foundation Model
The factors that decide which foundation model a feature runs on: model types and modality, capabilities against the performance requirements, context window, latency and cost, and the constraints and compliance rules that rule candidates out first.
Read articleExam Room · AI Practitioner
Flash Card: What GenAI Gets Used For
The generative-AI use cases the exam guide names, and the AWS service that usually runs each one.
Read articleExam Room · AI Practitioner
Flash Card: Amazon Bedrock AgentCore
The managed runtime and supporting services for running agents in production on AWS: a serverless runtime that hosts whatever framework built the agent, plus memory, a tool gateway, scoped identity and a trace of every call.
Read articleExam Room · AI Practitioner
Flash Card: Strands Agents
The open-source AWS framework for writing an agent in code: you supply a model, a set of tools and a prompt, and the model plans the steps. It is the code-first neighbour of the managed agent services, and it is a library rather than a service, which is where Amazon Bedrock AgentCore comes in.
Read articleExam Room · AI Practitioner
Flash Card: Vector Storage on AWS
Four AWS services can hold embeddings and answer a similarity query, and each one is the named answer for a different reason. Learn them by the phrase in the scenario that gives each away.
Read articleExam Room · AI Practitioner
Flash Card: Temperature, Top-P and Max Tokens
The four sliders on a model call that change the answer without changing the model, and what each one is actually for.
Read articleExam Room · AI Practitioner
Flash Card: Amazon Bedrock Guardrails
A configurable safety layer that inspects prompts on the way in and responses on the way out, sitting beside the model rather than inside it, with one policy type per feature of responsible AI it is there to protect.
Read articleExam Room · AI Practitioner
Flash Card: AWS AI Service Cards
AWS's own transparency document for a managed AI service, written by AWS about AWS. Learn it by who holds the pen, because that is what separates it from the card you write yourself.
Read articleExam Room · AI Practitioner
Flash Card: The Generative AI Security Scoping Matrix
AWS's five-scope framework for placing a generative AI workload by how much of the stack you built. Learn the five scopes, learn the one rule that decides placement, and most governance scenarios sort themselves.
Read articleExam Room · AI Practitioner
Flash Card: AgentCore Identity and Policy
Two names on the security objective that answer two different questions: who is asking, and what that caller may do once it is let in. Neither replaces IAM.
Read articleExam Room · AI Practitioner
Flash Card: What Bedrock Does With Your Prompts
Almost everyone new to hosted models assumes the service is quietly learning from what they type into it. It is not, and five lines settle what actually happens to a prompt: who sees it, how it is encrypted, which country it is processed in, where it is written down, and what a fine-tune leaves behind.
Read articlePop quizzes
Exam-style multiple choice, one question a card, with the reasoning behind the tempting wrong answer.
Exam Room · AI Practitioner
Pop Quiz: A Rule, Not a Prediction
A classifier that decides which allergen warnings go on the label, 99.4% accurate on four years of past labels. The accuracy figure is doing all the persuading, and it is the wrong number to be persuaded by.
Read articleExam Room · AI Practitioner
Pop Quiz: Batch or Asynchronous
300MB scanned bundles, eight minutes each, a few dozen a day at no predictable hour, and the person who uploaded one wants a notification when it is done. Two of the four inference types are described as 'not real-time', and only one of them fits.
Read articleExam Room · AI Practitioner
Pop Quiz: No Labels, No Target
Three years of subscriber behaviour, no labels, no target column, and a marketing team who want to know whether there are natural groups worth treating differently. Four of the five techniques on offer need something the data does not have.
Read articleExam Room · AI Practitioner
Pop Quiz: Which Managed AI Service
Recorded calls to text, transcripts checked for personal data and scored for sentiment, summaries in Spanish, and a voice bot on the front door. Four jobs, four services, and two pairs that get swapped more often than any others.
Read articleExam Room · AI Practitioner
Pop Quiz: Managed API or Self-Hosted
A proprietary model the provider offers only through an API, four application developers and nobody who runs ML infrastructure, and a few hundred summaries a day arriving in two bursts. Two of the five routes would leave the insurer running GPU instances nobody asked for.
Read articleExam Room · AI Practitioner
Pop Quiz: The Model Is Not Wrong, the World Moved
A demand model that was accurate at launch has been getting steadily worse for fourteen months. The columns are the same, the input distributions are the same, and nobody has touched the pipeline. What did change was the business.
Read articleExam Room · AI Practitioner
Pop Quiz: Ninety-Nine Percent and Blind
A fraud classifier reports 99.2% accuracy. The fraud team says nothing is being caught. Both are true at once, and the reason is printed in the dataset rather than in the model.
Read articleExam Room · AI Practitioner
Pop Quiz: Generative or Agentic
Three systems in a support team, all three built on the same kind of model, and only one of them is agentic AI. Sorting them takes one question: does the system act, or does it hand its output to somebody who acts?
Read articleExam Room · AI Practitioner
Pop Quiz: Which Job Suits a Foundation Model
Four projects land on one desk in the same week: catalogue images and a promo clip, call summaries, email routing into eleven queues, and a shopping assistant. Three of them are foundation-model work. One only looks like it.
Read articleExam Room · AI Practitioner
Pop Quiz: Input Tokens, Output Tokens, and the Bill
A summarisation service sends six thousand tokens of document and gets two hundred tokens of summary back, and the bill tracks the documents rather than the summaries. The team expected the opposite. Working out why is the whole of the token-based pricing model in one sum.
Read articleExam Room · AI Practitioner
Pop Quiz: What MCP Is For
An agent that reaches a ticketing system, a document store and an HR API has three hand-written integrations behind it, and a fourth system means a fourth. Model Context Protocol is the standard that stops the count from growing that way, and it is not the four things it gets mistaken for.
Read articleExam Room · AI Practitioner
Pop Quiz: Naming the Generative AI Limitation
An audit team runs the same prompt through the same model three times and gets three different answers. All three are correct. Naming what went wrong is harder than it looks, because the word most people reach for is the one that does not apply.
Read articleExam Room · AI Practitioner
Pop Quiz: Which Metric Convinces Finance
An ecommerce team has a ROUGE score, a human-preference win rate, and six weeks of a live recommendation assistant. The CFO wants to know whether to fund it again next year. Two of those three numbers describe the model, and none of them describes the return.
Read articleExam Room · AI Practitioner
Pop Quiz: Bedrock, SageMaker AI, or JumpStart
A named open-weight model, a VPC it has to run inside, instance types the team wants to choose, and no training at all. Five services from the same list, and only one of them is a way to put those weights on infrastructure you pick.
Read articleExam Room · AI Practitioner
Pop Quiz: Which Stage of the FM Lifecycle
The assistant shipped a month ago and users are now clicking thumbs-up and thumbs-down under every answer. Somebody in the stand-up asks what that activity is called and where it goes. Name the stage.
Read articleExam Room · AI Practitioner
Pop Quiz: Context Engineering or Fine-Tuning
An internal assistant answers refund questions confidently and gets them wrong, because the policy changed last month and nothing on the call tells the model so. One team wants to fine-tune, another wants a bigger model. The cheapest correct lever is neither.
Read articleExam Room · AI Practitioner
Pop Quiz: Facts That Change Weekly
Delivery windows and stock policy get rewritten every Monday, and the assistant quotes last month's. The team has costed a monthly fine-tuning run. There is a cheaper approach that is also correct on the Tuesday.
Read articleExam Room · AI Practitioner
Pop Quiz: Where the Embeddings Go
Subscribers and catalogue already live in Amazon Aurora PostgreSQL. The team wants semantic search over product descriptions and has said plainly that it will not run, secure, back up and pay for a second data store. Which store holds the vectors?
Read articleExam Room · AI Practitioner
Pop Quiz: Examples or Reasoning
A prompt works out refund amounts against a tiered policy with a pro-rata rule and a cap, and gets the arithmetic wrong about one time in five. Eight worked examples have already been added and the error rate has not moved. What changes next?
Read articleExam Room · AI Practitioner
Pop Quiz: Naming the Prompt Attack
A supplier uploads a spec sheet with one odd sentence buried in it, and a week later the assistant is reading the internal pricing table out to anyone who asks. Four named risks describe part of what happened. Only one names the mechanism.
Read articleExam Room · AI Practitioner
Pop Quiz: ROUGE or BLEU
Two candidate models, 500 incident reports, and a human-written summary of each one already sitting in a spreadsheet. The team wants a number per model by Friday. Which automatic metric scores a summariser?
Read articleExam Room · AI Practitioner
Pop Quiz: Distillation or Continued Pre-Training
A support assistant on a flagship model answers well and nobody wants to change that. It costs too much per call and takes too long to reply, at forty thousand calls a day. Which training route makes it cheaper without making it worse?
Read articleExam Room · AI Practitioner
Pop Quiz: One Prompt or an Agent
A nightly report runs the same four steps it ran last night, and somebody has proposed building it as an agent. When every step is known before the job starts, the loop that makes AI agents useful has nothing left to decide.
Read articleExam Room · AI Practitioner
Pop Quiz: The Metric the Business Reads
A self-service assistant is up for renewal with a strong ROUGE score and containment up 30%. Repeat contacts within a day are also up, and the agents taking the handovers say they arrive worse. Which number settles it?
Read articleExam Room · AI Practitioner
Pop Quiz: Capping the Response
A product-description generator writes two or three paragraphs more than the page has room for, and resends the same three thousand tokens of brand guidelines on every call. Two settings deal with both halves of that bill, and neither of them is temperature.
Read articleExam Room · AI Practitioner
Pop Quiz: Rolling Back a Prompt
Five services share one piece of customer-facing wording. Somebody edited it last Tuesday, the answers got worse, and nobody can say what it used to say or put it back without five deploys.
Read articleExam Room · AI Practitioner
Pop Quiz: The Smaller Model and the Emissions Target
A published emissions target, a summarisation feature to build, and a sign-off form with a box for the energy cost. Four ways to answer it, and only one of them reduces the compute rather than accounting for it.
Read articleExam Room · AI Practitioner
Pop Quiz: Who Owns the Generated Image
Marketing wants generated images and generated copy on the public site. Legal wants to know what happens when a photographer says the picture is theirs. Four of the five controls on the table answer a different worry entirely.
Read articleExam Room · AI Practitioner
Pop Quiz: Imbalanced Data or Bad Labels
A fraud training set holds 200,000 legitimate transactions and 1,100 fraudulent ones, and a sample review finds two annotators disagreed on roughly one label in six. Two defects, and only one of them is fixed by rebalancing.
Read articleExam Room · AI Practitioner
Pop Quiz: Overfitting, Underfitting, or Too Few Examples
96 per cent on training data, 94 per cent on validation, and 71 per cent for one group of customers. Three of the four repairs on offer make it worse, and the number that tells you which is which is the gap between the first two.
Read articleExam Room · AI Practitioner
Pop Quiz: Interpretable or Explainable
Every declined application has to arrive with the reasons it was declined. Four candidate models, and the fluent one cannot tell you what it actually did.
Read articleExam Room · AI Practitioner
Pop Quiz: Model Card, Service Card, or Artifact
An auditor asks three things in one meeting: what your own model is for, what fairness considerations AWS documented for the managed service you call, and proof that AWS holds SOC 2. Three different artefacts, and the tempting answer merges two of them.
Read articleExam Room · AI Practitioner
Pop Quiz: The Thumbs-Down Button
A customer reads an answer from the support assistant, knows it is wrong about her own account, and has nowhere to say so. Four of the five fixes on the table are aimed at somebody who will never meet her.
Read articleExam Room · AI Practitioner
Pop Quiz: What a VPC Endpoint Actually Does
The team was told to keep Bedrock traffic off the public internet, so they added an interface VPC endpoint. The security register now says the traffic is encrypted and access is restricted. One of those entries is wrong.
Read articleExam Room · AI Practitioner
Pop Quiz: Config, CloudTrail, Inspector, or Artifact
An assessor wants proof that a bucket has blocked public access every day for a year. Four governance services sound like they answer that, and the tempting one only records the moments somebody touched the setting.
Read articleExam Room · AI Practitioner
Pop Quiz: When a Model Says It Is Confident
The assistant ends every answer with a confidence figure, so the team plans to auto-approve anything at 95% or above and send the rest to a person. It reads like a sensible gate. It is measuring the wrong thing.
Read articleExam Room · AI Practitioner
Pop Quiz: Does Picking a Region Settle Residency
The assistant was deployed in Sydney because the rule says customer data stays in Australia. Legal wants that confirmed in writing. Three things can still move the data or outlive it, and none of them show up in the Region setting.
Read articleExam Room · AI Practitioner
Pop Quiz: Lineage, Catalogue, or Model Card
One audit email, four questions, and every one of them starts with where did this come from. Four different artefacts answer them, and the tempting move is to reach for the Model Card four times because it sounds the most official.
Read articleExam Room · AI Practitioner
Pop Quiz: Access Control or Integrity
Three requirements land on one corpus in the same meeting: analysts must not see another country's customers, source documents must survive seven years unaltered, and thin batches must not load. Three different properties, three different services, and the first one is the one people misfile.
Read articleExam Room · AI Practitioner
Pop Quiz: Which Log Records What Was Said
A customer quotes back an answer the assistant supposedly gave them, and the team is sure it never said that. Four logs are running and only one holds the words.
Read article