Task 1.1: Describe core AI concepts and define terminology.

• ATraining, because the model updates its parameters every time it scores a new payment

  Incorrect

  Training is the phase in which the algorithm learns patterns from historical examples and adjusts its parameters. Scoring a payment does not change the model; it uses the parameters it already has.

• BData preprocessing, because each payment must be cleaned before it is scored

  Incorrect

  Preprocessing cleans and transforms raw data, mainly to prepare it for training. It is not the step in which the model produces a fraud score.

• CInference, which applies the trained model to new payments it has never seen

  Correct

  Inference is the process of a trained model generating an output, here a fraud score, from a new input. It is the recurring production cost of a model, separate from the one-time training effort.

• DModel evaluation, because each score is tested against validation data

  Incorrect

  Evaluation measures how well a model generalizes by testing it on a separate validation dataset before or after release. Scoring live payments is not evaluation; the true outcome is not known yet.

Training is where a model learns from historical data; inference is where the trained model is used to produce predictions on new data. A fraud model that scores each live payment is performing inference, which is why inference volume, not training, drives most of the ongoing cost of a model in production.

• AAs machine learning, because the product makes routing decisions on its own

  Incorrect

  Making a decision automatically is not what defines machine learning. Machine learning derives its decision logic from data; this product's logic was written by people and changes only when people edit it.

• BAs generative AI, because the product produces a different routing result for each claim

  Incorrect

  Generative AI creates new content such as text or images. Choosing one of a fixed set of teams from fixed conditions creates nothing new.

• CAs deep learning, because the 120 conditions work together like layers of a network

  Incorrect

  Deep learning trains neural networks on data. A list of hand-written conditions is not a neural network and is not trained.

• DAs rule-based automation, which is neither machine learning nor generative AI

  Correct

  Machine learning learns its rules from examples instead of being explicitly programmed, and generative AI creates new content. Fixed, hand-written conditions that never learn and never generate anything are traditional rule-based software.

Machine learning learns patterns from data without explicit instructions, and generative AI creates new content. A product whose behavior comes entirely from hand-written conditions is traditional automation, whatever the sales deck calls it. Knowing the difference lets a buyer compare like with like and ask the right questions about data, monitoring and cost.

• ARecorded phone calls between dispatchers and drivers about delays

  Correct

  Audio recordings have no predefined schema of rows and columns. They are unstructured data and must be transcribed or otherwise processed before their content can be analyzed.

• BThe fleet telemetry table of GPS position, speed and fuel level per truck

  Incorrect

  Telemetry stored as rows with fixed fields is structured data. AWS lists IoT sensor data and fleet telemetry among common structured sources.

• CThe CRM table of accounts with ID, region and value

  Incorrect

  A CRM account table has defined attributes in rows and columns, which makes it structured data that SQL and BI tools can query directly.

• DThe point-of-sale ledger of parcel shipments with date, weight and price

  Incorrect

  A ledger with a predefined schema of dates, weights and prices is structured data, readily used for reporting and traditional machine learning.

• EScanned PDF copies of signed customer contracts and amendments

  Correct

  Free-form documents and images of documents do not follow a schema. Contracts in this form are unstructured data.

Structured data follows a predefined schema, typically tables of rows and columns, and is easy for software to query. Audio, images and free-form documents are unstructured: they hold much of the value generative AI can use, but they need extra processing to become usable and are harder to govern.

• ARaise the health check target to 99% for every field in the database before retraining

  Incorrect

  A defaulted "retained" value is complete and valid, so a higher generic score would still pass the wrong outcomes. AWS advises tying data quality metrics to how the data will be used rather than to generic data health.

• BSwitch to a more advanced algorithm that is designed to tolerate noisy records

  Incorrect

  A model learns the relationships in its training data, and noise in that data degrades accuracy. When many canceled accounts are recorded as retained, any algorithm learns from the same wrong outcomes.

• CSet quality checks for this use case, such as outcome accuracy, and pass them before training

  Correct

  Data quality means different things for different uses. Working back from what the model must achieve, the team sets metrics and minimum thresholds, here the accuracy of the outcome field, and treats them as a gate before the data is used.

• DFix the billing default for new accounts, then retrain right away on the existing records

  Incorrect

  Fixing the source protects future records, but retraining now still teaches the model from the historical outcomes that are wrong, so its predictions stay unreliable.

A model is only as good as the data it learns from, and a generic health score can hide the defect that matters most. Quality requirements should come from the use case: for a model that predicts cancellations, the recorded outcome must be right. Checking those requirements before training stops a known data problem from becoming a model problem.

• AISO/IEC 23053, a global framework that gives organizations a shared vocabulary for AI systems

  Incorrect

  ISO/IEC 23053 is one of the global frameworks that give teams a unified AI vocabulary, so that people mean the same things by the same terms. A shared vocabulary is not the managed AI process with independent certification that the questionnaires ask for.

• BISO/IEC 42001, the certifiable standard for an AI management system

  Correct

  ISO/IEC 42001 specifies requirements for establishing, implementing, maintaining and continually improving an Artificial Intelligence Management System (AIMS), and accredited bodies certify organizations against it.

• CThe Responsible AI Lens, completed as a self-review

  Incorrect

  The Responsible AI Lens is advisory best-practice guidance. AWS states that it should not be used as a compliance or assurance checklist, and a self-review is not independent evidence.

• DThe ISO/IEC 42001 certification that AWS holds for its own AI services

  Incorrect

  AWS's certificate covers AWS's own AI management process for the services in its scope. Customers are not certified by association; the company needs its own certification.

Customers asking for independent evidence of a managed, improving AI process are describing an AI management system. ISO/IEC 42001 is the international standard for that system, and certification by an accredited body is the evidence they can rely on. Building on AWS can make the company's certification easier, but it does not transfer AWS's certificate.

• AThe model only knows what was in its training data

  Correct

  Large language models are trained on large volumes of data and generate answers from what they learned; information outside their training data, such as a new internal price list, is unknown to them unless it is supplied.

• BThe model is broken and must be replaced with a newer version

  Incorrect

  Nothing is broken; any model lacks information it was never given, so a newer model would have the same gap.

• CIt hides current prices on purpose

  Incorrect

  Models do not hold back known facts for security; the price list was simply never part of what the model learned.

• DThe model retrieves prices from the internet but read an old page

  Incorrect

  A plain model call does not search the internet; it generates text from its training.

Knowing where a model's knowledge comes from sets realistic expectations. For current or private information, the system must be given that information at question time, for example through retrieval.

• AThe model will remove past bias automatically, because algorithms are neutral

  Incorrect

  An algorithm learns from the data it is given; it does not know which past patterns were unfair.

• BThe model may repeat the past pattern, because it learns from history

  Correct

  A model learns the patterns in its training data. AWS guidance asks whether training data might inappropriately represent groups, because such data can carry unwanted bias into the system's decisions.

• CThe model will ignore history and judge each applicant from scratch

  Incorrect

  A trained model's behavior comes from its history; it has no other basis for its predictions.

• DBias is impossible because the data contains no names

  Incorrect

  Other fields can act as proxies for demographic attributes, so removing names does not remove bias.

Training on historical data means learning from historical decisions, good and bad. Leaders should expect past patterns to reappear unless the data and the system are examined and adjusted.

• ADocument extraction that turns each note into structured fields

  Correct

  Intelligent document processing extracts data from documents and enters it in the format the target system needs, turning unstructured scans into the structured fields a SQL database can store and query.

• BA generative model that writes a summary of each note

  Incorrect

  Generative AI creates new content such as text; a free-text summary is still unstructured and cannot be loaded as exact field values.

• CA forecasting model trained on past delivery volumes

  Incorrect

  Forecasting predicts future trends from historical data; it does not read the fields on individual notes.

• DSQL queries run directly against the scanned images

  Incorrect

  SQL works on structured data with a predefined schema; scanned images are unstructured, so their fields must be extracted before they can be queried.

Whether data is structured or unstructured decides which tools can use it. Scanned notes are unstructured; document extraction converts them into the structured fields that existing reports rely on, which a generated summary or a forecast cannot do.

• AA recommendation engine trained on visitors' past behavior

  Correct

  Machine learning finds patterns in large amounts of historical data without explicit instructions, and retailers use it to recommend products based on previous purchases, browsing history and search patterns.

• BA bestseller list chosen by sales rules, shown to every visitor

  Incorrect

  Hand-set rules give every visitor the same list; the bookstore needs predictions for each visitor, learned from history rather than written as rules.

• CA sales-trend dashboard that marketing reviews every month

  Incorrect

  A dashboard describes past results for people to read; it makes no prediction for each visitor, while machine learning uses historical data to predict for every visitor automatically.

• DAnomaly detection on visitors' purchase histories

  Incorrect

  Anomaly detection finds unusual patterns in data; it does not predict which titles each visitor is likely to buy.

A model trained on historical behavior can make a prediction for every visitor at the moment the page loads, which no set of hand-written rules can match at this scale. Recognizing this as a recommendation problem sets expectations for the data it needs.

• AThe reports only need SQL counts of words such as "dirty" or "rude"

  Incorrect

  Counting listed words is a basic rule-based approach that misses context such as negation and sarcasm, which AWS notes machines struggle with; meaning in free text needs language processing.

• BThe average star rating already stored answers the question

  Incorrect

  A star rating is structured data that gives one overall score; what guests say about cleanliness, staff or noise is in the review text, which structured reports cannot capture.

• CSentiment is in unstructured text, which needs language processing

  Correct

  Structured databases track quantities such as prices but cannot capture sentiment, which comes in unstructured text; analyzing it requires techniques such as natural language processing and text mining.

• DStoring the reviews as a text column lets the dashboards chart them

  Incorrect

  Storing text in a table does not interpret it; unstructured data lacks the order that conventional data-mining techniques need, so it must be analyzed with language processing first.

Much of what customers think is captured in unstructured text. Natural language processing turns reviews and comments into insights about specific topics that structured reports and simple word counts cannot provide.