Describe core AI concepts and terminology
AI, machine learning, and generative AI as related categories
Two AI systems arrive on a manager's desk: one labels each card payment as genuine or fraudulent, the other drafts replies to customer emails. Both are artificial intelligence (AI), yet only the second generates anything, and naming that difference correctly is the first vocabulary skill a business leader needs. This topic follows Task 1.1 of the AWS AIB-C01 exam guide, 'Describe core AI concepts and define terminology'[1]: what AI, machine learning (ML), and generative AI are, how a model learns from data and is then used, language, vision, speech, and recommendation capabilities, the AWS offerings Amazon Bedrock, Amazon SageMaker AI, and Amazon Quick, data forms and document processing, why data quality and honest evaluation decide outcomes, and which frameworks give an organization one shared vocabulary. The same guide lists two neighboring tasks separately[1], 'Task 1.2: Identify and select appropriate AI solution types' and 'Task 1.3: Apply GenAI concepts and techniques'; see Identify and select appropriate AI solutions and Apply generative AI concepts and techniques.
Related categories, from broad to specialized
The terms below are related categories, not synonyms: their relationships are not all strict containment. Of the terms below, machine learning, deep learning, foundation models, and large language models nest in that order. Generative AI works by using machine learning models[2], so it sits inside machine learning, and it overlaps deep learning, with one part inside and one part outside. Inside deep learning are its modern forms: modern generative AI emerged in the late 2010s with advancements in deep learning[2], and the foundation models below, which AWS describes as a form of generative AI[3], are large deep learning neural networks[3]. Outside deep learning are older statistical generative models, which predate it: primitive generative models have been used for decades in statistics[2], whereas neural networks and deep learning were recent precursors for modern generative AI[2]. Large language models are one class of foundation models[2], so they sit inside generative AI too, as the figure below shows; AWS places foundation models as a group under generative AI, though a single model can convert input into embeddings[4] or be intended for understanding languages[3] instead of creating new content, as the foundation-model and large-language-model entries below explain.
- Artificial intelligence (AI) is the broad category of systems that perform tasks such as perception, reasoning, and prediction. AWS describes AI as the broader concept of making machines more human-like[5], from smart assistants and chatbots to robotic vacuum cleaners and self-driving cars.
- Machine learning (ML) is one branch of AI: it learns from data without explicit programming[5]. Machine learning is AI, but not all AI is machine learning.
- Deep learning is a specialized form of machine learning that uses artificial neural networks[5], layers of connected software nodes (multilayer neural networks), for complex tasks such as image and speech recognition.
- Generative AI is a type of AI that can create new content and ideas[2], including conversations, stories, images, videos, and music. Deep learning laid the foundation for its modern forms.
- Foundation models (FMs) are large deep learning neural networks trained on a broad spectrum of generalized and unlabeled data[3] (data with no answers attached) and able to perform a wide variety of general tasks; AWS describes them as a form of generative AI, though they can also convert input into embeddings[4], numeric representations rather than new content.
- Large language models (LLMs) are one class of foundation models[2], focused on language tasks such as summarization, text generation, classification, conversation, and information extraction. Not every LLM generates new text: AWS notes that Cohere has two LLMs: one is a generation model, and the other is a representation model intended for understanding languages[3].
Generation is not the test for AI
These relationships settle a common misreading at once: a system does not have to generate content to count as AI. Classifying transactions or recommending products are prediction and recognition tasks, and ML models perform specific analysis tasks like classifying documents, labeling images, or predicting the maintenance schedule of factory equipment[5] without producing anything new. The overlap also runs the other way: a generative model can classify when asked to, because classification is among the language tasks LLMs handle[2]. Category membership describes how a system gets its capability, not the single task it happens to perform.
How a generative LLM produces language
A generative LLM, the kind that writes text, works in tokens, the units of text (whole words or word parts) that it reads and writes. During training it adjusts its parameter values until it correctly predicts the next token from the previous sequence of input tokens[6]; parameters are the internal numbers a model learns, explained in the next section. When it generates a response, it predicts one token after another from the prompt (the instruction and input you give it) and the tokens it has already produced. Foundation model and LLM are not interchangeable names: the foundation-model category also includes text-to-image models such as Stable Diffusion[3], so every LLM is a foundation model but not every foundation model is an LLM.
Why foundation models change the economics
Before foundation models, each new text task usually meant training a narrow model for it. A foundation model is reusable instead: it is faster and cheaper to use pretrained FMs to develop new ML applications rather than train unique ML models from the ground up[3], and building a foundation model from scratch is expensive and training may take months[3]. Broad pretraining has limits, though. Broad pretraining does not establish knowledge of your private contracts or customer records, and AWS notes that answers on certain subject matter may be unreliable[3]. Evaluate it for your specific use case[7] and adapt it where the evaluation shows a gap.
Broad capability is not artificial general intelligence
Artificial general intelligence (AGI) is a field of theoretical AI research that attempts to create software with human-like intelligence and the ability to self-teach, performing tasks it was not necessarily trained or developed for[8]. The AI systems businesses deploy are task-bounded, which AWS calls narrow or weak AI: even emerging generative AI applications with better memory retention are considered weak AI because they cannot be repurposed for other domains[8]. Reusing one foundation model for several business tasks, or letting an agent (an AI system that pursues a goal by taking actions) complete a workflow on its own, is useful task-bounded capability and does not by itself demonstrate AGI.
The takeaway: distinguish these categories by how a system gets its capability, and do not let generation, breadth, or autonomy stand in for a definition.
From learning algorithm to trained model to prediction
Algorithm, model, and inference are distinct terms, and the AWS AIB-C01 exam guide names all three among the fundamental AI concepts to recognize and explain in business contexts[1]. Keeping them apart tells you whether a proposal changes the learning method, the learned result, or only how the result is used.
Algorithm versus trained model
A learning algorithm is the procedure that finds patterns; a trained model is what the procedure produces. AWS illustrates the difference with a small case: given the input and output pairs (2,10), (5,19), and (9,31), the algorithm computes the relationship o = 3 × i + 4, and given the input 7 it predicts 25[5]. The procedure that searched for the relationship is the algorithm; the resulting formula is the model, and the numbers 3 and 4 are its learned parameters. In a real system the model encapsulates what it learned in its parameters[5], which training adjusts to shrink the gap between predictions and the known outcomes. Run the same algorithm on a different company's data and you can get a different model. So when a team proposes a training change, ask which of two things it is: changing the learning method, or updating the learned model with new training evidence.
Training versus inference
Training learns the parameters from examples; inference uses the finished model. Inferencing is the process of an AI model generating educated predictions or outputs from new data[9], using the learned parameters it ended training with. Forecasting next month's demand and drafting a reply to a customer are both inference, even when the input contains details the model never saw before.
Ordinary inference does not change the model. AWS explains model drift in exactly those terms: when a model's weights (its learned parameters) remain fixed based on previous AI training data[9] and the world shifts, the outputs become less accurate, and retraining is the remedy. Storing every prediction request in an archive therefore teaches the model nothing; learning from newly collected examples needs an explicit training or update step.
One AWS statement can look like a contradiction here. AWS also writes that foundation models can continue to learn from data inputs or prompts during inference[3], and the same passage explains what that means: carefully curated prompts produce better outputs. A prompt steers that one response, while the learned parameters stay as they were.
Predictions are estimates, and generation can vary
A learned prediction is an estimate, not a promise. Machine learning can be made deterministic[5], so a model may return the same prediction every time for the same input, yet a repeatable prediction is still not a promise about the real-world outcome: a demand forecast is useful even though actual demand may differ from it. Generative models add a different kind of variation, in the output itself: for any given sequence, a model determines a probability distribution of options for the next token and samples from this distribution[10], so the same prompt can produce different wording on different requests. Inference parameters, which are settings and not the learned parameters (those stay unchanged), narrow or widen that variation: when running model inference, you can adjust inference parameters to influence the model response[10]. Temperature is one such setting; it affects the shape of the probability distribution for the predicted output[10], and a lower temperature leads to more deterministic responses, and a higher temperature leads to more random responses[10].
The figure below traces the two phases: historical examples run through the learning algorithm to produce the trained model, new examples update the model's learned parameters only through a training or update step, and each new input passes through the unchanged model at inference to give a prediction that is an estimate.
To sum up: the algorithm learns, the model holds what was learned, inference applies it without changing it, and predictions are estimates, not guarantees.
Learning approaches: matching the method to your evidence
For example, suppose a retailer has three questions, each answered by a different way of learning. To flag fraudulent card transactions, it can learn from past transactions already confirmed as fraud or genuine, which is AWS's own example of training a system to classify fraudulent and genuine transactions[5] in supervised learning. To find natural groups among customers that nobody has grouped yet, it needs unsupervised learning, whose algorithms train on unlabeled data[5]. To choose a sequence of offers that keeps customers loyal, it learns from reward feedback associated with its offers, the reward-and-punishment paradigm[11] of reinforcement learning. The evidence the retailer holds decides the approach, and the decision tree in this topic's overview walks the same choice for every approach in the table below.
| Approach | Evidence it needs | What it produces | Example |
|---|---|---|---|
| Supervised learning | Examples paired with known outcomes (labels) | A category (classification) or a number (regression) | Fraudulent or genuine; a selling price |
| Unsupervised learning | Data with no labels | Groups, associations, or unusual observations | Customer segments; bread bought with butter |
| Semi-supervised learning | A few labeled examples plus many unlabeled ones | A model trained on both | Long documents too costly to label in full |
| Self-supervised learning | Unlabeled data from which targets can be derived | Patterns learned without human labels | Foundation-model pretraining |
| Reinforcement learning | Actions and the rewards that follow them | A strategy that maximizes reward over time | Offer sequences aimed at long-term retention |
Supervised learning: known outcomes
In supervised learning the sample data specifies both the input and the output[5]; the known output for each example is its label. That is why the fraud case above fits: its past transactions already carry known outcomes. Supervised models predict one of two kinds of target. Classification predicts a category, either one of two possible classes (binary classification) or one of more than two outcomes (multiclass classification)[12]; regression predicts a numeric value[12], such as a selling price. Before choosing it, a business has to name the outcome it wants predicted and confirm that reliable historical examples of that outcome exist.
Unsupervised learning: patterns without answers
Unsupervised learning algorithms train on unlabeled data[5] and find structure on their own: grouping similar items (clustering), such as news articles into topics, or finding items that occur together (association), such as a customer who buys butter being most likely to also buy bread[5]. AWS states the limit plainly: it cannot give precise predictions[5]. The groups it finds are similarity groups that need business interpretation. A cluster of customers who look alike is not a set of verified churn labels, and a segmentation run cannot stand in for a model trained to forecast an exact revenue figure.
Anomaly detection: a signal, not a verdict
Anomaly detection is examining specific data points and detecting rare occurrences that seem suspicious because they are different from the established pattern of behaviors[13]. It helps when you lack enough labeled examples of the failures you care about. For fraud with little labeled history, AWS describes an unsupervised anomaly detection model as one option, with the caveat that not all fraud events are anomalies, and not all anomalies are indeed fraud[14]. An unusual transaction or machine reading is therefore a reason to investigate, not proof of fraud or of a specific fault. A supervised classifier answers a different question: it recognizes the categories present in its labeled examples, while anomaly detection flags departures from normal.
Semi-supervised and self-supervised learning: when labels are scarce
Scarce labels do not stop learning. Semi-supervised learning relies on a small amount of labeled data and a large amount of unlabeled data[5], which suits material such as long documents that would be too time-consuming for humans to read and label[5]. Self-supervised learning goes further and derives the targets from the data itself: foundation models use self-supervised learning to create labels from input data[3], for example by learning to predict the next word in a text. Neither approach requires a person to supply the correct answer for every example, and neither removes the need to evaluate the result against the actual business task.
Reinforcement learning: rewards over time
Reinforcement learning (RL) learns from actions and their consequences instead of from paired correct answers. RL algorithms use a reward-and-punishment paradigm, are capable of delayed gratification, and may find that the best overall strategy requires short-term sacrifices[11]. That strength is also the business risk, because the system pursues whatever the reward measures: as the data scientist designs the rewards, they can influence the results[5]. Reward only immediate clicks and the system learns to chase clicks, even when the real goal is long-term retention.
The takeaway: start from the evidence. Known outcomes point to supervised learning, no labels to unsupervised discovery, unlabeled data from which targets can be derived (as in foundation-model pretraining) to self-supervised learning, a few labels plus many unlabeled examples to semi-supervised learning, and actions with delayed results to reinforcement learning.
Language, vision, speech, and recommendation capabilities
Suppose, for example, a retailer's service team receives written complaints, damaged-parcel photos, and recorded calls, and wants to show shoppers relevant products; the AWS AIB-C01 exam guide expects a business professional to recognize common categories of AI tools and their business applications[15], such as recommendation engines and computer vision. Business information arrives in different forms (written text, images, spoken audio, records of customer behavior), and AI capability families specialize in tasks and input types that can overlap. Each job matches one family below; the table is a common set, not an exhaustive list.
| Capability | What it works on | What it produces | Business example |
|---|---|---|---|
| Natural language processing (NLP) | Written or transcribed human language | Meaning: intent, sentiment, translation | Sorting support emails by intent |
| Computer vision | Images and video | Information found in the visual input | Spotting product defects before shipping |
| Speech recognition (speech to text) | Spoken audio | Text | Transcribing a support call |
| Recommendation engine | Interaction history plus item and user information | A selected or ranked set of existing items | A Recommended for you product list |
- Natural language processing (NLP) is technology that allows computers to interpret, manipulate, and comprehend human language[16], so organizations can classify, sort, filter, and understand the intent or sentiment hidden in language data[16]. Sentiment analysis interprets the emotion conveyed by textual data[16], and machine translation converts text or speech from one language to another[16].
- Computer vision enables computers to derive meaningful information from digital images, videos, and other visual inputs[17], with uses such as identifying product defects before products leave the factory[17].
- Speech recognition (also called speech to text) enables the recognition and translation of spoken language into text[18].
Transcription and sentiment are separate tasks
Speech recognition produces words, not their meaning. To learn how a caller felt, transcribe the call first and then run sentiment analysis (an NLP task) on the transcript; contact centers combine the two to derive actionable insights from call analytics[16]. A transcript alone does not measure customer sentiment.
Recommendation selects items; writing is a separate task
A recommendation engine chooses or ranks existing items for a person, using evidence such as past interactions, preferences, and item information. Amazon Personalize is one AWS example among the recommendation tools available: it generates recommendations primarily based on item interaction data, such as users clicking items, and in some cases also uses data from items and users such as genre or price[19]. Writing a personalized message about the chosen item is a different capability. Personalize's own email use case splits the work the same way: Personalize generates batch recommendations for all users on an email list, and then an AWS service or third-party service sends the emails[19]. A customer experience can combine selection with generated wording, but generating the wording does not perform the selection.
The takeaway: match the capability to the form of the information and to the task (understand, see, transcribe, or select), and expect real solutions to chain more than one.
Amazon Bedrock, Amazon SageMaker AI, and Amazon Quick
Suppose a company's developers want a foundation-model support assistant, its data scientists their own churn model, and its analysts AI answers from company data; the AIB-C01 exam guide describes Bedrock as a generative AI platform, SageMaker AI as custom ML, and Quick as AI-powered business assistants[15]. These three AWS offerings are listed in the exam's scope, and they serve different people at different layers. Overlapping AI capabilities do not make them interchangeable; the useful question is who is building or using what.
| Offering | Who typically uses it | What it provides |
|---|---|---|
| Amazon Bedrock | Teams building generative AI applications and agents | Managed access to a choice of foundation models plus the capabilities to build generative AI applications |
| Amazon SageMaker AI | Data scientists and developers | A managed environment to build, train, and deploy their own ML models |
| Amazon Quick | Business users | Ready-to-use AI assistance grounded in connected company data, plus analytics and workflow automation |
- Amazon Bedrock is a fully managed service that offers a choice of industry leading foundation models (FMs) along with a broad set of capabilities that you need to build generative AI applications[20]. Builders use a single API for inference regardless of the models they choose[20] and can create agents that execute complex business tasks by dynamically calling company systems and APIs[20].
- Amazon SageMaker AI is a fully managed machine learning service with which data scientists and developers build, train, and deploy ML models[21]. Technical teams use it for the custom-ML lifecycle described in the training and evaluation sections below. On December 03, 2024, Amazon SageMaker was renamed to Amazon SageMaker AI[21].
- Amazon Quick is an AI-powered service for automating tasks, analyzing data, building web applications, and conducting research[22], used through natural language chat. Its features include Amazon Quick Index, which connects your documents and data sources so AI responses are grounded in your information[22], Amazon Quick Sight for dashboards and business intelligence, and Amazon Quick Flows and Amazon Quick Automate for workflow automation. Amazon Quick evolved from Amazon QuickSight[22], which continues as the Quick Sight feature.
The takeaway: builders assemble generative AI applications on Bedrock, technical teams develop their own models on SageMaker AI, and business users work with Quick.
Structured, semi-structured, and unstructured data
Whether data is structured depends on its schema, not on whether it holds numbers or words. A schema is the predefined design of a dataset: the field names, data types, and constraints that every new record follows[23]. The AWS AIB-C01 exam guide asks candidates to distinguish between structured and unstructured data and explain the relevance of data types for AI[1]: which of the three data forms a source takes, and which type of information each input carries (its modality). Evaluating data readiness across sources, including the effects of data silos on AI initiatives[24], is a separate task in the guide, 'Task 4.2: Establish data and infrastructure foundations for AI'; data silos are digital systems where data is difficult to share or access by other services[25]. See Establish data and infrastructure foundations for AI.
| Data form | Defining feature | Examples | What AI must do with it |
|---|---|---|---|
| Structured | A predefined schema of named fields | Transaction rows with customer, date, and amount | Analyze records |
| Semi-structured | Self-describing tags and metadata without a rigid table | A media library with dates and topic tags | Filter by tags; analyze the content separately |
| Unstructured | No predefined schema | Narrative documents, images, audio, video, scanned pages | Understand the content |
Structured versus unstructured
Structured data is typically tabular with rows and columns that clearly define data attributes[23]. Unstructured data exists without any predefined schema[23]: plain text documents, video and audio files, images, and scanned documents. The difference is the schema, not the subject matter: a structured database can easily track quantitative information, such as product prices and order volume, and can also store customer feedback for sentiment analysis[26].
Text is not automatically unstructured. A short product-category or customer-ID field is part of a schema and is structured, while a digital contract with no fixed fields is still unstructured. The distinction tells you which job you are buying: record analysis for structured data, content understanding for unstructured data.
Semi-structured data: metadata is not meaning
Semi-structured data lacks a specific relational or tabular data model but includes self-describing elements such as tags, metadata, and nesting[23]. A recording library tagged with dates and topics can be filtered by those tags, but the tags do not tell you what anyone said in a recording; that still needs analysis of the content. Searching document filenames is likewise not the same as reading the obligations inside the contracts they name.
Modality: the type of information
A modality is a type of information, such as text, images, audio, or video, and a multimodal capability handles more than one. AWS describes multi-modal agents as AI systems that can understand and analyze data in multiple modalities[27]. A text-only capability receives only text, so it cannot inspect the damage shown in a claim photo or read the values in a scanned chart just because a user mentions the image. For a mixed-information task, list which modalities the decision depends on and check that any conversion, such as summarizing an image as text, keeps the evidence the decision needs.
The takeaway: identify the data form and every modality first, then pick a capability that reads that form and does the job you need (search, analyze records, or understand content).
Optical character recognition versus intelligent document processing
Reading a document's text and capturing its business fields are different jobs; the AWS AIB-C01 exam guide lists document extraction[15] among the common categories of AI tools. Optical character recognition (OCR) converts an image of text into a machine-readable text format[28], which makes a scanned page searchable. Intelligent document processing (IDP) goes further: it automates the process of manual data entry from paper-based documents or document images to digital form[29], for example extracting invoice data and entering it in the required format in the accounting system[29], and it helps with data extraction, validation, and classification[29]. OCR is one of the technologies IDP builds on[29]. A scanned invoice made searchable through OCR shows that text recognition occurred, not that the invoice number, total, and payment terms were captured correctly in the business record.
The takeaway: OCR makes a document's text searchable; when the job is to extract fields, ask for field-level extraction and validation as well.
Data quality dimensions and why they decide outcomes
A model learns whatever its training records say, mistakes included. AWS states that a machine learning model's performance depends on the data quality used for training, and that missing values, inconsistent data entries, and noise can significantly degrade model accuracy[5]; using poor training datasets can result in erroneous predictions in deployed models[30]. This section covers the dimensions used to describe data quality and what to do when one fails. The AWS AIB-C01 exam guide lists data strategies, data ownership, and data sharing frameworks[24] under 'Task 4.2: Establish data and infrastructure foundations for AI'; see Establish data and infrastructure foundations for AI.
The dimensions
AWS guidance describes overall data quality through facets such as consistency, completeness, correctness, accuracy, validity, and timeliness[31] and treats two of them together as the degree of data accuracy relative to real-world entities (correctness and accuracy)[31], so the Accuracy row below covers correctness too. Data-quality rules also check uniqueness, the share of unique values, which occur exactly once[32]. Each dimension catches a different defect, so a dataset can pass one and fail another.
| Dimension | The question it asks | A defect it catches |
|---|---|---|
| Accuracy | Do recorded values reflect reality? | A customer address that is simply wrong |
| Completeness | Is the required information present? | Missing income on loan applications |
| Timeliness | Is the data current enough for this decision? | Last year's prices feeding today's forecast |
| Consistency | Does the same business fact mean and look the same across records and sources? | Two stores recording returns differently |
| Validity | Do values follow the required format or range? | A date written in the wrong format |
| Uniqueness | Is each entity or event recorded only once? | The same customer entered twice |
Passing one dimension's check does not establish another's, as these pairs show:
- Accuracy versus timeliness. An old record can be perfectly accurate for its date and still unsuitable for a decision that needs current information.
- Validity versus accuracy. Validity rules restrict values within a specific range, format, or predefined set[33], but a value can pass a format check and still be wrong: a correctly formatted phone number can belong to someone else.
- Uniqueness versus completeness. Two copies of the same customer can both be accurate and have every required field filled in, so a completeness check passes them, yet together they inflate the customer count. Ask whether repeated records are genuinely separate observations before treating record volume as the size of the business population.
Fix the source before retraining
When predictions rest on inaccurate or missing records, correct the source data and reassess whether it suits the use before retraining. A more sophisticated model trained on contradictory customer records still faces that data-quality risk, and adding more examples from the same faulty process can enlarge the problem rather than supplying better evidence. Data cleansing means modifying and removing incorrect and incomplete data fields, identifying and removing duplicate information, and correcting formatting, missing values, and spelling errors[30]. When stores record the same business event differently, standardize how the data is defined, collected, and checked at the source; making the final AI reports look uniform changes the presentation, not the comparability of the underlying measurements.
Profiling finds issues; experts decide what they mean
Data profiling analyzes the statistical properties, distributions, and patterns within a dataset to identify anomalies, missing values, outliers, and other quality issues[34] before training. It needs people as well as tools: subject matter experts should interpret these findings to determine appropriate actions[34], and AWS lists failing to document data quality issues and their resolutions[34] as an anti-pattern. An unusually large transaction may be a typing error or a legitimate business event; deleting every large transaction before training a fraud model may discard the very cases the model needs to learn. Set quality expectations, decide each remediation with an expert rather than blindly accepting an automated suggestion, and record why.
The takeaway: name the failing dimension, fix it where the data is produced, and have business experts judge what profiling flags before anything is removed.
Representative data, label quality, and synthetic data
Volume is not coverage. This section covers three properties of training evidence that even a large dataset can lack: representation of the right people and situations, trustworthy labels, and synthetic examples that behave like the real thing.
Representative data
Training evidence has to look like the setting where the model will work. AWS guidance is to consider whether your datasets accurately reflect the real-world conditions where your system will be used, and to gather examples that represent your users[35]. Imbalance has a predictable effect: when some classes are much more frequent, the model can perform well on the majority class while failing to predict the minority class accurately[5]. A strong overall score can therefore hide weak performance for an underrepresented group, a rare outcome, or a missing business context. Collecting more of the already dominant customer segment adds volume without filling the gap; seek relevant examples of what is missing and record what you included and excluded from your datasets, so the limitations are known[35].
Label quality
A label tells a supervised learning system which answer to learn, so inconsistent or biased labels teach the wrong associations. For human labeling, AWS recommends training processes that teach consistent labeling, measuring how well different annotators agree with each other, checking for unwanted biases, and simple tests using examples with known correct answers[36], along with processes for fixing labels that do not meet your quality standards[36]. Machine-generated labels need the same scrutiny: check synthetic and model-generated labels against human judgment, and test whether they work well for underrepresented groups and edge cases where automated systems may fail[36]. Automation changes how labels are produced, not the need to prove they are reliable.
Synthetic data
Synthetic data is data generated by an algorithm rather than observed from real-world measurements[37], and it can supplement scarce examples. Its value depends on realism, not volume. AWS guidance is to verify that synthetic data keeps the statistical properties you need, does not exclude important edge cases or rare scenarios, and preserves key relationships in the data[38]; to search for bias in the synthetic data generation process[38]; to have domain experts evaluate whether synthetic examples capture the key phenomena and scenarios[38]; and to document when synthetic data should and should not be used as a substitute for real data[38]. Replacing a handful of rare real incidents with many generic generated ones can erase the unusual relationships the model most needs to recognize.
The takeaway: judge training evidence by coverage, label reliability, and realism, and document its limits; a bigger dataset with the same gaps is not better evidence.
Training, validation, and test sets
A model's training score says how well it learned the examples it was given, not how it will perform on the next customer. This section covers how teams separate the evidence so that an approval decision rests on examples the model has not already seen, and the ways that separation breaks.
Three sets, three jobs
AWS guidance is to split the data into at least three sets: training, validation, and testing[39].
- The training set teaches the model its parameters.
- The validation set supports development decisions: teams use the validation set to choose and tune the hyperparameters[39], the settings that control the learning process.
- The test set is held out until development is finished: when you have achieved a desired result, perform evaluation on the testing set[39].
A business approval should rely on the test result, because those examples were neither used to fit the model nor consulted again and again while refining it. Checking every development revision against the test set quietly turns it into a second validation set, and the independent assessment is lost. The same rule applies to prompt engineering: AWS recommends that you split a sample dataset into a development set for trying prompt changes and a hold-out test set to see if your prompt modifications generalize[40]. That separation tells a reusable improvement apart from a result tailored to familiar examples. The figure below traces the whole sequence, ending with the use-context check described at the end of this section.
Data leakage
Data leakage happens when your model gets data that it should not have access to, such as data samples that were used for training, or information that will not be available when the model is deployed in production[39]. Leakage makes results look better than reality, in two common ways:
- Information from the future. Historical data might be enriched with additional columns or values that were not present at some point in time[39]. A claim-risk model scored when a claim is first reported must not learn from the settlement amount, which the business only knows once the claim closes.
- Duplicates across sets. When a dataset contains too many duplicate samples, different subsets might have samples in common, and overfitting might be mistaken for generalization[39].
When a result looks impressive, ask two questions: was every input available at the real decision point, and are the evaluation examples genuinely separate from the training examples?
Overfitting and underfitting
The gap between training and test results is the diagnostic. Overfit models perform well on the training set but not on the test set; underfit models produce inaccurate results on both the training data and the test set[41]. An overfit model memorized training-specific detail or noise; an underfit model never learned the relevant relationship. AWS identifies overfitting by a high error rate in the testing data and a large gap between accuracy in the training data and test data[41]. In AWS's terms, underfit models exhibit high bias and overfit models exhibit high variance[41]: bias here is the error of a model that never learned the relationship, and variance is performance that holds on the training set but not on new data. Neither is the unfair treatment of groups discussed under representative data and the fairness dimension. AWS notes that for a sufficiently complex model more model training leads to less bias, but variance can increase[41]. Longer training or a higher training score is therefore not automatically better. The two patterns need different technical fixes, and the business question is always whether the model generalizes to new examples.
Check the use context
An aggregate test score covers the population the test set came from, not every place the model will be used. AWS gives two examples. A credit model trained on data for a whole country should, when the bank expands into one city, be evaluated on test data for that city slice as well[39]; a slice is a subset of evaluation data for a particular group, place, or segment. And a grocery demand model trained on 2019 data and tested on early 2020 data would probably score well on the testing set but generate poor results in production because consumer patterns changed significantly during the COVID-19 pandemic[39], which is why AWS suggests an additional recent test set as a safeguard for model approval. After a major business shift, ask for evaluation on suitably recent examples.
The takeaway: approve on held-out, leak-free evidence that covers the groups, places, and recent conditions where the model will run; a single impressive aggregate score is not that evidence.
ISO/IEC 22989, 23053, and 42001
Three ISO/IEC standards come up in AI conversations, and each answers a different question. This section covers what each one is for, so you can tell a vocabulary standard from a system framework and from a management-system standard. The AWS AIB-C01 exam guide lists governance work under 'Task 3.2: Establish AI governance structures and ensure regulatory compliance'[42]; see Establish AI Governance and Regulatory Compliance.
| Standard | Kind of document | The question it answers |
|---|---|---|
| ISO/IEC 22989 | Concepts and terminology | What do our AI terms mean? |
| ISO/IEC 23053 | Framework for AI systems that use ML | What are the parts of an ML-based system, and how do they relate? |
| ISO/IEC 42001 | Management-system requirements | How does the organization manage its AI? |
- ISO/IEC 22989 establishes terminology for AI and describes concepts in the field of AI[43], and it can be used in support of communications among diverse, interested parties or stakeholders[43]. When teams argue about what model, training, or inference mean, this is the dedicated reference.
- ISO/IEC 23053 provides a conceptual framework and shared terminology for describing AI systems that use machine learning, and defines the components and functions of ML-based AI systems[44]. Deep learning, as a subset of machine learning, is explicitly addressed within the standard's framework[44]. Use it to explain how the elements of a system relate.
- ISO/IEC 42001 specifies requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS)[45]. An AIMS is a set of interrelated or interacting elements of an organization intended to establish policies and objectives, as well as processes to achieve those objectives, in relation to the responsible development, provision or use of AI systems[45]. It applies to organizations of any size involved in developing, providing, or using AI-based products or services[45], not only to organizations that build models.
A management system is not a model check
ISO itself draws the contrast: 22989 establishes terminology and 23053 a framework for describing ML systems, while ISO/IEC 42001 is a management system standard that puts in place policies and procedures for the sound governance of an organization in relation to AI, using the Plan-Do-Check-Act methodology, rather than looking at the details of specific AI applications[45]. Plan-Do-Check-Act is a repeating cycle of four steps (plan, do, check, act) that serves the standard's requirement for establishing, implementing, maintaining and continually improving[45] the management system. Two contrasts follow from it. Reaching for 42001 to settle a terminology dispute picks a management-system standard over the dedicated vocabulary standard. And a model benchmark report is not an AI management system: it measures one application's outputs, while 42001 concerns the organization's coordinated policies, objectives, and improvement processes. An organization should consider both kinds of evidence in light of its AI role and use case.
The takeaway: 22989 for the words, 23053 for the parts of an ML system, 42001 for how the organization manages AI.
NIST AI RMF, AWS CAF for AI, and AWS responsible AI dimensions
Beyond the ISO standards, three more frameworks give business and technical teams a shared vocabulary: one for risk, one for adoption, and one for responsible AI. The AWS AIB-C01 exam guide lists Task 3.1, 'Apply responsible AI principles to business decisions'[42] and Task 4.1, 'Assess AI business readiness and maturity'[24] separately; see Apply Responsible AI Principles to Business Decisions and Assess AI business readiness and maturity.
NIST AI Risk Management Framework
The NIST AI Risk Management Framework (AI RMF) is intended to be voluntary, rights-preserving, non-sector-specific, and use-case agnostic[46], offered to organizations that design, develop, deploy, or use AI systems. Its Core is composed of four functions: Govern, Map, Measure, and Manage[46]:
- Govern establishes the governance that the other three functions rely on.
- Map establishes the context to frame risks related to an AI system[46].
- Measure employs quantitative, qualitative, or mixed-method tools[46] to assess those risks.
- Manage entails allocating risk resources to mapped and measured risks[46], which is where responses are prioritized.
Govern is not a one-time first step: governance is designed to be a cross-cutting function to inform and be infused throughout the other three functions[46], and framework users may apply these functions as best suits their needs[46], tailoring them to their application and context. Because the framework is voluntary guidance, adopting it is not regulatory permission to deploy; applicable legal requirements still apply on their own terms.
AWS Cloud Adoption Framework for AI
The AWS Cloud Adoption Framework for AI (AWS CAF for AI) treats AI adoption as an organizational capability challenge. Adoption depends on foundational capabilities in business, people, governance, platform, security, and operations[47], the framework's six perspectives. The word governance names three elements here, each defined in its own framework: NIST's cross-cutting Govern function, this CAF for AI perspective, and an AWS responsible AI dimension below; security is likewise a CAF perspective and part of the privacy and security dimension. A capability here is an organizational ability to use processes to deploy resources (such as people, technology, and other tangible or intangible assets) to achieve an outcome[48]. The perspectives connect business outcomes and workforce capability with oversight, enabling technology, protection, and reliable operation, which gives business and technical stakeholders one structure for discussing gaps. An assessment limited to the Platform perspective leaves the other five perspectives unassessed, so on its own it cannot establish overall AI readiness.
AWS responsible AI dimensions
AWS defines responsible AI using a core set of dimensions[7], each a distinct concern:
| Dimension | AWS definition |
|---|---|
| Fairness | Considering impacts on different groups of stakeholders |
| Explainability | Understanding and evaluating system outputs |
| Privacy and security | Appropriately obtaining, using, and protecting data and models |
| Safety | Preventing harmful system output and misuse |
| Controllability | Having mechanisms to monitor and steer AI system behavior |
| Veracity and robustness | Achieving correct system outputs, even with unexpected or adversarial inputs |
| Governance | Incorporating best practices into the AI supply chain, including providers and deployers |
| Transparency | Enabling stakeholders to make informed choices about their engagement with an AI system |
Distinguish transparency from explainability. Transparency is disclosure that lets people decide how to engage with a system; explainability is the ability to understand and evaluate its outputs. A transparency disclosure alone therefore does not establish that individual decisions are understandable. More generally, strong results on one dimension, such as veracity, do not by themselves establish adequate performance on the others, so no single quality score can stand in for all eight.
The takeaway: the NIST AI RMF organizes risk work, AWS CAF for AI organizes adoption capabilities, and the responsible AI dimensions name the separate concerns a trustworthy system has to meet.
Matching business situations to core AI concepts
The AWS AIB-C01 exam guide asks candidates to recognize and explain fundamental AI concepts in business contexts[1]. The table below gives illustrative business situations, the concept each one points to, and a neighboring concept or misreading that sounds right but does not apply, with the mechanism that rules it out.
| Business situation | Concept that applies | Neighboring misreading and why it fails |
|---|---|---|
| A system classifies or recommends but writes nothing | AI or ML without generation | Ruling it out as AI; generation is one kind of AI, not the definition |
| A new text task and no model yet | Reusing a pretrained foundation model | Commissioning a model from scratch; foundation models are pretrained for reuse |
| An archive of prediction requests keeps growing | Inference does not update the model | Calling the archive retraining; storing requests changes no parameters |
| Call recordings, and the goal is customer mood | Speech recognition, then sentiment analysis | Transcription alone; it outputs words, not their meaning |
| A numerical revenue forecast is needed | Supervised regression on known outcomes | Customer segmentation; it discovers groups, not a numeric target |
| An unusual transaction is flagged | Anomaly detection, a signal to investigate | Treating it as confirmed fraud; not all anomalies are fraud |
| Experts cannot label every document | Self-supervised or semi-supervised learning | Requiring full manual labels before any pretraining |
| Reward set on immediate clicks, goal is retention | Reinforcement learning reward design | Assuming the system pursues the real goal; it pursues what the reward measures |
| Customer ID and product-category fields | Structured data | Calling all text unstructured; schema-defined text fields are structured |
| Claim photos sent to a text-only model | Modality fit | Judging visual damage from text that never described it |
| Invoices are now searchable | OCR, not field extraction | Accepting searchability as proof that payment fields were captured |
| Duplicate customer records | Uniqueness | Relying on a completeness check; duplicates can have every field filled |
| Weak results for a minority segment | Representative data | Adding more majority-segment data; volume does not fill the gap |
| The test set is used on every revision | Independent testing | Calling the final score an independent assessment |
| Settlement data in an initial claim-risk model | Data leakage | Trusting the score; it uses information known only later |
| Strong national score, new city | Slice evaluation | Accepting the aggregate; local performance is unexamined |
| Teams dispute what AI terms mean | ISO/IEC 22989 | ISO/IEC 42001, a management-system standard |
| NIST AI RMF adopted before launch | Voluntary risk guidance | Treating adoption as regulatory permission to deploy |
| Only platform capacity reviewed for readiness | AWS CAF for AI perspectives | Treating one perspective as the whole readiness assessment |
| A model disclosure is published | Transparency | Treating it as explainability of individual decisions |
The takeaway: name the mechanism behind each situation (what the system learns from, what form the data takes, what the evidence proves, what the framework is for) before choosing between neighboring concepts.
AI vs machine learning vs generative AI
| Question | Artificial intelligence (AI) | Machine learning (ML) | Generative AI |
|---|---|---|---|
| What it is | The broad category of systems that perceive, reason, and predict | The branch of AI that learns patterns from data | AI that produces new content from learned patterns |
| Where it sits | The outer category | Inside AI; deep learning is a specialized form of ML | Inside ML; overlaps deep learning, and modern forms use deep learning |
| Typical output | A decision, a recognition, or an action | A prediction: a category, a number, or a group | New text, images, audio, or code |
| Business example | Smart assistants, robotic vacuums, self-driving cars | Classifying a transaction as fraudulent or genuine | Drafting a reply to a customer email |
| Must it generate content? | No | No | Yes, generation is what defines it |
Decision tree
Sharp facts the exam loves — give these one last read before exam day.
Cheat sheet
Sharp facts the exam loves — scan these before test day.
- AI, machine learning, and generative AI describe related but different capabilities
Artificial intelligence is the broad category of systems performing tasks such as perception, reasoning, and prediction; machine learning is an approach within AI that learns patterns from data. Deep learning is a form of ML using multilayer neural networks, and modern generative AI uses learned patterns to produce content such as text, images, or audio. A system can classify transactions or recommend products without generating new content, while a generative model can also perform tasks such as classification when prompted.
Trap Using content generation as the defining requirement for every AI application, which excludes predictive and recognition capabilities.
4 questions test this
- A manufacturer runs two AI review tracks. Systems that produce new text, images, or audio go through a generative AI track with extra content-safety review. Predictive and recognition systems go throu
- An insurance company wants to automate the first review of auto claims. It needs a capability that assigns each claim a likelihood of fraud from past claim outcomes, and it wants to avoid paying for f
- A bank's innovation board funds only projects it can categorize accurately for its AI portfolio report. A team proposes a system that scores each card transaction in real time and flags likely fraud f
- A retailer's leadership team wants a single sentence in its AI policy that defines which projects fall under AI governance. Draft wording says that AI means any system that generates new content. The
- An algorithm is a method, while a trained model contains learned relationships
A learning algorithm is the procedure used to discover patterns, whereas a trained model is the resulting representation of those patterns and learned parameter values. The same learning procedure can produce different models when trained on different business data. When discussing an AI proposal, distinguish changing the learning method from updating the learned model with new training evidence.
5 questions test this
- A bank's model governance policy requires full validation whenever the learning method behind a model changes. A lighter performance check applies when the same method is retrained on newer data. The
- A retailer is ending a contract with a consultancy that built its demand forecasting system. The retailer must keep producing forecasts on day one after handover without retraining anything. The consu
- An insurer's regulator can ask it to reproduce any automated claim decision from the past seven years. The insurer retrains its claims model every quarter with the same learning method, and each retra
- A payments company built a fraud model on transactions from its home country and now plans to launch in a second country with different shopping habits. A product manager says that because the learnin
- A hotel group's pricing model was trained four years ago, and its rate recommendations now miss current booking patterns. The data team still trusts the learning method and wants to keep costs low. A
- Training learns patterns, while inference applies a model to a new request
Training uses examples to learn model parameters; inference uses an already trained model to produce a prediction or response for an input. Forecasting next month's demand or generating a reply is inference, even when the input contains information the model has not previously encountered. Ordinary inference does not itself update the model's trained parameters; learning from newly collected examples requires an explicit training or updating process.
Trap Treating a growing archive of prediction requests as automatic retraining, which confuses collecting examples with updating model parameters.
5 questions test this
- A logistics company is planning a parcel delivery-time model. The project sponsor must assign two budgets: one for gathering five years of delivery records with actual arrival times, and one for syste
- A lender's credit model, trained on ten years of applications, receives an application from a profession that rarely appears in its training data. A loan officer insists the model must be retrained be
- A bank's chatbot answers customer questions with a deployed language model. Compliance staff are worried that customer messages could teach the model to repeat one customer's details to others. The ba
- A software company plans a generative AI assistant for its 40,000 customers. Finance must estimate which costs will rise as customers send more questions each month. The company will use an existing p
- An email security vendor's phishing filter was trained two years ago and has scored millions of messages since. Attackers now use new techniques, and more phishing emails slip through. The vendor's ac
- Foundation models provide broadly reusable capabilities rather than one narrowly trained task
A foundation model is pretrained on broad, diverse data and can be adapted or prompted for multiple downstream tasks. This allows a business to reuse an existing capability instead of funding training from scratch for every application. Broad pretraining does not establish that the model knows a company's private information or meets its specific quality requirements, so business-specific evaluation and, where necessary, adaptation remain important.
Trap Commissioning a new model from scratch for each text task, which overlooks the reusable capabilities of pretrained foundation models.
4 questions test this
- A hospital network plans several generative AI tools for patient communication, appointment summaries, and staff training content. The CIO worries that each tool would need its own costly model build.
- A retail company is evaluating a foundation model to answer customer questions about its return policies. In a demo, the model writes fluent general answers but cites return windows that differ from t
- A regional publisher wants AI to draft article summaries and headline options in its house style within six months. A product lead proposes training a new language model from scratch so the output ref
- A legal services firm wants AI to summarize contracts, draft client emails, and classify incoming requests. The IT director proposes training three separate models from scratch, one per task. The mana
- Language, vision, and speech capabilities address different forms of business information
Natural language processing interprets or generates human language, supporting sentiment analysis, intent recognition, and translation. Computer vision extracts information from images or video, supporting visual inspection and object recognition. Speech recognition converts spoken words into text, which can then be analyzed for meaning; transcribing a call and determining customer sentiment are different tasks that can be combined.
Trap Using speech transcription alone to measure customer sentiment, which produces words rather than an interpretation of their meaning.
4 questions test this
- A bank wants to monitor recorded advisor calls for signs of customer dissatisfaction and needs results in a compliance dashboard. The project sponsor wants to buy only what is necessary. One vendor se
- A quick-service restaurant chain wants drive-through customers to speak their orders to an automated assistant. The assistant must enter each requested item into the point-of-sale system and then read
- An online marketplace wants to flag product photos that show prohibited items and to translate seller descriptions into five languages. The trust and safety lead and the localization lead are sharing
- A contact center records every customer call. Leadership wants a weekly report showing how frustrated customers feel about a new billing change. The team has deployed speech recognition and now has ac
- A large language model generates language by predicting successive tokens
A large language model (LLM) is a model trained on extensive language data to learn patterns that support understanding and generating text. A generative LLM produces a response by predicting successive tokens from the supplied context and the tokens already generated. LLMs are language-focused members of the broader foundation-model category, which also includes models for non-language content such as images; foundation model and LLM are therefore not interchangeable names.
- A learned prediction is an estimate rather than a guaranteed outcome
A learned prediction is a probabilistic estimate based on patterns, not a promise that the predicted event will occur. A demand forecast can therefore be useful even though actual demand differs from its predicted value. Generative responses can also vary across repeated requests with the same prompt because generation can select among possible next tokens; identical wording is not guaranteed simply because the request is unchanged.
4 questions test this
- A telecom company's support chatbot uses a generative model to answer billing questions with each customer's account details. QA automation sends the same test prompt five times and fails the build wh
- A marketing team uses a generative AI model to draft product descriptions. When a manager submits the same prompt twice, the model returns two differently worded drafts. The manager reports a defect a
- An insurer uses a generative AI assistant to draft claim denial letters. Compliance discovers that letters for the same claim facts are worded differently each time they are generated, and regulators
- A beverage distributor's demand model forecast 12,000 cases for July, and actual sales were 11,200. A board member argues that the model is broken because its prediction was not exact and proposes can
- Broad task capability does not by itself establish artificial general intelligence
Artificial general intelligence (AGI) denotes the theoretical goal of generalized human-level cognitive ability, including learning and solving problems across domains beyond the tasks a system was developed for. Task-bounded AI operates within capability limits even when it performs many useful activities. Reusing a foundation model for several business tasks or letting an agent complete a workflow autonomously does not by itself demonstrate AGI.
- Recommendation engines select relevant items rather than merely personalize wording
A recommendation engine uses evidence such as customer interactions, preferences, and item information to select or rank relevant products or content. Purchase history and product descriptions can help determine which existing items to show a customer. Generating a personalized message about an item is a different capability: it creates wording rather than performing the selection or ranking, although a business experience can combine both.
3 questions test this
- A clothing retailer's marketing lead says the company already has recommendations because a generative model writes a personalized greeting with each customer's name in every email. Click-through rate
- An online bookstore wants each customer's home page to show the titles they are most likely to buy, chosen from its existing catalog. A product manager proposes using a generative model to write a per
- A travel company wants its app to show each user the five hotels they are most likely to book in a destination, and to add a short generated summary of each hotel's amenities. The product team debates
- Amazon Bedrock, Amazon SageMaker AI, and Amazon Quick serve different business capability needs
Amazon Bedrock provides managed access to foundation models and capabilities for creating generative AI applications and agents. Amazon SageMaker AI provides a managed environment for technical teams to develop, customize, train, deploy, and manage models across the custom-ML lifecycle. Amazon Quick is a ready-to-use business AI service combining assistance grounded in connected data, analytics, and workflow automation. These categories distinguish application-building capabilities, a model-development environment, and business-facing assistance; overlapping features do not make the offerings interchangeable.
4 questions test this
- A manufacturer has three AI needs. Its data science team must build and train a custom defect-detection model on its own images. Its product team wants a generative AI assistant in the customer portal
- A retailer's business analysts want to ask natural-language questions about sales dashboards and automate routine reporting tasks. They have no ML engineers and want a ready-to-use service grounded in
- An insurer's data science team uses Amazon SageMaker AI to train and deploy custom claims-fraud models. After adopting Amazon Quick for business analytics, the CFO proposes retiring SageMaker AI becau
- A bank's data science team needs to build a credit risk model from its own labeled loan data, tune it, deploy it, and manage it across the model lifecycle. The CIO asks whether Amazon Bedrock or Amazo
- Supervised learning uses known outcomes to learn categorical or numerical predictions
Supervised learning learns from inputs paired with known target outcomes, such as historical transactions marked fraudulent or genuine. Classification predicts a category, including two-category and multiple-category outcomes; regression predicts a numerical value such as a selling price. A business must identify both the outcome it wants predicted and whether reliable historical examples of that outcome exist.
Trap Using customer segmentation as the training objective for an exact revenue forecast, which discovers groups rather than learning the required numerical target.
4 questions test this
- A real estate company wants to predict the selling price of each listed home from features such as size, location, and age. It has ten years of completed sales with final prices. An analyst proposes g
- A bank wants to sort incoming customer emails into five service categories, such as billing, cards, and loans. It has two years of emails already tagged with the correct category by agents. The operat
- A utility company wants to forecast each customer's electricity use for next month in kilowatt-hours. It has five years of monthly meter readings. A consultant suggests a classification model that lab
- A subscription company wants to know which customers will cancel next month so it can offer retention discounts. It has three years of customer records showing who cancelled and who renewed. The marke
- Unsupervised learning discovers relationships without examples of a required answer
Unsupervised learning finds patterns in data without supplied target labels. Clustering groups similar customers or documents, while association discovery identifies relationships such as products frequently purchased together. These findings support exploration, but the resulting groups need business interpretation rather than being treated as predefined outcomes the system was taught to predict.
Trap Treating newly discovered customer clusters as verified churn labels, which substitutes similarity groups for observed business outcomes.
4 questions test this
- A bank wants to design product bundles, such as a checking account with a credit card, but has never offered bundles before. It holds five years of records showing which products each customer holds.
- A telecom company clustered two million customers by usage and found seven segments. The marketing director plans to use the cluster output as each customer's probability of buying a new premium plan
- An insurer ran clustering on five years of policyholder data and found a group with unusual coverage combinations. A product manager wants to label this group as high-risk and raise their premiums at
- A city transit agency wants to discover distinct rider travel patterns in two years of fare-card trip records to guide a route redesign. Riders are tagged only by fare type, such as student, senior, o
- Limited manual labels do not mean all learning must stop
Semi-supervised learning uses a small labeled dataset together with a larger unlabeled dataset, which can be useful when expert labeling is costly. Self-supervised learning derives learning targets from the data itself, as in learning to predict missing or subsequent text, and is widely used in foundation-model pretraining. These approaches differ from collecting a human-provided correct answer for every training example; neither removes the need to evaluate results against the intended business task.
Trap Requiring experts to label every document before any pretraining can begin, which overlooks learning signals obtainable from unlabeled data.
4 questions test this
- A media company wants to build a language model for its archive. A project manager insists that editors must label every article with the correct next sentence before any training can start, which wou
- A software company is evaluating a foundation model for internal search. An executive objects that the model cannot be trustworthy because no human labeled the billions of documents it was pretrained
- A support center wants a model that sets ticket priority. Senior agents labeled 2,000 tickets by priority, and 400,000 older tickets have no labels. A vendor proposes semi-supervised learning, and the
- A hospital wants to classify radiology reports by urgency. Radiologists can label only a few hundred reports, and the hospital has tens of thousands of unlabeled reports. The chief medical officer wan
- Reinforcement learning learns actions from rewards rather than paired correct answers
Reinforcement learning improves a strategy through actions and feedback expressed as rewards, with the aim of improving cumulative outcomes over time. A locally attractive action can be less useful than a sequence that earns a better long-term result. Business leaders must understand what the reward encourages, because a system can learn to pursue the rewarded outcome even when that outcome is a poor proxy for the real business goal.
Trap Rewarding only immediate engagement for a long-term retention goal, which teaches the system to optimize a different outcome.
3 questions test this
- A courier company uses reinforcement learning to assign drivers to orders. The reward counts deliveries completed each hour, and drivers now skip distant high-priority orders that carry service-level
- A logistics company is piloting reinforcement learning to schedule warehouse shifts. A product manager says the system will work out what the business wants on its own, so leaders do not need to agree
- A streaming service uses reinforcement learning to choose which shows to recommend. The current reward is the number of clicks in each session. Leadership's real goal is long-term subscriber retention
- Anomaly detection identifies unusual behavior rather than proving its cause
Anomaly detection identifies observations that differ from expected patterns, which can help surface unusual transactions or equipment behavior when exhaustive examples of failures are unavailable. An unusual observation is a signal for investigation, not by itself proof of fraud or a particular equipment fault. A known-outcome classifier instead learns the categories represented in its labeled examples, so the two capabilities answer different business questions.
Trap Treating an unusual transaction as a confirmed fraud label, which turns a deviation signal into an unverified conclusion.
3 questions test this
- A manufacturer's inspection logs contain thousands of labeled examples of three known defect types on its packaging lines. It now wants to name those known defects automatically and also surface unusu
- A manufacturer wants to detect equipment problems on a new production line. It has no history of failures for this equipment, but it does have sensor data from normal operation. The plant manager also
- A payments company has few confirmed fraud labels and deploys anomaly detection to surface unusual transactions. A manager proposes automatically blocking every flagged transaction and recording each
- Data structure depends on its schema, not whether it contains numbers or words
Structured data follows a predefined schema, such as transaction rows with named fields for customer, date, and amount. Unstructured data includes narrative documents, images, and recordings whose meaning is not organized as fixed record fields. Short text fields can be structured, and a digital document can remain unstructured; the distinction helps determine whether a business needs record analysis or content understanding.
Trap Classifying all text as unstructured, which ignores schema-defined fields such as customer identifiers and product categories.
4 questions test this
- A hospital's operations team wants to forecast bed demand from admission records that include patient ID, ward, admission date, and diagnosis code. A data vendor insists the records need natural langu
- An insurance company stores adjuster reports as free-form narrative documents and also keeps a claims table with policy number, claim date, and status. Leadership wants to know which source needs cont
- A utility company wants to analyze two sources: a meter table with account number, reading date, and kilowatt-hours, and thousands of scanned letters in which customers describe billing disputes. A ma
- A bank's product team is building a dashboard of complaint trends. The complaint system records complaint category, product, branch, and date in named fields, and it also stores a long free-text descr
- Metadata organizes content without making all of its meaning structured
Semi-structured information combines organizational cues with content that does not fit a rigid table, such as a media collection with dates and topic tags. Those tags can support filtering and discovery, while understanding what a recording actually says still requires analysis of its content. A business should distinguish searching known metadata from extracting new meaning from the underlying material.
Trap Using document filenames as substitutes for analyzing contract obligations, which searches labels rather than the terms inside the documents.
4 questions test this
- An e-commerce company stores product photos in a library where each image has metadata for product ID, upload date, and photographer. The merchandising team wants to find photos that show a product be
- A pharmaceutical company keeps research reports in a repository where each report has tags for project code, author, and approval date. A compliance officer must confirm which reports mention a specif
- A hotel group stores two years of guest reviews as JSON records with a star rating, stay date, property code, and a free-text comment. Leadership asks how often guests at its downtown property complai
- A media company's archive of interview videos is tagged with recording date, location, and topic. Editors can already filter by those tags. The content director now wants to find every clip in which a
- An AI solution must support the information modalities the business needs it to use
A modality is a type of information, such as text, images, audio, or video; a multimodal capability can process more than one type. A text-only capability cannot directly interpret visual evidence in a product photograph or a scanned chart merely because a user mentions it. For a mixed-information task, establish which modalities must be understood and whether any required conversion preserves the evidence needed for the decision.
Trap Using a text-only summary to judge visual damage that was never described in text, which removes the evidence needed for the assessment.
4 questions test this
- An auto insurer wants AI to help adjusters assess vehicle damage from photos that policyholders upload. A proposed tool accepts only text, so the team plans to send it the claim form's written descrip
- A manufacturer wants AI to triage maintenance requests. Technicians submit a typed note, a photo of the faulty part, and sometimes a short audio clip of machine noise. Plant leadership wants triage de
- A property management company wants AI to summarize move-out inspections. Inspectors record a video walkthrough of each unit and write brief notes, and deposit deductions depend on visible damage such
- A retailer is evaluating an AI assistant for store managers. Managers want to upload shelf photos and ask whether displays follow the planogram, and they also want to type questions about stock levels
- Recognizing text and extracting business fields are different document-processing capabilities
Optical character recognition converts an image of text into machine-readable text. Intelligent document processing goes further by extracting, classifying, and validating business information such as invoice numbers, totals, and payment terms for downstream workflows. When the goal is to populate a business record rather than simply make a document searchable, assess field-level extraction and validation as well as text recognition.
Trap Accepting a searchable invoice as proof that its payment fields were correctly captured, which confuses text recognition with business-data extraction.
5 questions test this
- An accounts payable team scanned all supplier invoices and ran optical character recognition, so the invoices are now searchable in the document system. The controller now wants invoice numbers, total
- An insurer digitized its paper claim forms with OCR, and adjusters can now search for any word in a form. The claims director wants claim amounts and policy numbers loaded into the claims system witho
- A bank's mortgage team receives pay stubs and tax forms as images. A completed pilot made them searchable, and underwriters still retype income figures into the loan system. Leadership wants income va
- A logistics company receives thousands of bills of lading as scanned PDFs. Finance wants shipment weights and freight charges entered into the billing system automatically, and audit requires that tot
- A hospital wants patient intake forms to update its records system automatically. The forms are scanned, and a pilot made them searchable as text. The records team notices that the system still has no
- Incorrect and missing records are data-quality problems before they are model problems
When predictions rely on inaccurate records or missing business information, prioritize correcting the source data and assessing its suitability before retraining. Accuracy means that recorded values reflect reality, while completeness concerns whether required information is present. Timeliness concerns currency and relevance, so an accurate old record may still be unsuitable for a decision that needs current information. Adding more examples from the same faulty process can enlarge the problem rather than supply better learning evidence.
Trap Funding a more sophisticated model for contradictory customer records, which leaves the incorrect learning evidence unchanged.
5 questions test this
- An online retailer's product recommendation model suggests discontinued items. The catalog feed marks discontinued products inconsistently, and many product records have no category. Merchandising wan
- A grocery chain trains a replenishment model on supplier delivery records. Suppliers updated delivery schedules three months ago, but the records used for training are from last year. The records are
- A telecom company's churn model gives poor predictions. An audit finds that many customer records list the wrong contract end date and that a third of records lack a plan type. The CTO proposes buying
- A bank's credit risk model was trained on income values that branch staff typed by hand, and an audit shows many entries are off by a factor of ten. The model team proposes adding records from three m
- A health insurer's claims model flags too many valid claims for review. Investigation shows that a legacy intake system stored some procedure codes incorrectly and left provider specialties blank on m
- Consistent definitions and valid values make data comparable across business sources
Data consistency requires equivalent business information to have compatible meanings and representations across records and sources. Validity checks whether values follow required formats or acceptable ranges; a value can pass a format check and still be factually wrong. When stores record the same business event differently, standardize collection definitions and quality checks rather than merely standardizing the AI's final reports.
Trap Standardizing forecast report layouts to reconcile incompatible store inputs, which changes presentation without making the underlying measurements comparable.
4 questions test this
- A national retailer trains a returns forecast on data from its stores. Some stores log a return when the customer drops off an item, and others log it only after inspection approves a refund, so retur
- A logistics company merges shipment records from two acquired carriers. One carrier records weight in pounds and the other in kilograms, and both fields pass the format check for a numeric value. The
- A restaurant chain wants one demand forecast across all locations. Some stores record sales by order and others by individual menu item, and some log times in local time while others use headquarters
- A hospital group combines readmission data from five facilities to train a risk model. Each facility defines a readmission differently, using windows of 14, 30, or 60 days. All records pass the data w
- Large historical datasets can still omit the people and situations that matter
Training evidence must represent the intended users and operating conditions, not simply contain many records. Underrepresented populations, rare outcomes, or missing business contexts can lead to uneven performance even when overall results look strong. Seek relevant coverage of those gaps and document exclusions; collecting more of the already dominant examples does not necessarily address them.
Trap Expanding only the majority customer segment to fix weak minority coverage, which increases volume without filling the missing experience.
4 questions test this
- A hospital wants a model to predict patient deterioration. The training data comes from adult wards, but the hospital plans to use the model in its pediatric ward as well. A vendor notes that the data
- A food delivery company trained an arrival-time model on four years of orders, nearly all from its southern cities in mild weather. It will launch in northern cities this winter, where snow is common.
- A bank trained a loan approval model on ten years of applications, almost all from urban customers. The model performs well overall but poorly for rural applicants, a segment the bank is expanding int
- An e-commerce company's support chatbot was trained on English-language tickets from its largest market. The company is now expanding to Spanish-speaking customers and plans to add two more years of E
- Training labels require quality control whether people or AI produced them
A label tells a supervised-learning system what answer to learn, so inconsistent or biased labels can teach the wrong associations. Require clear labeling guidance, checks against known examples, and review of disagreements by appropriate domain experts. Machine-generated labels also need validation against human judgment, especially for underrepresented groups and difficult cases; automation changes how labels are produced, not the need to establish their reliability.
Trap Treating agreement among automatically generated labels as independent proof of correctness, which can reproduce the same underlying labeling error.
4 questions test this
- An insurer is training a claims triage model and uses an automated tool to pre-label claims as simple or complex. Adjusters review only claims the tool marks uncertain. The model performs worse on cla
- A bank uses a language model to label thousands of transaction descriptions by spending category. Three different prompts produce the same label for most transactions, and the project lead argues that
- A medical imaging startup has radiologists label scans as normal or abnormal. Two radiologists often disagree on borderline scans, and one consistently labels more scans as abnormal than colleagues. T
- A retailer uses contractors to label product reviews as positive, negative, or neutral for a sentiment model. Early results show inconsistent labels for sarcastic reviews. The product owner wants reli
- Synthetic examples are useful only when they preserve the business phenomena being learned
Synthetic or augmented data can supplement scarce examples, but its usefulness depends on realism, relevant relationships, and coverage of important edge cases. Require domain validation and checks for bias rather than accepting generated volume as evidence of quality. Document where synthetic examples can substitute for real observations and where their limitations make real data necessary.
Trap Replacing rare real incidents with generic generated examples, which may remove the unusual relationships the system needs to recognize.
4 questions test this
- A payments company has only a few hundred confirmed cases of a rare fraud pattern. A vendor offers to generate one million synthetic fraud transactions and replace the real cases with them, saying vol
- A bank plans to train its support chatbot on 200,000 conversations generated by a language model because it has only 5,000 real chat transcripts. In a pilot, the bot handles routine questions but stum
- A bank wants to train a credit model for a new small-business product but has little real data. A team proposes using a generative model to create synthetic applicants. The chief risk officer is conce
- A manufacturer wants a model to detect a rare equipment failure. It has 40 real failure records and plans to generate synthetic sensor data to augment them. The engineering lead wants the model to wor
- Data profiling reveals issues, but business expertise determines what they mean
Data profiling examines characteristics such as missing values, distributions, unusual observations, and inconsistent categories before training. Subject matter experts must interpret those findings because an unusual value may be either a recording error or a valid business event. Establish quality expectations and record the rationale for remediation instead of blindly accepting an automated suggestion or removing every unusual case.
Trap Removing every unusually large transaction before fraud training, which may discard legitimate examples of the events being investigated.
4 questions test this
- A utility company profiles smart-meter data before training a consumption forecast. The tool flags extreme usage spikes on several days. Engineers know that some spikes were caused by a heatwave and o
- An insurer profiles policy data before training a pricing model. The tool reports that a vehicle category appears under three spellings and that some drivers are listed as 110 years old. The pricing m
- A retail company profiles customer data for a churn model and finds many missing ages and a few transactions with negative amounts. The data engineer proposes accepting the profiling tool's automatic
- A card issuer is preparing data to train a fraud model. Automated profiling flags a cluster of very large transactions as outliers and suggests removing them. Some may be data entry errors, but some m
- Duplicate records can distort evidence even when every field is correct
Uniqueness is the data-quality dimension concerned with avoiding unintended duplicate records for the same business entity or event. Recording the same customer twice can inflate customer counts even when both records are accurate and complete. Assess whether repeated records represent genuinely separate observations before accepting record volume as evidence of the size of the business population.
Trap Using a required-field completeness check as the sole acceptance criterion for customer counts, because duplicate records can contain every required field.
- Training, validation, and testing provide different kinds of evidence
Training data teaches the model, validation data supports development decisions, and a separate test set assesses performance after those decisions are made. A business approval should rely on examples not used to fit or repeatedly refine the model. The same independence principle applies to prompt engineering: use development cases to refine instructions and held-out cases to verify the final prompt. This distinguishes a reusable improvement from a result tailored to familiar examples.
Trap Using the final test set for every development revision, which makes the claimed independent assessment part of the development process.
3 questions test this
- An insurer is building a claims severity model. The team plans to use one dataset both to tune model settings and to report final accuracy to the regulator, because splitting the data would leave fewe
- A software company tested several large language model prompts for drafting sales emails. The product manager picked the winning prompt based on scores from a test set and then made six more edits, ch
- A retailer's data science team tuned a demand model over twenty iterations, checking each revision against the same test set to pick the best version. They now present that test score to the steering
- Information unavailable at decision time must not inflate historical evaluation results
Data leakage occurs when learning or evaluation uses information that would not legitimately be available for the real prediction. Examples include later business outcomes present in historical records and duplicate examples appearing across training and testing. When reviewing impressive results, ask whether the evidence reflects the information available at the actual decision point and whether evaluation examples are genuinely separate.
Trap Including completed-claim settlement information in an initial claim-risk prediction, which supplies knowledge the business would only obtain later.
2 questions test this
- A retailer's product review classifier shows near-perfect test accuracy. An analyst finds that many reviews were posted on several sites and appear in both the training and test sets. The product owne
- A subscription company trained a churn model using a feature that records whether a customer called to cancel. Test accuracy is excellent. The marketing director plans to use the model to target reten
- Training success must be compared with performance on unfamiliar examples
Overfitting occurs when a model learns training-specific detail or noise and performs well on training examples but poorly on new examples. Underfitting means the model has not learned the relevant relationship and performs poorly on both training and test examples. These patterns call for different technical investigations; business leaders should request evidence of generalization rather than assuming longer training or higher training accuracy is always better.
Trap Demanding more training solely to maximize an already strong training score, which can deepen memorization instead of improving generalization.
4 questions test this
- A logistics company's delivery time model performs poorly on both its training data and its test data. A manager suggests that the problem is overfitting and wants to reduce the amount of training to
- A bank's fraud model was retrained with many more features and now shows 100 percent accuracy on training data. Its accuracy on a held-out test set dropped from 92 to 80 percent. The project sponsor c
- A marketing analytics team reports that its new lead scoring model achieves 99 percent accuracy on training data but only 70 percent on a held-out test set. The sales director asks the team to keep tr
- A retailer is contracting a vendor to build a weekly demand forecasting model. The vendor proposes that the retailer accept the model once training accuracy reaches 99 percent. The merchandising direc
- Historical evaluation must address the specific context in which the model will be used
A national or historical average test score does not establish performance for a particular city, customer group, or newly changed market. Request evaluation on relevant subsets and, after a major business shift, suitably recent examples before relying on the broader result. The decision is whether the evaluation evidence supports the intended application, not whether a single aggregate score is impressive.
Trap Accepting a national average test result as sufficient evidence for entry into a poorly represented city, which leaves local performance unexamined.
3 questions test this
- A retail chain's markdown model was tested on three years of data and achieved strong overall accuracy. The chain recently launched a new online-only product line, and the test data contains almost no
- A health insurer's claims model performs well on its overall test set. The insurer plans to use it for a newly acquired group of members over age 80, who make up less than 1 percent of the test data.
- A national bank's credit model has a strong average test score across the country. The bank plans to expand into a city where few customers appear in the training and test data. The expansion lead wan
- ISO/IEC 22989 and ISO/IEC 23053 support different levels of shared AI understanding
ISO/IEC 22989 establishes AI concepts and terminology for communication among diverse stakeholders. ISO/IEC 23053 provides a framework for describing the components and functions of AI systems using machine learning, including deep learning. Use shared vocabulary to resolve inconsistent terminology and the ML-system framework to explain how system elements relate; neither serves the same purpose as an organizational AI management system.
Trap Using ISO/IEC 42001 as the primary glossary for a terminology dispute, which selects a management-system standard rather than the dedicated vocabulary standard.
4 questions test this
- A healthcare technology company is preparing training materials so clinicians, engineers, and lawyers can discuss AI projects consistently. The training must first align everyone on basic AI terms and
- A bank's architecture review board must describe its machine learning credit systems consistently, covering components such as data, the trained model, and processing, and how they function together.
- A manufacturer has adopted ISO/IEC 22989 terminology across its AI teams and uses ISO/IEC 23053 to document its ML pipelines. The CEO now wants organization-wide policies, objectives, and continual im
- A global insurer's business and technical teams use terms such as model, training data, and AI system inconsistently, which slows approvals for AI projects. The chief data officer wants an internation
- ISO/IEC 42001 addresses organizational AI management rather than an individual model's accuracy
ISO/IEC 42001 specifies requirements for establishing, maintaining, and continually improving an AI management system. It applies to organizations that provide or use AI, not only those that develop models, and uses a Plan-Do-Check-Act approach to managing risks and opportunities. Its focus is coordinated policies, objectives, and processes across the organization rather than a substitute for evaluating an individual AI application's outputs.
Trap Using a model benchmark report as the organization's AI management system, which measures outputs without establishing management policies and improvement processes.
3 questions test this
- A retailer uses third-party AI services for pricing and customer support but builds no models itself. The compliance director wants to know whether ISO/IEC 42001 applies to the company. A manager argu
- A bank's AI team reports that its fraud model achieved excellent accuracy on an industry benchmark. The chief risk officer wants to align the bank's AI program with ISO/IEC 42001, and the AI team prop
- A logistics company plans to deploy several AI tools from different vendors and to build one forecasting model in-house. The CEO wants a single organizational approach for managing AI risks and opport
- The NIST AI RMF provides a voluntary structure for managing AI risks across the lifecycle
The NIST AI Risk Management Framework is voluntary, use-case agnostic guidance for incorporating trustworthiness into AI design, development, use, and evaluation. Its core functions are Govern, Map, Measure, and Manage: establish governance, understand context and risks, assess those risks, and prioritize responses. Govern is cross-cutting rather than a one-time first step; organizations tailor the risk-management outcomes and approaches to their application and context.
Trap Treating adoption of the NIST AI RMF as regulatory permission to deploy, which confuses voluntary risk guidance with applicable legal requirements.
3 questions test this
- A health insurer has adopted the NIST AI Risk Management Framework for a new claims AI tool. A product manager says that adopting the framework gives the company regulatory permission to deploy the to
- A bank's AI program office wants to use the NIST AI RMF to structure risk work for a new loan assistant. The program manager asks which functions to use to understand the system's context, assess its
- A retailer is applying the NIST AI RMF to a pricing AI system. The project team treats Govern as a step to finish in the first week and then plans to move on to Map, Measure, and Manage without revisi
- AWS CAF for AI treats adoption as an organizational capability challenge
AWS CAF for AI applies six perspectives to AI adoption: Business, People, Governance, Platform, Security, and Operations. These connect business outcomes and workforce capability with oversight, enabling technology, protection, and reliable operation. A technology assessment alone therefore cannot establish overall AI readiness; the perspectives provide a common way for business and technical stakeholders to discuss gaps without prescribing one executive owner for every organization.
Trap Using platform capacity as a complete AI-readiness assessment, which leaves people, governance, and other organizational capabilities unassessed.
3 questions test this
- A retailer wants to scale generative AI from pilots to production. Its platform team has delivered strong infrastructure, and leadership wants to confirm readiness before committing budget. The COO as
- A bank's AI program is stalling. The data science team has strong tools, but business units do not sponsor use cases, staff lack AI skills, and no one owns AI risk oversight. The chief transformation
- A manufacturer's CIO assessed AI readiness by reviewing GPU capacity, data platforms, and cloud tooling, and concluded the company is fully ready to scale AI. The CEO notices that business sponsors, w
- AWS responsible AI dimensions describe distinct concerns that a single quality score cannot replace
AWS describes responsible AI through fairness, explainability, privacy and security, safety, controllability, veracity and robustness, governance, and transparency. Explainability concerns understanding and evaluating outputs, while transparency enables informed stakeholder choices; controllability concerns mechanisms to monitor and steer behavior. Fairness considers impacts on different groups, and veracity and robustness address correct outputs even with unexpected inputs, so strong performance on one dimension does not establish all the others.
Trap Using a transparency disclosure as evidence that individual decisions are understandable, which substitutes stakeholder information for output explainability.
3 questions test this
- An insurer's AI claims assistant passed a bias review across customer groups. The board wants assurance on other responsible AI concerns before launch, specifically whether adjusters can understand an
- A retailer's chatbot scores highly on answer accuracy in testing. Before launch, the responsible AI lead finds that the bot gives harmful advice when users enter unexpected adversarial prompts, and th
- A bank publishes a notice telling loan applicants that AI is used in credit decisions and describing the system's intended use and limitations. A regulator later asks the bank to show why a specific a
Also tested in
References
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain1.html
- https://aws.amazon.com/what-is/generative-ai/
- https://aws.amazon.com/what-is/foundation-models/
- https://docs.aws.amazon.com/bedrock/latest/userguide/key-definitions.html
- https://aws.amazon.com/what-is/machine-learning/
- https://aws.amazon.com/what-is/large-language-model/
- https://aws.amazon.com/ai/responsible-ai/
- https://aws.amazon.com/what-is/artificial-general-intelligence/
- https://aws.amazon.com/what-is/inferencing/
- https://docs.aws.amazon.com/bedrock/latest/userguide/inference-parameters.html
- https://aws.amazon.com/what-is/reinforcement-learning/
- https://docs.aws.amazon.com/machine-learning/latest/dg/types-of-ml-models.html
- https://aws.amazon.com/what-is/anomaly-detection/
- https://aws.amazon.com/blogs/machine-learning/overcome-the-machine-learning-cold-start-challenge-in-fraud-detection-using-amazon-fraud-detector/
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01.html
- https://aws.amazon.com/what-is/nlp/
- https://aws.amazon.com/what-is/computer-vision/
- https://aws.amazon.com/what-is/speech-to-text/
- https://docs.aws.amazon.com/personalize/latest/dg/what-is-personalize.html
- https://aws.amazon.com/bedrock/faqs/
- https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html
- https://docs.aws.amazon.com/quick/latest/userguide/what-is.html
- https://aws.amazon.com/what-is/structured-data/
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain4.html
- https://aws.amazon.com/what-is/data-silos/
- https://docs.aws.amazon.com/AmazonRDS/latest/AuroraUserGuide/postgresql-ml.html
- https://aws.amazon.com/blogs/machine-learning/generative-ai-and-multi-modal-agents-in-aws-the-key-to-unlocking-new-value-in-financial-markets/
- https://aws.amazon.com/what-is/ocr/
- https://aws.amazon.com/what-is/intelligent-document-processing/
- https://aws.amazon.com/what-is/data-cleansing/
- https://docs.aws.amazon.com/wellarchitected/latest/devops-guidance/metrics-for-data-lifecycle-management.html
- https://docs.aws.amazon.com/glue/latest/dg/dqdl-rule-types-Uniqueness.html
- https://aws.amazon.com/what-is/data-integrity/
- https://docs.aws.amazon.com/wellarchitected/latest/machine-learning-lens/mlops03-bp01.html
- https://docs.aws.amazon.com/wellarchitected/latest/responsible-ai-lens/raidp02-bp01.html
- https://docs.aws.amazon.com/wellarchitected/latest/responsible-ai-lens/raidp02-bp03.html
- https://aws.amazon.com/what-is/synthetic-data/
- https://docs.aws.amazon.com/wellarchitected/latest/responsible-ai-lens/raidp02-bp04.html
- https://docs.aws.amazon.com/prescriptive-guidance/latest/ml-operations-planning/splits-leakage.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/design-a-prompt.html
- https://aws.amazon.com/what-is/overfitting/
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain3.html
- https://committee.iso.org/standard/74296.html
- https://committee.iso.org/standard/74438.html
- https://committee.iso.org/standard/42001
- https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
- https://docs.aws.amazon.com/whitepapers/latest/aws-caf-for-ai/ai-cloud-transformation-value-chain.html
- https://docs.aws.amazon.com/pdfs/whitepapers/latest/aws-caf-for-ai/aws-caf-for-ai.pdf