AWS Certified AI Business Strategist (AIB-C01) Course

Machine learning (ML) is part of artificial intelligence: it learns patterns from examples to make predictions. A training algorithm is the procedure that learns; a model is the learned result. The algorithm and the model are not the same thing.

A rule-based system follows conditions that people write explicitly. For example: if an invoice amount exceeds the approval limit, require manager approval. The decision comes from a stated policy, not from examples of earlier invoices.

Artificial intelligence (AI) is the broad field of building systems that perform tasks associated with intelligence, such as reasoning, perception, planning, and language processing. AI describes a wide range of capabilities, not one way of building software.

The clearest distinction is the source of the behavior. In a rule-based system, people explicitly write the conditions and conclusions. For example, an expert writes: if a device is reachable locally but cannot contact its gateway, flag a possible gateway connection problem. The system applies that supplied relationship; it did not discover it from examples.

Model inference means running a trained machine learning model on an input to obtain an output. For example, a delivery-time model can use a new shipment’s details to estimate how long delivery will take. Inference makes the model’s learned patterns useful for a particular case.

Amazon SageMaker AI helps teams build, train, and deploy machine learning models. These machine learning capabilities belong to the broader Amazon SageMaker platform. This lesson follows a retailer that wants to predict how many items it will sell, so its application can help staff plan stock.

A model is a system that uses learned patterns to produce outputs from inputs. A foundation model is trained on broad data so that it can support many tasks. Depending on the model, those tasks can include drafting text, summarizing documents, or generating images.

Rule-based AI uses knowledge that people write as explicit rules. It applies those rules to available facts to derive a conclusion or select an action. This supports reasoning without requiring model training: the relationships come from human authors.

Amazon Comprehend is a natural language processing (NLP) service: it analyzes human language in text. It uses machine learning (ML), in which models learn patterns from data rather than relying only on rules written by people. ML is a subset of artificial intelligence (AI), the broader field of systems that perform tasks associated with human intelligence.

Generative AI produces content using patterns learned during training. Those patterns can support writing a reply, summarizing a report, creating an image, or suggesting code. Generating content does not guarantee that it is original, correct, or suitable for its intended use.

Amazon Comprehend provides specialized APIs for analyzing text. An API is an interface an application calls to request a defined task. For example, an application can ask Comprehend to detect sentiment or identify people and places mentioned in a review. The result describes the supplied text rather than composing a new message.

Structured data organizes information into consistent fields with defined data types. A schema describes that layout: which fields exist, what they mean, and which types of values they accept. It can also define rules such as required fields or unique identifiers.

Unstructured data includes free-form text, photographs, recordings, and video. Its meaningful content does not follow a predefined analytical schema. A photograph, for example, does not provide ready-made fields describing every object it shows.

Structured data organizes information into predefined fields with consistent meanings and data types. This arrangement is called a schema. For example, each order record might have an order ID, an order date, and a numeric total.

Semi-structured data organizes records with named fields and relationships. JSON is a common example. A key names a field, and its value holds information: in {"text":"Delivery was quick","label":"positive"}, the keys connect a customer comment with its sentiment label.

A record describes one thing, such as a purchase. Structured data commonly represents records as rows in a table. The table’s schema defines shared columns, their meanings, and their data types. Every row follows those column definitions, even when an allowed value is missing.

Semi-structured data uses named keys, tags, or other markers to organize a record without requiring a uniform relational table. Unstructured data contains meaning that is not divided into predefined analytical fields, such as the ideas expressed in a customer review.

Amazon RDS is a managed service for relational databases. A relational database organizes data into tables. Each row is one record, such as an order. Each column is a named attribute, such as the order's amount. You define the tables your application needs.

Amazon S3 is object storage. An object contains stored data, such as a spreadsheet export, photo, or audio recording, together with metadata describing the object. A bucket is a container for objects. This lesson uses general purpose buckets.

Amazon RDS is a managed relational database service. It suits records such as customers, orders, and payments that an application needs to look up, change, and connect. A relational database organizes records in tables and uses SQL, a database language, to query them. For example, it can find a customer's orders and total their payments.

A scanned invoice is easy for a person to read, but its contents are not yet separate fields a program can reliably use. The page might contain a customer name, an invoice date, and a table of purchases. In an image, these are pixels rather than named fields and individual records. This is unstructured document data, even when the page looks neatly organized.

Data quality means fitness for a specific purpose. It is not a permanent badge attached to a dataset. A daily inventory snapshot may be suitable for forecasting future demand but too old for an AI assistant that promises customers what is available right now.

Data representativeness means how well a dataset reflects the people, cases, and conditions an AI system is intended to serve. The target population is that intended set of users or cases. Representativeness is relative to a purpose: records from urban stores do not establish coverage of rural stores, even when there are millions of records.

A label is an answer attached to a data item for a particular task. A support ticket might have the label “billing question,” while an image might have the label “bicycle.” Label quality means those answers fit the task, use suitable evidence, and follow consistent rules.

Amazon SageMaker Data Wrangler provides a visual interface for exploring and preparing data for machine learning. Data quality matters because a model learns patterns from its training data. Missing measurements can hide useful signals, while inconsistent category names can make one real-world category look like several different ones.

Amazon SageMaker Ground Truth helps prepare labeled datasets for machine learning. A label is an answer attached to an example, such as “damaged” on a product photograph. In supervised learning, the model uses these example-and-answer pairs to learn patterns for making predictions.

AWS Glue Data Quality evaluates data against rules you choose. Two important properties are completeness: whether required values are present, and validity: whether values meet defined requirements. A value can be present but still be invalid.

A dataset that passed yesterday can fail today. A source can stop sending a field, a transformation can introduce duplicates, or incoming records can become outdated. Ongoing quality controls repeat measurable checks so that earlier approval does not become permanent approval.

Amazon SageMaker Data Wrangler helps you prepare datasets through an interactive visual interface. You can import data, inspect it, apply transformations, and view the results. Its main focus is helping data scientists and analysts get data ready for machine learning (ML), although it also supports analytical preparation.

Supervised learning learns from examples that pair an input with a known answer. The input is described by features: information available to help make a prediction. The known answer is the target, also called the label.

A train-validation-test split divides historical data into separate sets, called partitions. In supervised learning, each example has inputs and a known answer, or label. Separating examples helps distinguish learning from checking how well that learning works on unseen data.

A model learns from training data: examples used to adjust its predictions. Its results on those examples are its training performance. Those results show how well it fits familiar examples, not necessarily how useful it will be on new ones.

ISO/IEC 23053 provides a framework for describing AI systems that use machine learning. It describes system components, their functions, and their relationships so that teams can discuss the same system without depending on a supplier’s product names.

ISO/IEC 42001:2023 defines requirements for an artificial intelligence management system (AIMS). This is the combination of policies, responsibilities, processes, and documented information that helps an organization manage AI responsibly and consistently. It is not a model, a software tool, or a one-time approval.

ISO/IEC 23053 provides a framework for describing AI systems that use machine learning. It helps people explain system components, their functions, and their relationships across technologies and organizations. Its central question is: How does this ML-based AI system work? For example, a description can connect data preparation, model training, and the application of a model to new inputs.

A rule-based system checks supplied facts against conditions written by people, then produces an action or conclusion. A rule has an if–then shape: if an invoice amount is greater than 100, then require manager approval. Here, 100 is an example policy threshold, not a cloud service limit. Domain specialists define the policy; developers or policy authors express it as executable rules.

Determinism means that identical inputs, initial state, and execution conditions produce identical results. An invoice calculation can be deterministic when it uses the same line items, tax rate, and rounding policy each time.

Determinism describes repeatability: identical inputs, initial state, and execution conditions produce identical results. It answers, “Will this execution produce the same result again under the same conditions?”

AWS Lambda runs your code without requiring you to provision or manage servers. A Lambda function can run in response to an event: a message describing something that happened, such as a file upload.

Orchestration means coordinating work across several steps. AWS Step Functions and agents in Amazon Bedrock can both coordinate work, but they differ in who decides the next step: the developer’s workflow definition or a model’s planning.

An AI agent is software that selects actions toward a goal, observes their results, and decides what to do next. An answer-only application returns information or suggested steps; it does not carry out those steps. For example, explaining how to investigate a delayed order differs from actually looking up its shipment status.

Tool use lets a foundation model request capabilities outside its learned knowledge, such as looking up a shipment. The model produces a request; application code decides whether to allow it, runs the operation, and returns the result.

Tool use lets an AI application use capabilities outside the model's learned knowledge, such as looking up an order, calculating a total, or creating a ticket. The model can request an operation; application code executes it and returns the result. This makes the application more capable, but does not by itself make it an AI agent.

Task planning connects a goal to the actions needed to achieve it. A goal describes the desired outcome; a plan describes the work needed to reach it. For an agent asked to resolve a delayed order, success might mean giving the customer a verified resolution allowed by the business policy—not merely producing a convincing reply.

An AI agent uses a model, instructions, tools, and execution logic to take actions toward a goal. A single agent can often handle a task by choosing among several tools. Using several tools does not mean there are several agents.

An AI agent uses a model to help choose actions toward a goal. A text-only assistant can explain how to check an order; a tool-using agent can select an order lookup tool, request current information, and use the result to decide what to do next. A tool is a callable capability, such as a function or an API.

Amazon Bedrock provides managed access to foundation models. An application sends a request to a selected model and receives a response, such as generated text. AWS operates the underlying inference infrastructure. Model access supplies a reasoning or generation capability; it does not, by itself, host the application's complete agent program.

AWS Step Functions coordinates work across services. Instead of making each service decide which service to call next, you define a state machine: a workflow made of states and rules for moving between them.

Model drift describes changes in a model’s operating environment that can make its learned patterns less useful. The model itself can stay unchanged while customers, transactions, or behavior evolve. Good results before deployment therefore do not remove the need for production monitoring.

Model evaluation checks whether a model produces suitable results for its task. It must continue after deployment because users, incoming data, and business needs can change. A model that worked well at launch may not remain suitable.

Amazon CloudWatch collects monitoring data so you can follow the health of resources and applications. A metric is a series of numerical measurements over time: CPU utilization, request count, or prediction latency, for example. Latency measures how long a response takes.

Model retraining produces a new model version using refreshed, corrected, or revised training data. A deployed model normally keeps using what it learned during training; making predictions does not automatically teach it new patterns. Retraining creates a candidate model, not an automatic replacement.

Shadow AI is AI use without the organization's required authorization, oversight, or controls. It can include a personal chatbot account, a model API connected without review, or an unreviewed meeting assistant. The issue is missing governance, not how sophisticated the tool is.

An Acceptable Use Policy (AUP) defines which uses are acceptable, not just which actions are technically possible. For AI tools, an organization's policy tells staff which tools they may use for work and what they may do with them. These internal rules apply alongside the provider's terms, acceptable-use rules, and applicable law.

Shadow AI is AI used for work outside the organization’s approved oversight. For example, someone may paste work information into a personal AI account without realizing that the tool has not been reviewed. A visible AI tool register gives staff a shared place to check which tools they may use and under what conditions.

Allowlisting means an entity or action must be explicitly permitted before use within a controlled environment. For AI tools, the starting point is a visible tool register: a shared record of which tools may be used and under what conditions. Being available online does not make a tool approved for work.

AWS Service Catalog lets an organization offer approved IT products for deployment on AWS. A product defines an offering, such as a configured AI application and its supporting resources. A portfolio is a collection of products managed together. This gives people a clear place to find approved options rather than choosing tools without organizational review, a practice often called shadow AI.

An organization can classify AI tools as approved, blocked, or under evaluation. Publishing these decisions helps people distinguish permitted tools from shadow AI: AI used without organizational approval or oversight. These labels are governance decisions, not built-in AWS permission settings.

Prompt context is background information a model receives to understand a task. It can include reference documents, purchase details, previous conversation turns, or tool results. A support assistant answering a return question needs the applicable policy and purchase facts, not just the question. Supplying context helps it answer for that situation rather than relying only on patterns learned during training.

A prompt template is a reusable specification for a model request. It separates instructions that stay the same from values that change on each use. Reusing that structure helps an application ask for the same task consistently, rather than rebuilding its instructions for every request.

Iterative prompt refinement means improving a prompt through a repeatable cycle: define the desired result, test the prompt, identify failures, and revise it. The goal is better performance on the task, not simply an answer that sounds more polished.

A language model processes tokens, not words directly. A tokenizer splits text into units and maps them to numerical identifiers in its vocabulary. A token can represent a whole word, part of a word, punctuation, or other text. Spaces also affect tokenization.

A model's context window is the capacity available for one inference request. For text, it is measured in tokens: words, parts of words, punctuation, or other text fragments. Tokens are not the same as words, and different models can tokenize the same text differently. Always check the selected model version and API.

Conversation memory is information an application retains from earlier interactions. It can remain in a database between requests, even when no model is running. Saving a message preserves it for possible later use.

Conversation memory helps an assistant follow references such as “compare that with the cheaper option.” The application supplies earlier messages or a representation of them so the model can identify the options. The practical question is how much useful history can fit in the next request.

A language model can produce a fluent answer without knowing your organization's policies or its latest product changes. Retrieval-Augmented Generation (RAG) helps by finding external information relevant to a question and supplying it to the model before it answers.

An embedding is a numerical representation produced by a machine learning model. For text retrieval, it commonly takes the form of a fixed-length vector: an ordered list of numbers representing learned characteristics of a query or passage. Meaning is encoded across the numbers; individual numbers usually have no simple human-readable meaning.

Retrieval Augmented Generation (RAG) gives a language model relevant source material when it answers a question. Amazon OpenSearch Service can provide the retrieval part: finding passages that are related to the question.

Fine-tuning adapts an already trained model using targeted examples. Training updates its learned parameters, or trains additional adaptation parameters, rather than building its broad capabilities from scratch. The result is a reusable adaptation that persists across requests when you use that adapted model.

Fine-tuning and retrieval-augmented generation (RAG) improve responses in different ways. Fine-tuning trains a pretrained model on targeted examples, updating its learned parameters or adding trained adaptation parameters. This creates a reusable change that persists across requests—for example, learning to produce a particular kind of summary.

Semantic search finds information by matching the meaning of a request to the meaning of source material. An employee asking “Can I expense a train ticket?” could find a policy titled “Reimbursement for business travel,” even though the wording is different.

Text generation produces written content from instructions and context. Its business value comes from helping people complete useful work, not simply producing more words. Common outputs include drafts, summaries, translations, customer responses, and code.

Semantic search locates existing information by meaning. Text generation composes text in response to instructions. The useful distinction is the work being done: finding relevant source material versus writing a response.

Amazon Connect is a cloud-based contact center service for handling customer conversations, including voice and chat. It combines tools for directing contacts to the right place with AI capabilities that help customers and service teams resolve issues.

Amazon Personalize is a managed machine learning service that helps applications recommend products or content relevant to individual customers. For example, a shop can suggest products a customer may want, while a streaming application can suggest videos they may enjoy. AWS manages the underlying machine learning infrastructure.

Amazon Q has been in maintenance since June 30, 2026: it is closed to new customers, while existing customers can keep using it. AWS provides its migration guide at https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html. This lesson explains the assistants as AWS offered them, rather than recommending them for new workloads.

A key performance indicator (KPI) is a selected measurement that shows whether an important objective is being met. Start with a business need, identify the work AI could help with, and connect that capability to an outcome. For example, summarizing support conversations could reduce the time employees spend writing case notes.

Amazon Q entered maintenance on June 30, 2026. It is closed to new customers; existing customers can keep using it. This lesson explains Q as AWS offered it, not as a recommendation for new workloads. AWS's migration guide is https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html.

Lifecycle notice: Amazon Q Business has been in maintenance since June 30, 2026. It is closed to new customers; existing customers can continue using it. AWS provides migration guidance at https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html. This lesson explains the service as AWS offered it, not as a recommendation for new workloads.

Amazon Bedrock provides managed access to foundation models: models already trained on broad datasets that can support tasks such as generating text, summarizing documents, or creating images. You choose a supported model and call it from your application through an API. Available tasks depend on the model.

AI compute is the processing capacity used to train models or run them to make predictions. Amazon EC2 and Amazon SageMaker AI can both provide this capacity. The main difference is how much infrastructure your team controls and operates.

Total Cost of Ownership (TCO) is the complete cost of a solution over an agreed evaluation period. A fair comparison starts with a common baseline: the same business task, expected demand, output quality, availability, recovery needs, and security or compliance requirements. Use the same time period and currency for every option.

Total cost of ownership (TCO) asks, “What will this solution cost across its lifecycle?” Cost-benefit analysis asks, “Are the additional benefits worth the additional costs compared with a baseline?” The methods work together: TCO supplies cost estimates that cost-benefit analysis can use.

Total Cost of Ownership (TCO) measures the complete cost of owning and operating a solution over an agreed period. For an AI application, that includes more than service charges: development, integration, staff training, human review, and ongoing support can also contribute.

Cost-benefit analysis asks whether an approach creates enough extra business value to justify its costs and consequences. Start with a business objective, such as reducing the effort needed to process documents, rather than an objective such as deploying AI. A model's accuracy is not itself a business benefit: value comes from better decisions or useful work completed.

Vendor lock-in occurs when changing suppliers requires substantial effort, expense, or disruption. Migration may still be possible. The problem is the switching cost: everything needed to leave one provider and make a replacement work.

AWS Marketplace is a digital catalog for finding, buying, deploying, and managing third-party software, data, and services. For AI, it can help you find an existing product to buy or a provider to partner with, rather than building everything yourself.

AWS Cost Explorer shows AWS costs and usage over time. For an initiative such as an AI search pilot, the organization’s total bill is too broad: it includes unrelated work. A useful report separates the initiative’s spending from everything else.

AWS Pricing Calculator answers, “What would this proposed workload cost under these assumptions?” It produces planning estimates from service configurations and expected usage, such as compute hours, storage, and request volumes. You can estimate a new AI workload without having spending history for it.

Return on investment (ROI) expresses the net financial return from an initiative as a percentage of its investment cost. To estimate AI ROI, first define a baseline: what would happen without the initiative, such as continuing the existing manual process. Count improvements beyond that baseline, not all the value the business produces.

Payback period asks, “How long until we recover the initial investment?” Recovery happens when accumulated financial benefits, after ongoing costs, cover the upfront spending. Its answer is a duration, such as months or years. Shorter payback means earlier recovery.

Lifecycle environmental impact means the resources consumed and impacts created across an AI system’s life. Include data preparation, experiments, training, deployment, inference, storage, and retirement—not just the final training run. Inference is using a model to produce an answer or prediction.

Return on investment (ROI) expresses financial benefit after costs relative to the investment. It helps compare AI initiatives, but a positive ROI is not an automatic instruction to fund one. A percentage does not show the total financial gain or the amount of money at risk: a small project can have a higher percentage return but deliver less total value.

A stage-gate process is a way to govern investment as evidence grows. A stage contains work that reduces uncertainty, such as investigating a business problem or testing a prototype. A gate is a decision about whether that evidence justifies the next commitment. Completing the work does not automatically earn more funding.

Inference latency is the time needed to obtain an output from a deployed AI model. It applies to a fraud classifier as well as a text-generating assistant. The measurement boundary matters: model execution latency measures model processing, while end-to-end latency measures the application's full wait, including network travel, authentication, retrieval, queueing, and response delivery.

Failover transfers work from an unhealthy component to a healthy alternative. It supports high availability: keeping a service usable despite failures. It requires failure detection, an alternative ready to work, and a way to redirect requests.

Blue-green deployment keeps the current release and its replacement in parallel environments. Blue serves users while green is prepared and tested. A cutover redirects production traffic to the validated replacement instead of updating the application that is already serving users.

A Key Performance Indicator (KPI) is a selected measurement used to judge progress toward an objective. To measure change over time, first establish a baseline: the performance of the existing workflow before AI is introduced. Collect it before rollout, rather than trying to reconstruct it after the workflow has changed.

Labor productivity means useful output per hour of human effort: accepted units of output ÷ total labor hours. An accepted unit meets defined quality requirements, such as an invoice that passes validation. AI-generated drafts, suggestions, or documents are not accepted output merely because they exist.

Task completion rate measures how often an application achieves the user's intended outcome. The unit is a task, not a message or model call. Booking one appointment may require several messages, tool calls, and backend operations, but it is still one task.

User satisfaction measures how people perceive an application’s experience, not just the quality of its underlying AI model. Clear answers, useful context, waiting time, and appropriate help all shape that experience. A key performance indicator (KPI) is a measure chosen to track progress toward an important objective.

Task completion rate asks, “Did users achieve the intended outcome?” It measures the share of eligible attempted tasks that succeeded. User satisfaction asks, “How did users feel about the experience?” It reflects users’ judgments, commonly gathered through ratings or surveys. Both are useful key performance indicators for an AI application, but they describe different things.

Amazon QuickSight is a business analytics service for creating analyses and publishing interactive dashboards. A key performance indicator (KPI) is a measure selected to show progress toward a business goal. For an AI-assisted support team, the goal might be lower support costs, happier customers, or more productive employees—not simply more AI requests.

When work improves after an AI rollout, AI is not necessarily the cause. Easier tasks, extra staff, or seasonal changes could also help. To measure incremental AI impact, look for the improvement beyond what would have happened without AI.

A pre-AI baseline records outcomes from the existing process before AI is introduced. It answers, “How well do we perform today?” For a support service, measure successful issue resolution rather than simply counting messages sent. AI activity is not the same as business improvement.

AWS Pricing Calculator produces planning estimates from the services you select and the usage you expect. It applies supported pricing dimensions to those inputs to estimate charges. Creating an estimate does not deploy resources, reserve capacity, or purchase a discount commitment.

Payback period is the time needed for an initiative’s accumulated financial benefits, after ongoing costs, to recover its upfront investment. It helps you understand how long the money committed to a project remains unrecovered.

Sensitivity analysis asks how a calculated result changes when its assumptions change. For an AI business case, start with one measurement period and a clear model: ROI = (benefits − costs) ÷ costs × 100%. Compare benefits with a credible baseline without the AI investment.

A leading indicator is an early signal that may predict a later result. A lagging indicator measures a result already achieved. Start by naming the intended business outcome, then work backward to conditions that could make it happen.

AI cost estimates connect expected consumption to pricing units. For token-priced text generation, a token is a unit of text, not necessarily a whole word. Input includes instructions, conversation history, and supplied documents; output is the generated response. Input and output can have different rates. Estimate their charges separately using matching pricing units, then add them together.

Cost allocation attributes spending to the projects or teams responsible for it. An AI project's costs can include model training, inference, data storage, and networking. Pricing tells you what those services cost; allocation tells you whose spending they represent.

AWS Budgets lets you define a spending target and receive warnings about costs. A cost budget has a budget amount and a period, such as a monthly target of $1,000. The amount is a planning target, not prepaid credit or a spending restriction.

AWS Budgets and AWS Cost Explorer help you manage cloud costs, but they answer different questions. Choose between them by asking whether you need to receive a spending warning or examine spending data.

Right-sizing means choosing resource types and amounts that meet an AI workload’s needs without unnecessary capacity. Capacity includes processing power, memory, storage performance, and network capacity. The smallest or cheapest resource is not automatically the right choice: a model must fit in memory, and its work must finish within the required time.

Porter's Five Forces examines industry structure: the competitive pressures that constrain an industry's ability to earn profits. It asks, What pressures shape competition in this market? It is not simply a ranking of competing companies.

Porter's Five Forces explains how industry structure puts pressure on profitability: how much businesses can retain after their costs. It examines rivalry, new entrants, substitutes, supplier power, and buyer power—not just direct competitors. Before applying it, define the industry, customer segment, geography, and planning horizon. A retailer using AI faces different pressures from a company selling AI software.

SWOT stands for strengths, weaknesses, opportunities, and threats. It helps an organization choose actions that could improve its position against competitors or alternatives. The aim is not simply to fill four boxes: it is to connect what the organization can do with what customers and the market need.

First-mover advantage is the potential benefit of acting before competitors. In AI, this can mean solving an important problem for a particular customer group before others do—not inventing a model. A head start is about timing; a lasting advantage is about retaining better business outcomes after competitors respond.

The Business Model Canvas describes how a business creates, delivers, and captures value. Its nine connected blocks describe a business model, not a technical architecture. For an AI opportunity, compare the current canvas with a proposed future canvas rather than starting with a model or cloud service.

Mass customization means producing offerings that fit individual customers while retaining the efficiency of shared operations. The business starts with reusable components, supported choices, and repeatable processes rather than designing everything from scratch. Components can be physical modules, software features, or approved learning materials.

Servitization shifts the offering from a product alone toward an ongoing service that helps the customer achieve a useful result. A compressor manufacturer might sell equipment with a maintenance service, or provide compressed-air availability under a service agreement. The customer values reliable production, not simply owning a compressor.

AI experimentation means trying an approach and checking whether its results are useful. On premises, a team may need to buy and install hardware before it can begin. With the AWS Cloud, teams can obtain computing and storage on demand, reducing the wait between an idea and a working experiment.

Amazon Q Business has been in maintenance since June 30, 2026. It is closed to new customers; existing customers can continue using it. This lesson describes the service as AWS offered it, not as a recommendation for new workloads. AWS's migration guide is available at https://docs.aws.amazon.com/amazonq/latest/qbusiness-ug/qbusiness-availability-change.html.

A data network effect is a reinforcing relationship between participation and product value. More people use a product, their activity supplies useful signals, and learning from those signals improves the product for other people or future interactions. Better outcomes can then attract new users or encourage existing users to return.

The Technology Adoption Lifecycle describes why organizations adopt a technology at different times. Innovators tolerate substantial uncertainty to explore what is possible. Early adopters accept implementation risk to pursue strategic advantage. Both invest before extensive proof is available, but exploration and business advantage are different motivations.

Cost-benefit analysis helps decide how much to invest in AI and which investments to make first. Start with a business objective, such as reducing processing delays, and a baseline, such as today's workflow. Over a shared evaluation period, expected net benefit is estimated additional benefit minus additional cost. Include implementation and ongoing work, not just model charges.

Data minimization means using only the information needed for a specific business purpose. It is not simply collecting fewer records. It includes choosing which fields, documents, and events are necessary, how detailed they must be, who receives them, and how long they are kept.

Amazon Macie is a fully managed service that helps discover sensitive data in Amazon S3. It uses machine learning and pattern matching to identify information such as personal identifiers and financial data in supported objects.

Algorithmic fairness concerns whether an AI-supported decision treats people and groups equitably in its intended context. High overall accuracy does not establish fairness: a model can be accurate for most people while repeatedly disadvantaging one population.

Model explainability helps people understand why a model produced an output and which inputs influence its behavior. A prediction such as “high fraud risk” states the result; an explanation identifies factors behind that result. This helps people investigate decisions rather than simply accept a score.

AI transparency means giving people relevant, understandable information about an AI system. A useful disclosure answers four questions: where is AI used, what is it for, what are its important limitations, and who is responsible for helping when something goes wrong?

AI transparency means providing relevant, understandable information about an AI system and how it is operated and governed. It helps people know where AI is used, what it is intended to do, what its limitations are, and who is responsible for it. Governance means the rules, responsibilities, and oversight surrounding its use.

Model robustness is the ability to maintain acceptable behavior when inputs differ from those used during development. Real inputs include spelling mistakes, background noise, missing values, and unfamiliar formats. High accuracy on clean examples alone does not show that a model handles these variations.

Response policies define what an AI application may return. A model's generated answer is a candidate response, not an automatically approved answer. Instructions such as “avoid harmful content” help guide generation, but the application must enforce its release rules outside the model.

Model robustness concerns whether a model maintains acceptable behavior when inputs or operating conditions change. For example, a support assistant should understand a request despite a small spelling mistake. Robust behavior does not mean giving identical answers to everything: when the facts change, the answer may need to change too.

Risk appetite is the risk an organization is willing to take or retain to achieve business objectives. Leadership approves that direction; teams translate it into safeguards and release decisions. It is not a target amount of harm to cause.

The interpretability-performance tradeoff is the choice between how well a model serves its task and how readily people can understand its behavior. A small decision tree can expose the rules behind a prediction. A large ensemble or neural network is usually harder to inspect directly, but complexity does not guarantee better predictions. Measure the difference rather than assuming it exists.

Selective disclosure means choosing what information to reveal, to whom, at what level of detail, and for what purpose. In responsible AI, the goal is meaningful transparency: enough accurate information to understand, question, and oversee a system, without unnecessarily exposing protected information. It does not mean publishing everything or keeping everything secret.

An Algorithmic Impact Assessment examines the consequences of using an algorithmic system for a particular purpose. It asks more than whether a model is accurate: who benefits, who could be harmed, and whether the proposed use is justified. The same model can have different consequences when recommending entertainment and when influencing access to employment.

Privacy by Design means treating privacy as a requirement from the start, not as a final inspection. For an AI project, this starts during business-case development and project planning, before choosing datasets, models, cloud services, or deployment patterns. Privacy requirements help determine whether the proposed benefit can be achieved acceptably.

A governance gate is a decision point where an AI project must meet defined requirements before moving forward. It is more than a progress meeting: reviewers use evidence to decide whether work can proceed, needs changes, or should stop. Organizations can use different gate names, but the purpose is to prevent unresolved risks from silently becoming production problems.

Human-in-the-Loop means people deliberately participate in an AI workflow instead of treating its output as automatically authoritative. People can review recommendations before action, evaluate generated responses, or audit completed work. The aim is not simply to add an approval click: it is to give people a real opportunity to influence the outcome.

A hallucination occurs when generative AI produces information that is false, invented, or unsupported by the evidence it should use. An assistant might invent an API parameter or claim that a policy promises a refund when the policy contains no such promise. One response can mix supported facts with unsupported claims.

A confidence score expresses how strongly a system supports a prediction, extracted value, or response. Its meaning depends on its source. A classification score, a document similarity score, and an answer-support score measure different things. Before using a score, document what it measures and what decision it will inform.

Amazon Augmented AI (Amazon A2I) helps applications add human review to machine learning predictions. The model produces a prediction; people review selected cases before the application relies on them. This is a human oversight safeguard, not a replacement for the model.

Decision rights describe who is authorized to make a particular decision and under what conditions. They answer questions such as who may approve an AI release, what evidence is required, and which approvals must be obtained.

AI governance defines who may decide what, which limits apply, and who must answer for outcomes. A cross-functional governance body brings together business leads, technical experts, security, legal, and compliance. Each contributes a different view: business value, technical feasibility, security threats, legal obligations, or evidence of compliance. Together, they can address choices that no single specialty can settle alone.

A RACI matrix makes cross-functional ownership visible. Its rows describe activities or deliverables, and its columns name organizational roles. Each relevant cell shows how that role participates, so an AI initiative does not rely on everyone assuming someone else owns the work.

A RACI matrix connects activities to organizational roles. Responsible roles do the work; the Accountable role owns the outcome. Consulted roles provide input, and Informed roles receive updates. To reveal capability gaps, look beyond the letters: can the assigned roles actually fulfill their duties?

Separation of Duties assigns incompatible responsibilities to independent actors. Responsibilities are incompatible when combining them lets someone judge their own work or bypass review. For an AI initiative, the goal is to prevent one person or team from controlling the proposal, approval, and oversight without independent challenge.

Compliance means meeting applicable requirements and being able to demonstrate that you meet them. For an AI workload, examine both the data it handles and what the business does with its outputs. Summarizing public documents raises different questions from using personal information to support hiring or lending decisions.

AWS Artifact provides access to AWS security and compliance documents, including SOC reports, ISO certifications, and PCI reports. These documents help you evaluate AWS controls: the safeguards and processes AWS uses to meet specified requirements. They are evidence for a compliance review, not approval of your AI application.

The General Data Protection Regulation (GDPR) sets requirements for processing personal data: information relating to an identified or identifiable person. AI prompts, customer records, device identifiers, and identifiable conversations can all contain personal data. GDPR can apply outside the EU when organizations offer goods or services to people in the Union or monitor their behavior there; citizenship and server location alone do not determine applicability.

A Privacy Impact Assessment (PIA) examines how collecting, using, sharing, retaining, and deleting personal data could affect people. For AI, it asks whether the proposed use is justified and whether a less intrusive approach would work. Begin during business-case development and design, before deployment makes changes expensive.

AWS Audit Manager has been in maintenance since March 31, 2026 (2026-03-31). It is closed to new customers; existing customers can keep using it. This lesson explains the service as AWS offered it, rather than recommending it for new workloads. AWS's migration guide is at https://docs.aws.amazon.com/audit-manager/latest/userguide/audit-manager-availability-change.html.

Access control decides whether a requester may perform an action on a resource under the current conditions. In an AI system, the requester might be a person reviewing model results or a workload reading training data.

Least privilege means giving an identity enough access to do its intended work, without unrelated capabilities. Authentication establishes who is making a request; authorization determines what that identity may do. An authenticated AI application should still be refused access to data it does not need.

An IAM permissions policy is a document that defines allowed or denied actions on AWS resources. An identity-based policy attaches to an IAM user, group, or role. When someone uses a role, its permissions help determine what they can do. Signing in identifies the caller; a permissions policy helps decide whether a requested action is authorized.

AWS Identity and Access Management (IAM) controls access to AWS actions and resources. An identity is who makes a request, such as an IAM user or an assumed role. IAM permissions policies can grant that identity permission to perform particular actions. For example, a team's role can receive permission to invoke an AI model.

AI systems handle data such as training files, prompts, model files, and generated responses. Encryption converts readable data, called plaintext, into an unreadable form called ciphertext. Decryption recovers the plaintext using the appropriate cryptographic key.

Encryption changes readable data, called plaintext, into unreadable ciphertext. Decryption uses the appropriate key to recover the plaintext. It answers the question: can someone understand these bytes without the key?

Encryption at rest protects data while it is stored, such as documents or AI training files in Amazon S3. Encryption changes readable data, called plaintext, into unreadable ciphertext. Decryption makes it readable again using the appropriate key.

A risk matrix helps people prioritize harmful events by comparing two things: likelihood, how likely the event is within a stated period, and consequence, how serious the harm would be if it occurred. It supports judgment; it does not predict exactly what will happen.

The NIST AI Risk Management Framework (AI RMF) helps organizations identify, assess, and respond to AI risks. Its four functions—GOVERN, MAP, MEASURE, and MANAGE—connect organizational expectations to decisions about particular AI uses. They are connected, iterative activities, not four approval stages completed once.

The EU AI Act sets legal requirements for AI based on risks to health, safety, and fundamental rights. Classification starts with the application’s intended purpose: what it is designed to do, as described in its instructions, documentation, and marketing. A model’s size, generative capabilities, or cloud service does not establish the application’s classification.

The NIST AI Risk Management Framework (AI RMF) is voluntary guidance for managing AI risks. The EU AI Act is legislation that creates enforceable requirements for organizations within its scope. Both address AI risks, but they have different authority: guidance helps structure decisions; law establishes requirements that apply regardless of an organization's preferences.

A risk matrix helps an organization prioritize governance work. It compares the likelihood of a particular harm with the severity of its consequences. An internal rating can guide where to spend extra testing effort, require senior review, or strengthen monitoring.

Audit logging records security-relevant activity so people can review what happened and who was responsible. Unlike a performance metric such as response time, an audit record describes an action and its context. For production AI, accountability can require evidence about both changes to the system and its use.

Algorithmic bias is systematic unfairness in an automated system’s behavior or outcomes. It is not simply any prediction error, and it is different from the statistical bias in the bias–variance trade-off. A system can be accurate overall yet consistently serve one population worse than others.

Subgroup analysis measures a system’s behavior separately for meaningful groups, rather than relying only on an aggregate, or overall, score. Larger groups contribute more to an overall accuracy score, so good performance for them can hide poor performance for a smaller group. In the example below, 96% overall accuracy hides much lower accuracy for Language B.

Intellectual property (IP) rights protect certain intangible assets. Copyright protects original expression, patents protect qualifying inventions, trademarks protect source-identifying names and symbols, and trade secret law protects qualifying confidential information. Each has different requirements and limits.

Data provenance is the documented history and context of data: where it came from, how it was collected, which permissions apply, and what happened before it reached a model. It gives investigators evidence to consult when harmful material or an intellectual property concern is reported.

Amazon Rekognition is an AWS service that analyzes images and videos. Its content moderation features detect potentially inappropriate visual content, such as nudity or violence. This helps reviewers find material that may need attention without manually inspecting every upload first.

Change readiness asks whether an organization can implement and sustain a specific change. It is not the same as maturity, which describes capabilities established over time. A team might be ready for an internal AI assistant that drafts answers for employee review, but not for an assistant that sends answers directly to customers.

An executive sponsor is a senior leader accountable for an initiative’s business success, with authority to commit resources and resolve competing priorities. For AI, sponsorship means more than supporting model development. It means aligning affected leaders around a useful business outcome and preparing the organization to achieve it safely.

Decision rights define who may decide an issue, under what conditions, and with whose input. For an AI initiative, funding, data access, production release, and suspension are different decisions. Assign a named accountable decision owner to each: someone answerable for ensuring the decision is made within the agreed rules. This prevents decisions from drifting between teams or being made by whoever happens to have technical access.

A data strategy describes how an organization intends to use data to achieve its goals. For example, improving demand forecasts requires reliable sales records that AI teams can find and understand. The strategy gives that work a purpose; it does not, by itself, authorize every use of those records.

A proof of concept (POC) tests a specific proposition under defined constraints. Instead of asking whether AI can transform customer support, ask whether an AI assistant can draft useful answers from approved documentation. The purpose is to gather evidence for a decision, not simply to produce an impressive demonstration.

A Capability Maturity Model describes how an organizational capability develops from informal, person-dependent work into a managed, improving practice. It helps establish a baseline, choose a suitable target state, and identify gaps. It assesses how reliably an organization works, not just whether it owns AI tools.

Gap analysis compares what an organization must be able to do with what it can reliably do today. A capability gap is a specific shortfall between those two states. The target comes from a business use case, not a general ambition to be excellent at every aspect of AI.

A skills matrix places people, teams, or roles against the competencies needed for their work. It records required proficiency, current demonstrated proficiency, supporting evidence, and an assessment date. Its purpose is to reveal workforce gaps that need action, not simply list qualifications.

Weighted scoring is a way to compare investments when no single measure captures everything that matters. A team chooses evaluation criteria, gives each criterion a weight, and scores each proposal. The combined result makes the team's priorities visible; it does not prove which proposal deserves funding.

A data silo exists when data cannot be readily found, accessed with permission, understood, or combined across team or system boundaries. A department can have accurate records that remain unusable to other teams. Separate storage is not itself a silo: distributed datasets can work together through documented interfaces, shared descriptions, and governed access.

A data silo forms when one team's data is difficult for other teams to discover or understand. Amazon DataZone provides a shared business catalog where producers publish descriptions of data assets and consumers search for useful data. It helps teams share knowledge about data without requiring all the underlying data to move into one store.

Teams preparing data for AI need to answer two different questions: Which dataset should we use? and What data are we allowed to read? Amazon DataZone helps people discover and request data. AWS Lake Formation governs permissions on data lake resources.

A data strategy sets goals such as trustworthy data and responsible sharing. A data contract turns those goals into explicit expectations between a producer, which supplies data, and consumers, which use it. It describes what the data means, what users can expect, and who acts when something goes wrong.

A catalog holds metadata: information that describes data. Amazon DataZone provides a business-facing catalog where producers publish data assets and consumers discover them, understand their meaning, and request access. It helps teams share data across organizational boundaries instead of relying on personal contacts to find it.

A data lake holds datasets that teams can reuse for analysis and AI. Reuse creates an ownership question: who decides which people and workloads may read each dataset? A data owner is responsible for those decisions, including the permitted purpose and scope of access.

AWS Clean Rooms helps organizations analyze their combined datasets without giving one another direct access to the source records. For example, partners can measure how their customer populations overlap without exchanging unrestricted copies of their customer files.

Compute executes the instructions behind an AI workload. Choosing capacity starts with understanding the work, not choosing the largest machine. Training uses examples to adjust a model's learned parameters. Inference uses a trained model to produce predictions or generated output from new input.

An Amazon EC2 instance is a virtual server. Its instance type determines its mix of computing resources. For AI workloads, important resources include CPU capacity, system memory, and any accelerators. An instance family emphasizes a particular resource mix; the size identifies a capacity option within that family.

Object, block, and file storage are different ways to present data to an application. Choose the interface the application needs, not just enough space for its data. AI projects often use more than one model: a collection of training images and a shared working directory have different access needs.

An AI application has several network paths: data moves to the workload, users send requests, and results return. Each path needs its own requirements. A private data path and a public user-facing endpoint can belong to the same application.

Amazon VPC gives you a logically isolated virtual network in an AWS Region. You choose its IP address ranges and divide them into subnets. Each subnet belongs to one Availability Zone; a VPC can contain subnets across multiple Availability Zones.

Stakeholder mapping identifies who can affect an initiative, who will be affected by it, and what they need to participate. For a cloud-based AI initiative, this includes leaders, engineers, security and data teams, and the people whose daily work will change. Keep their profiles in a maintained stakeholder register, rather than treating a list of names as the finished map.

A guiding coalition is a group with the authority, expertise, and credibility to lead change across an organization. Commonly associated with Kotter’s change-management approach, it connects executive decisions to everyday work. For enterprise AI adoption, it is more than a committee that receives project updates: it makes decisions and helps remove obstacles.

A Community of Practice brings people together around a shared area of work, regular interaction, and practical knowledge they build together. An AI community helps colleagues discuss real problems, learn from attempts, and share useful practices across teams. It is more than a series of presentations.

A Center of Excellence (CoE) is an organizational capability, not a cloud service or just a technology team. It brings specialists together so AI initiatives connect business goals, sound engineering, and risk management. Shared guidance and reusable resources help teams avoid solving the same problems separately.

A Change Impact Assessment is a structured review of what will change, who will be affected, and what support they need. For AI adoption, installing a tool is only part of the change. Employees may gain new review duties, managers may need different performance expectations, and technical teams may take on ongoing monitoring work.

The ADKAR Model, developed by Prosci, describes five outcomes individuals need to adopt and sustain change: Awareness, Desire, Knowledge, Ability, and Reinforcement. Installing an AI tool does not create these outcomes automatically. Leaders use the model to ask what someone needs, rather than treating all hesitation as resistance.

Psychological safety is a team's shared expectation that people can ask questions, admit uncertainty, report mistakes, and challenge ideas without ridicule or retaliation. During AI adoption, someone might say, “I do not understand this evaluation” or “This assistant gives unsafe advice.” Safety means a respectful, fair response—not a promise that every suggestion will be accepted.

A growth mindset treats capability as developable rather than fixed. Someone who cannot yet check whether an AI answer is reliable can learn through instruction, practice, feedback, and collaboration. Current performance is a starting point, not a permanent label.

Status quo bias means giving the current arrangement an advantage simply because it already exists. A team might reject AI-assisted ticket classification because manual sorting feels familiar, without checking how well either approach works. Familiarity is not evidence that a workflow is safer or better.

Role-based learning starts with the work people must perform, not with a course catalogue. Identify each role’s tasks, decisions, and risks, then compare current skills with the skills needed. Define an observable outcome: something a learner can demonstrate, such as checking a generated answer against an approved source.

Bias means systematic patterns in AI data, predictions, or decisions that can disadvantage people or groups. Fairness concerns equitable treatment under explicitly chosen criteria. A difference between groups is a reason to investigate; deciding whether it is unfair also requires understanding the task and its consequences.

Human-centered design starts with what people need to accomplish, not with what an AI model can generate. Work with intended users, domain experts, and people affected by decisions to understand their goals and constraints. For a customer-support team, the goal might be resolving problems correctly and respectfully—not simply closing more requests.

Task decomposition means starting with a business outcome and dividing the work into manageable tasks. For example, “resolve a customer complaint” includes gathering records, understanding the complaint, checking policy, preparing a response, and approving an action. These tasks need different capabilities, even when one employee currently performs them all.

Automation bias is the tendency to rely on an automated result instead of checking whether it is right. It can happen with dashboards and rules engines as well as AI. A history of good results, fluent explanations, or precise-looking scores can encourage trust without proving that the current result is correct.

The AWS Cloud Adoption Framework (AWS CAF) guides organizational transformation, not just the movement of applications to the cloud. Its journey has four phases: Envision, Align, Launch, and Scale. These connect business goals with the capabilities needed to achieve them. CAF is guidance, not a deployable AWS service.

A Minimum Viable Product (MVP) is a deliberately limited product that lets intended users complete a meaningful task while testing a business assumption. Its purpose is validated learning: using observed results to find out whether an idea creates value. Minimum means narrow scope and limited investment; viable means useful, safe, and dependable enough for the intended use.

Experiment tracking preserves evidence about what a team tried, what happened, and why it chose the next step. That evidence connects envision, experiment, launch, and scale. These phases support iteration, not an irreversible path toward deployment.

Benefits realization means checking whether a delivered capability actually improves the business and whether that improvement lasts. A business case forecasts value; benefits realization checks what actually happens. For AI-assisted customer support, the aim might be faster resolution without worse service, rather than simply deploying an assistant.

Phased rollout introduces AI in bounded stages instead of everywhere at once. Each stage has a defined task, user group, accountable owner, and decision date. The aim is to grow proven business value while preparing people, data, and operations for wider use—not simply to increase AI traffic.

The hub-and-spoke model shares responsibility between a central enabling team and teams in business units. For enterprise AI, the hub is often an AI center of excellence (COE). The spokes are department or product teams that understand their users, data, and business needs. This is an operating model, not a requirement for a particular network layout.

A feedback loop uses evidence about an AI system’s results to guide changes, then checks whether those changes helped. Collecting ratings is only the beginning. The loop needs accountable review, a controlled change when justified, and follow-up measurement.

AWS CloudTrail records AWS account activity so reviewers can investigate who did what, when, and from where. This includes supported actions made through the AWS Management Console, command line tools, SDKs, and AWS services.

An AI application can change when developers update application code, prompts, model settings, or dependencies. A release pipeline turns those changes into a defined sequence of automated checks and release steps, rather than relying on someone to repeat a set of manual instructions.

An operational readiness review (ORR) is a decision process for checking whether a workload can be operated responsibly in production. An AI pilot shows that an idea can work, often with curated data, few users, and close supervision. Production approval requires evidence that the service can handle its intended use, failures, and ongoing responsibilities.

Resiliency means handling disruption without an unacceptable loss of service or data. For an AI service, this can mean keeping predictions available when infrastructure fails. A failure domain is a boundary within which resources can be affected by the same failure.

AWS Resilience Hub helps you assess whether a workload's architecture can meet your recovery goals after a disruption. Its purpose is to identify weaknesses before an outage threatens business continuity: keeping important business activities running or restoring them within acceptable limits.

High performance means completing work within the response-time and processing-capacity targets the application needs. For an AI assistant, users care about waiting for answers. For a batch prediction job, the goal might be finishing all records before a deadline.