Task 4.2: Establish data and infrastructure foundations for AI.

• AA separate copy of the data for each AI project team, refreshed by hand

  Incorrect

  Ad hoc copies multiply the silos, add security exposure and leave each project to repeat the same assembly work.

• BA more advanced prediction model that tolerates missing inputs

  Incorrect

  A model that tolerates gaps does not give the organization access to the data it needs; the bottleneck is fragmentation.

• CRestricting AI projects to one department's data

  Incorrect

  Limiting projects to single silos avoids the problem rather than solving it, and discharge prediction needs data from several departments.

• DUnified, governed access to data across the seven systems

  Correct

  AWS's generative AI guidance calls for breaking down data silos with a unified data system and unified access to data, independent of where it resides, while maintaining security and compliance.

Fragmented data environments and unclear ownership slow AI down. A unified, well-governed way to access data across systems lets many AI initiatives draw on the same foundation instead of each rebuilding its own extracts.

• AA rule that the AI team fixes any data errors it finds

  Incorrect

  The AI team consumes the data but does not control how it is created. Fixing errors downstream leaves their cause in place.

• BA data lake that copies all of the product data into one central location

  Incorrect

  Centralizing storage can improve access, but it does not decide who is accountable for the data's correctness.

• CClear data ownership, with roles and standards for each data domain

  Correct

  Data governance determines roles, responsibilities and standards for data usage, including who can take what action on which data. Clear ownership gives each data domain someone accountable for its quality.

• DAn annual data quality audit performed by an external firm

  Incorrect

  An audit can measure quality, but without owners there is still no one responsible for correcting and preventing errors.

AI depends on data that someone is accountable for. A governance model that assigns owners and defines roles, responsibilities and standards turns recurring disputes into clear accountability for data quality.

• AA complete production platform with automated pipelines for every future use case

  Incorrect

  Production-grade pipelines come when solutions move to production. For proofs of concept AWS advises a minimal platform built on managed services.

• BA self-managed GPU cluster to train the airline's own foundation model

  Incorrect

  Pretraining a foundation model is the costliest option and is not needed to run proofs of concept on existing models.

• CNo foundation, letting each team use its own personal accounts and tools

  Incorrect

  Unmanaged personal accounts bypass security and governance and create shadow AI from the start.

• DA secure landing zone, developer permissions and a managed experimentation space

  Correct

  AWS's maturity guidance lists, for the Experiment level, setting up foundation infrastructure such as a landing zone and the permissions developers need, and an environment for experimentation such as an Amazon Bedrock playground.

The infrastructure should match the stage. Proofs of concept need a secure, governed place to experiment quickly with managed services; production platforms and custom training infrastructure come later, when a use case has proven its value.

• AEach team's own data handling now, standardized after the experiments

  Incorrect

  Results built on differently collected and labeled data cannot be compared, and AWS places a common approach to data before the experiments, at the Envision stage.

• BPublic datasets in place of the bank's own data for the experiments

  Incorrect

  Public data does not reflect the bank's own customers and processes, so results would say little about its use cases, and it leaves the teams without a common way to prepare the bank's data.

• CA structured approach to collecting, cleaning and labeling data

  Correct

  AWS's data strategy guidance for the Envision stage is to define a structured approach to collecting, cleaning and labeling data for generative AI experiments.

• DCleaning and labeling of every dataset in the bank before any experiment

  Incorrect

  Preparing all the bank's data first would delay the experiments; the guidance is a defined approach for the data the experiments need.

Data foundations start with consistent practices. A shared approach to collecting, cleaning and labeling data makes experiments comparable and their data reusable, without waiting for every dataset to be perfect.

• AMore analysts to check data by hand before each project

  Incorrect

  Manual checks do not scale with growing data and still miss errors.

• BAutomated validation and cleansing, monitored continuously

  Correct

  AWS's generative AI data guidance calls for automated data validation, cleansing processes and continuous monitoring to maintain data integrity at scale, under a data governance framework.

• CAccepting data errors, because models will learn around them

  Incorrect

  Poor data leads to poor AI outcomes; models do not reliably correct bad inputs.

• DA one-time cleanup of all current data with no ongoing checks

  Incorrect

  Quality degrades again without continuous validation.

Data quality is the bedrock of effective AI. Building validation and monitoring into pipelines keeps quality consistent as data and AI use grow.

• AFederated data platforms serving every business unit at enterprise scale

  Incorrect

  Federated enterprise-scale data platforms are a Scale-stage investment.

• BA custom-trained foundation model on all company data

  Incorrect

  Pretraining a model is unnecessary for RAG experiments on existing models.

• CEarly data pipelines for both data types, plus vector databases for RAG

  Correct

  For the Experiment stage, AWS lists implementing early-stage data integration, including structured and unstructured data pipelines, and setting up vector databases for RAG experiments.

• DNo technology work, because experiments run on spreadsheets

  Incorrect

  RAG experiments need data pipelines and a vector store, not spreadsheets.

Infrastructure should match the stage. RAG experiments need pipelines for the relevant data and a vector store; enterprise-scale platforms come later.

• AElastic infrastructure that scales with AI data workloads

  Correct

  AWS's guidance lists scalable, elastic data infrastructure that can handle varying AI workloads and data volumes among the key success factors for generative AI data architecture.

• BFixed capacity sized to this year's average data load

  Incorrect

  Fixed capacity sized to an average cannot absorb unpredictable spikes.

• CBuying hardware for the largest imaginable future load now

  Incorrect

  Over-buying ties up capital in capacity that may sit idle.

• DLimiting AI projects to what current storage can hold

  Incorrect

  Constraining projects to old capacity limits the value AI can deliver.

AI workloads are variable and growing. Elastic infrastructure lets capacity follow demand instead of forcing a choice between shortages and idle hardware.

• AA separate personalization model trained on each system's data

  Incorrect

  Each model would see only part of a shopper's history, which keeps the silo problem in the models instead of solving it in the data.

• BUse of the loyalty data alone, as the largest of the three sources

  Incorrect

  Dropping the online and in-store records leaves the incomplete purchase history the retailer is trying to fix; AWS guidance is to break down data silos, not to work around them.

• CReliance on the personalization model to learn which profiles match

  Incorrect

  AWS treats accurate, complete and consistent data across sources as the bedrock of effective AI; duplicate profiles fed to a model become poor inputs, not something it reliably corrects.

• DMatching and linking the records into one view per customer

  Correct

  AWS Entity Resolution matches, links and enhances related records stored across applications and data stores to build a unified view of customer interactions; copying the systems into one place without linking keeps the duplicates.

Breaking down data silos is more than moving data into one store. When each system identifies customers differently, records must be matched and linked before AI can see one customer's full history.