Establish data and infrastructure foundations for AI
Three foundations, worked backward from the initiative
An AI initiative depends on its data: the AWS Well-Architected Generative AI Lens states that the value driven from generative AI applications depends on the ability to use both structured and unstructured data[1]. AWS notes that data silos exist in software applications, within different departments, and in rapidly growing environments[2], and that their isolation can lead to duplicates, inaccuracies, increased manual processes, and business performance issues[2]. Silos, and the duplicates and inaccuracies they cause, are not problems with the AI model; each is a gap in the foundations the initiative rests on.
This topic covers those foundations, judged for one AI initiative at a time. AWS data governance guidance recommends that you attach data work to already-funded business initiatives and understand which data domains, sources, and elements are needed to support them[3], and the AWS generative AI workload assessment asks which data types are crucial for your generative AI workloads and what percentage of them is currently accessible[4]. Working backward from the initiative in this way produces requirements that fall into three groups, which this topic calls the three foundations, as the figure below shows:
- Data readiness: is the data fit for this use case, and can the AI reach it? Fitness covers the data quality dimensions (timeliness as refresh requirements), relevance and representation, and quality controls at the source; together, these are the fitness checks a dataset must pass for one use case. Reach covers accessibility across data silos, including unstructured information.
- Data strategy and governance: who decides about the data, and how is it found and shared? This covers ownership and stewardship, governance style, shared business definitions, data catalogs, data sharing frameworks, lineage, and lifecycle rules.
- Technology foundation: can the platform carry, connect, and protect the workload? This covers workload demand, integration with business systems, the choice of managed AI platform, and protection and observability.
These groups match the three skills the AWS AIB-C01 exam guide lists under Task 4.2: Establish data and infrastructure foundations for AI[5]: data readiness; data strategies, ownership, and sharing frameworks; and technology and infrastructure requirements.
Related tasks in the AWS AIB-C01 exam guide
This topic focuses on data and infrastructure foundations. The AWS AIB-C01 exam guide lists 'Task 4.1: Assess AI business readiness and maturity' beside Task 4.2; its skills assess business readiness to adopt AI across critical dimensions (for example, leadership alignment, data quality, cultural preparedness, technical infrastructure, governance frameworks) and apply AI maturity models[5], so data and infrastructure are two dimensions among several there; see Assess AI business readiness and maturity. Your role as a business strategist is to assess these foundations and state requirements; building data pipelines and configuring infrastructure stays with technical teams.
The takeaway: judge every foundation against a named initiative, checking data readiness (reach and fitness), data strategy and governance, and the technology foundation.
Data quality dimensions as acceptance criteria
This section turns the data quality dimensions into acceptance criteria for a data-readiness scorecard: a short list of measures, each with a target, that decides whether a dataset is ready for one use. Accuracy, completeness, consistency, uniqueness, timeliness, and validity are commonly referred to as the six data quality dimensions[6], and AWS Well-Architected guidance builds an overall data quality score by individually assessing each facet, then aggregating and normalizing them into a single metric[7].
| Dimension | The question it asks | Example acceptance measure |
|---|---|---|
| Accuracy | Are the values correct? | Share of sampled records confirmed against the authoritative source |
| Completeness | Is the required information present? | Share of records with every required field filled |
| Consistency | Do records and systems represent the same fact the same way? | Share of shared fields that agree across systems |
| Timeliness | Is the data current enough for this use? | Age of the latest update against the agreed limit |
| Uniqueness | Is each business entity or event recorded only once? | Duplicate rate |
| Validity | Do values meet the required formats and business rules? | Share of values passing validity rules |
The right-hand column shows illustrative measures, not a standard list; choose the measure and target that fit the business process.
Each dimension catches a different defect
Strength in one dimension does not establish strength in another. A value can pass a format rule (valid) and still be factually wrong (inaccurate), and an accurate customer record says nothing about whether the same customer also appears twice. A required-field completion rate close to 100 percent therefore does not show readiness while duplicate customers and invalid identifiers remain. AWS guidance for small and medium businesses recommends starting with uniqueness (no duplicates), completeness (critical fields present), and timeliness (data current enough to act on)[8], and notes that duplicate leads and mismatched customer records can inflate audience size[8]: duplicates inflate counts and give repeated evidence undue weight even when each copy is correct.
Volume is not a quality dimension. AWS guidance on generative AI data states that volume alone is not enough; careful cleaning, filtering, and preprocessing are critical[9]. A high record-count target cannot serve as the acceptance criterion, because a large dataset can be complete and still inaccurate, or consistent and still out of date.
Choose the dimensions the process relies on
The AWS generative AI workload assessment asks how you ensure and measure the quality of your data[4] and lists completeness, accuracy, consistency, timeliness, and relevance[10] among typical data quality concerns. Relevance is not one of the six dimensions above: it asks whether data aligns with the specific generative AI task[10], so it is assessed with representation as a separate fitness check, described next. Set targets only for the dimensions the business process relies on, and record why each target matters for the intended use.
The takeaway: write one acceptance measure per relevant dimension, and never let record count stand in for any of them.
Relevance and representation of the target population
Data readiness also asks whether the data describes the right people and situations. A dataset can pass all six quality dimensions and still be the wrong foundation, because it was collected from a different population, region, channel, or period than the one the AI initiative will serve.
- Relevance asks whether the data relates to the specific task. AWS guidance for retrieval pipelines says to collect high-quality, relevant data and make sure that the dataset is relevant to the application's domain[9], and the workload assessment lists relevance: confirming that data aligns with the specific generative AI task[10] as a data quality concern.
- Representation asks whether the data covers the intended users, operating conditions, and business situations. AWS guidance on fine-tuning data states that clean, representative, and appropriately sized datasets are essential[9], and the workload assessment asks how you ensure diversity in your datasets to prevent bias, for example by auditing datasets for representational bias[4].
For example, suppose a lender's historical loan records come almost entirely from urban branches and it now plans an AI-assisted service for rural applicants; the workload assessment's example answer to such gaps is actively sourcing data from diverse demographics and regularly auditing datasets for representational bias[4]. The records may be accurate, complete, and free of duplicates, yet they say little about the new applicants. Cleanliness does not establish relevance, so clean historical records are a candidate foundation, not a default representative dataset. Evaluate coverage of the target setting before accepting them. The AWS AIB-C01 exam guide lists recognizing that bias can occur at multiple stages of the AI lifecycle and the importance of ongoing monitoring for bias drift[11] under Task 3.3; see Identify Enterprise AI Risks and Direct Mitigation.
The takeaway: before accepting a dataset, ask whose situations it records and whether those are the situations the AI will meet.
Data freshness matched to the rate of business change
Freshness is the timeliness dimension from the table above applied to refresh planning: timeliness becomes a planning requirement once you ask how often each source must be refreshed. The rule: set the refresh expectation from how quickly the business information changes, not from whatever schedule a source already follows.
The AWS generative AI workload assessment asks how frequently you need to update or refresh data from each source[10], and its example answer mixes very different cadences: near real-time updates for transactional databases, daily batch updates for document repositories, and monthly or quarterly updates for public datasets. AWS data governance guidance likewise says to identify the applications the data must feed, including requirements for freshness and privacy[3].
A dynamic knowledge source and a stable reference collection need different refresh approaches. An assistant that answers questions about live company policies needs an update whenever a policy changes; an archive of retired product manuals may need almost none. Using the document archive's publication schedule for a live policy assistant is a mistake, because that schedule may not meet the assistant's currency requirement (how up to date its answers must be; currency in the sense of being current, not money). For retrieval-based assistants, AWS guidance notes that regular updates and maintenance of the data repository are essential to preserve the accuracy and relevance of the information[9].
Freshness also decays after launch: data that suited the pilot when it started can fall out of date later. Readiness requirements should therefore include the update frequency and ongoing quality monitoring, not only a one-time check; AWS lists performing one-time data quality checks without continuous monitoring[12] as a common anti-pattern.
The takeaway: match update frequency to the use case's need for current information, and keep monitoring it so the data stays usable.
Reliable inputs: quality controls at the source
This section covers where quality problems get fixed: where the data is created, not at the AI output. When several sites collect the same business information differently, the AI receives inconsistent inputs, and correcting each output centrally hides the defect without removing it.
The AWS Cloud Adoption Framework (AWS CAF) governance perspective tells organizations to identify root causes of data quality problems and improve relevant processes at the source[13] and to establish data quality standards: identify key quality attributes, business rules, metrics, and targets[13]. AWS guidance for small and medium businesses describes data quality management as a repeatable loop that prevents bad data from entering your systems, fixes issues that slip through, and makes quality visible[8], not a one-time cleanup project. Its loop has four working steps, which the figure below shows, followed by documenting what the critical fields mean:
- Find issues: profile the highest-impact dataset for duplicates, missing values, invalid formats, and stale records.
- Fix and standardize: correct what you find and standardize as you go, so you do not fix the same problem every month[8].
- Prevent at entry: add controls where data enters your business[8], such as required fields, dropdowns instead of free text, and format checks.
- Monitor with rules: track a small set of rules and assign someone to review exceptions[8].
Applied to the opening case of sites that collect the same business information differently: before expanding an AI initiative to every site, agree one definition of each shared business term, one collection practice, and one set of quality checks across sites, and give a named person responsibility for recurring defects. Consistent controls across locations make reliable inputs an ongoing capability instead of a one-time cleanup.
The takeaway: standardize definitions, collection, and checks where the data is created, and assign responsibility for recurring defects; central output correction leaves the source defect in place.
Data silos and their two kinds of cause
A data silo is a system or departmental collection whose data is difficult for other users or applications to access or share. AWS defines data silos as digital systems where data is difficult to share or access by other services[2] and notes that the isolation diminishes the value of that data and can lead to duplicates, inaccuracies, increased manual processes, and business performance issues[2]. For AI the cost is context: an application can use only the business information it can reach, and records copied by hand between systems must be reconciled.
The diagnostic question is what keeps the data out of reach. There are two kinds of cause, and the diagnosis decides the remedy:
- Technical cause. Legacy systems create data silos, as they are often not integrated into wider company technology stacks[2]. The remedy is to integrate or migrate systems: AWS notes that organizations might need to migrate the data to a centralized hub or implement middleware that connects the legacy systems[2].
- Organizational cause. Silos also form if business units don't actively share their data with others[2], and AWS lists departmental data ownership conflicts[2] among typical bottlenecks. The remedy is to agree ownership and sharing rules (the data sharing framework later in this topic) and to build a sharing culture, because an organization can only overcome data silos if its employees engage with these systems[2].
Technical connectivity alone does not resolve an organizational cause: a new connection does nothing if the owning department will not agree to share. AWS advises that an organization first look to the wider cause of why data silos occur[2] and then take action to unify disparate systems, implement data integration tools, and systematically break down data silos[2]; as the figure below shows, both remedies lead to unified, governed access.
Neither remedy involves the model. Funding model retraining to compensate for business records that remain inaccessible changes nothing about what the AI can reach, and adding processing capacity to a single-database pilot connects no additional source.
The takeaway: diagnose whether a silo is technical or organizational before choosing a remedy.
Unified, governed access across existing locations
This section covers the strategic answer to fragmented enterprise data: a unified data strategy that connects and governs the sources an AI initiative needs. Here a data strategy means the plan for which data the organization needs, in what condition and with what access; unified, governed access is the part of that plan that answers the data readiness reach question (can the AI reach the data it needs?). When AI must use information spread across disconnected systems, resolve the fragmentation once, as an enterprise foundation, rather than leaving each initiative to negotiate its own isolated workaround.
Unified does not mean a single physical repository. AWS describes centralized data management as removing data from isolation whether the data itself is co-located or spread across multiple sources[2], and the AWS Well-Architected Generative AI Lens says organizations must address data silos and enable unified access to data, independent of where it resides[1]. The same lens recommends a unified data system that integrates structured and unstructured data from across the organization[1]. Some data may be consolidated; much can stay where it is, provided access, formats, and governance are consistent.
A unified data strategy therefore resolves three kinds of fragmentation:
- Access: a governed way for authorized users and applications to reach each required source.
- Formats: consistent representations, so data from different sources can be processed together.
- Governance: consistent ownership, definitions, and rules of use, which the data strategy and governance sections below describe.
The takeaway: fund unified, governed access as a shared enterprise foundation; it can span existing locations.
Unstructured information in the data inventory
Structured data has a standardized format, or predefined schema[14], typically rows and columns; unstructured data is information with no set data model, or data that has not yet been ordered in a predefined way[14], such as documents, emails, images, and recordings. A data assessment that inventories only the reporting warehouse misses much of the business knowledge an AI application may need.
AWS guidance states that the value driven from generative AI applications depends on the ability to use both structured and unstructured data[1], and warns that identifying, classifying, and organizing unstructured data can lead to security risks, cost overheads, and issues with storage, interpretation, and compliance[1]. The workload assessment's list of data sources a generative AI workload may need includes internal structured databases, unstructured text data from documents, emails, and social media, and audio and video files[10]. The AWS CAF tells data curators to catalog key data products, including structured and unstructured data[13].
Include the relevant unstructured formats in every step of the data foundation:
- Discovery: does the inventory find the documents, messages, and media at all?
- Classification: which of them contain sensitive information?
- Quality assessment: are the documents current and authoritative, or superseded drafts?
- Authorized access planning: who may use them, and may the AI retrieve them for a given user?
Plan, too, how users and AI applications will find these files and understand their context, such as which policy version a document represents and who owns it.
The takeaway: an AI data assessment covers documents, messages, and media as well as tables, with the same discovery, classification, quality, and access steps.
A data strategy that works backward from initiatives
This section begins the data strategy and governance foundation, which the AWS AIB-C01 exam guide lists as the skill to describe the importance of data strategies, data ownership, and data sharing frameworks that provide the foundation for AI[5]. A data strategy is the plan for which data the organization needs, in what condition and with what access, to support its business priorities; unified, governed access, described above, is the part of that plan that answers the data readiness reach question (can the AI reach the data it needs?).
AWS describes the most common strategic challenge for data governance as aligning the program to business initiatives instead of proposing the value of data governance directly[3]; proposals such as easier data discovery, made in isolation, are solutions in search of a problem[3]. AWS recommends that you:
- Build a data governance roadmap that shows support for targeted business initiatives[3], then identify where initiatives need the same data.
- Understand what fit-for-purpose data looks like for each chosen business initiative[3].
The output is a data roadmap that names, for each targeted initiative, the data domains and sources it needs, the quality levels and access capabilities required, and a measure of whether the required information is fit for its intended purpose. Needs shared by several initiatives become shared investments. This makes the data foundation an enabler of business value rather than a collection program looking for a future purpose.
The success measure follows from the same logic. Total enterprise data collected is a poor measure, because collection volume does not show that any business priority is supported. Better measures track whether targeted initiatives have the data they need; the Generative AI Lens suggests tracking data quality scores, time-to-model deployment, AI application performance metrics, and business value generated from AI initiatives[1].
The takeaway: a data strategy earns its funding by enabling named business initiatives, so measure it by their data readiness, not by volume collected.
Data owners, data stewards, and technical teams
Data governance covers the processes and policies that ensure data is in the proper condition to support business initiatives and operations[3] and determines roles, responsibilities, and standards for data usage[3]. The AWS CAF governance perspective says to define and assign key roles, including data owners, stewards, and custodians[13]; custodians here are the technical teams that hold and operate the data's storage and systems.
| Role | What it does | Route these matters to it |
|---|---|---|
| Data owner | Makes policies about the data, including who should have access to it and under what circumstances, how to interpret and apply regulations, and key term definitions[3] | Access conditions, permitted uses, business definitions |
| Data steward | Comes from the business, implements the data governance process in projects, monitors compliance, and escalates issues[3] | Recurring quality or governance issues found in daily work |
| Data engineers and other technical teams | Select and implement the tools to secure data, integrate data from various sources, manage data quality, and find the right data[3] | The supporting mechanisms |
AWS adds that data owners are also responsible for the data sets' technical administration and access controls[3]. That makes the owner accountable for how controls are set; it does not make whoever runs the storage the owner. A storage administrator has technical custody of the data, and custody does not establish authority over permitted business use, so a storage administrator is not the default business data owner.
The working rule: route policy decisions to the responsible owner, and route recurring quality or governance issues through the stewardship process. Technical teams provide the mechanisms, but the business responsibilities must still be explicit.
The takeaway: owners decide, stewards apply and escalate, and technical teams operate; name each role for every data domain an AI initiative depends on.
Centralized, federated, and decentralized governance
Governance style decides where data decisions are made. AWS recommends that a data governance program balance centralization and decentralization (including self-service)[3] and that you empower domain teams as much as possible while maintaining coherence across domains (such as the ability to link data together)[3]. AWS describes three styles:
| Style | Where decisions sit | Fits when |
|---|---|---|
| Centralized | Central organizations are ultimately responsible for mission statements, policies, tool choices, and more; day-to-day actions are often pushed into lines of business[3] | Common control needs dominate |
| Federated | Individual business units operate in the way that best matches their needs, while a smaller centralized team solves problems that repeat frequently, such as enterprise-wide data quality tools[3] | Business domains need local expertise and autonomy |
| Self-serve or decentralized | Each department does what it needs for the specific project while aligning with centralized policies[3] | Project needs call for departmental choices within centralized policies |
The three styles differ in who makes the operating choices: in centralized governance, central organizations set policies and tools; in federated governance, business units choose how they operate, aided by a smaller central team; in self-serve governance, departments choose per project within centralized policies[3].
Federated governance is the style a data mesh uses. A data mesh is a distributed data management model built on domain ownership, data productization, self-service, and federated governance[15]; in it, leadership determines global standards and policies that you can apply across domains, while each domain keeps a large degree of autonomy, so long as data products have a standardized external interface[15]. Here a domain is a business area responsible for its own data[15]; a data product is a structured, reusable component[15] for other teams. The AWS CAF suggests that organizations consider adopting a federated (data mesh) approach to governance[13].
Federated does not mean independent. Incompatible departmental dictionaries and access rules, without shared standards, do not constitute effective federation, because incompatible rules undermine cross-domain use. In a data mesh, every domain publishes data products using the same shared schemas and metadata conventions[15].
The takeaway: federated governance gives domains autonomy under shared central policies, so information still links across domains.
Shared business meaning: glossaries and dictionaries
Connecting two datasets lets you read both; it does not make them mean the same thing. If sales counts a customer as active after any purchase in the past year while support counts a customer as active only while a contract is open, a combined report mixes two definitions without saying so.
Two tools give teams a consistent interpretation. A business glossary defines business terms (such as "active customer") and how they relate. A data dictionary describes each field; AWS guidance describes one that records what each field means, where it comes from, who owns it, and what good looks like[8]. The AWS CAF says to specify standards, including data dictionaries, taxonomies, and business glossaries[13] and to identify what datasets need to be referenced and model the relationships between reference data entities[13], where reference data means shared code lists such as product identifiers or reporting periods. In a data mesh, each data product is self-describing, with well-described syntax and semantics that follow standard naming conventions[15] and interoperable, adhering to a common set of governance standards[15].
Reconcile meanings such as customer status, product identifiers, and reporting periods before using information together. A common file format is not proof of interoperability: two files can share a format and still carry conflicting business meanings. Technical connectivity enables technical reachability, whereas shared semantics make the combined information interpretable and comparable.
The takeaway: before combining datasets for AI, agree what their key terms mean and record it in a glossary and data dictionary.
Data catalogs for discovery and context
A data catalog is an organized inventory of data assets and their metadata, where metadata means data about the data: ownership, business meaning, origin (provenance), and quality information. The AWS CAF describes using metadata to build an inventory of data products in a Data Catalog that helps data consumers quickly locate relevant data products as well as understand their context, such as provenance and quality[13]. Despite the capital letters, this means the general capability, an inventory of all the data that an organization collects and processes[16], not one product.
For AI teams, a catalog replaces rediscovery with discovery. AWS data governance guidance notes that with a centralized data catalog, data can be found easily, access can be requested, and data can be used to make business decisions[3]. The AWS CAF also says to identify and capture relevant technical and business metadata, including lineage[13] and to identify lead curators with responsibility for moderating the Data Catalog[13], so the catalog stays accurate as the data changes.
A catalog is not a governance program on its own. AWS states plainly that having a data catalog does not constitute a data governance program[3]: the listing does not supply the policies, stewardship, and quality processes that make the data trustworthy. Folder names are not a catalog either, because a storage location alone does not explain ownership, meaning, or fitness for use.
The takeaway: a catalog helps people find, understand, and request candidate data; it is one part of governance, not a substitute for it.
Data sharing frameworks and conditions of use
A data sharing framework is the set of rules that decides who may use which data, for what purpose, and under what protection and compliance conditions. Its aim is data democratization: making relevant data discoverable and accessible to the people and applications with a legitimate need. Democratization means easier authorized access, not unrestricted access.
Access and control as two levers
AWS describes two levers to make governance an enabler of innovation: access and control[3]. With too much control, the data gets locked up in silos and users are not able to access the data when they need it[3], which pushes people toward unapproved systems that IT teams do not know about (shadow IT[17]); with too much access, data risks becoming unregulated across applications and data stores, increasing unauthorized access risk[3]. The Generative AI Lens describes the access side as enabling self-service data access for AI teams through intuitive data catalogs, automated data discovery, and standardized APIs[1]. An organization-wide read permission is the wrong way to speed collaboration, because broad availability does not establish authorization for every user.
Conditions of use
Data governance outlines who can take what action, upon what data, using what methods, and in what situations[3]. For AI, assess privacy, confidentiality, consent, and applicable use restrictions alongside availability and quality; the workload assessment asks what measures you have to ensure data privacy, obtain proper consent, and maintain confidentiality[4]. Sensitive information, such as personally identifiable information (PII), which can be used to distinguish or trace an individual's identity[18], may need protection or exclusion: AWS guidance says to make sure that PII or other sensitive information is masked or anonymized before model training[19] and to apply data protection controls that provide an appropriate level of control for each class of data defined in your classification policy[20].
The access-request workflow
Clear access-request workflows and role-appropriate permissions, as the figure below shows, mean routine legitimate use does not depend on ad hoc negotiation. In a data mesh, each data product registers itself with a centralized data catalog for easy discoverability[15], and data consumers request access to the data products and seek approvals or changes directly from data owners[15]. AWS lists role-based access control systems, attribute-based access controls, and auditable data trails[2] among the controls that keep access to authorized personnel. For generative AI specifically, make sure that generative AI models can only retrieve information that the user is authorized to access[19].
Make ownership, approval, and ongoing responsibilities clear across organizational boundaries, so authorized access is paired with an authorized AI purpose. Moving the data is not the same as agreeing its use: a successful interdepartmental file transfer is not evidence of an adequate sharing agreement, because transport does not define permitted reuse.
The takeaway: a sharing framework makes legitimate access routine and keeps every use tied to an approved purpose and its conditions.
Data lineage and version records
Data lineage records where data originated and how it changed on the way to its current form, as the figure below shows; a version record identifies the particular state of a dataset that was used. Together they make AI data use traceable.
AWS Well-Architected guidance recommends a lineage tracker system that tracks changes for each release to enable reproducibility, speed up problem diagnosis, and improve regulatory adherence[21]. If a model starts producing unexpected results, such a system should let you trace back to the exact version of the model and the data it was trained on[21]. The Generative AI Lens lists versioning and lineage tracking for reproducibility[1] among key data considerations, and the workload assessment asks how comprehensive your documentation for data sources and lineage is, and how many versions of critical datasets you maintain and for how long[4].
Lineage and versions serve three useful purposes here:
- Impact assessment: when a source defect is found, identify which datasets and AI outputs used it.
- Investigation: trace a quality problem back through each transformation to where it started.
- Reproduction: rebuild the dataset state behind an earlier result.
Current data and historical evidence are different things. A dataset can be suitable for today's operational use without being evidence of what supported an earlier AI output, because it may have changed since; the version record, not the current copy, shows what the AI relied on.
A common shortcut fails here: recording the source department's name is ownership metadata, not lineage, because it describes neither the transformations nor the dataset state used.
The takeaway: keep lineage and version records for AI data so you can assess impact, investigate problems, and reproduce prior results.
Data lifecycle policies: retention and disposal
A data lifecycle policy sets how long data is retained, when it is updated, and when and how it is disposed of, in line with business needs and compliance obligations. The AWS CAF tells organizations to develop data lifecycle policies, and implement continuous compliance monitoring[13], and AWS notes that data governance frameworks cover rules that apply to internal data systems, including data collection, sharing, storage, deletion, and more[2].
AI creates new data assets that need these rules too: training and evaluation copies, retrieval collections, and the interactions themselves, since user-generated data includes user questions and prompts, the model's responses, and any explicit feedback[9]. Include these assets in the lifecycle policies so old copies do not become unmanaged exceptions. Unclear rules have a cost: AWS warns that incomplete records and unclear retention practices make it harder to respond to GDPR or CCPA requests, customer disputes, or audits[8]. GDPR is the European Union's General Data Protection Regulation[22] and CCPA the California Consumer Privacy Act[23], both privacy laws.
Retention and disposal work together, and they do not conflict with the lineage section above. Retention supports legitimate operational and evidence needs, such as keeping the version records that show what supported an earlier AI output; the workload assessment's example answer keeps the last five versions of each critical dataset, with a retention period of 18 months for each version[4], an illustration rather than a rule. Disposal controls data that no longer has a justified use. Indefinite accumulation is neither.
The takeaway: give every AI-related data asset a retention and disposal rule tied to its purpose and to compliance obligations.
Sizing the technology foundation to workload demand
This section begins the technology foundation: can the platform carry the workload? The business strategist does not design the infrastructure, but does state the demand it must meet, because the technical design follows from those numbers.
The AWS generative AI workload assessment asks for the expected volume of data or requests for the use case[10] (its example: 500 transactions per second), the key performance requirements, for example response time, throughput, and accuracy[10], and your current data processing capacity and how much you project needing for future generative AI workloads[4]. Four useful parameters for describing the demand are:
- Data volume: how much data the AI must store and process.
- Arrival rate: how many requests arrive per second or per hour.
- Concurrent demand: how many requests must be served at the same moment, especially at peaks.
- Required response speed (response time in the workload assessment): how quickly each result is needed.
Response speed separates two patterns. Real-time inferencing is the process of generating AI inference instantly based on input requests[24], used in time-critical systems, whereas batch inferencing is useful in non-time-critical applications, where you can wait for inference outputs[24]. A customer chat assistant needs the first; an overnight document summary needs the second.
Averages hide peaks. An average of 86,400 requests a day is one request per second, but if most of them arrive within a lunchtime hour, the platform must serve many times that rate during the hour. Approving capacity from average daily volume alone conceals these short periods of high concurrent demand, and a successful low-volume demonstration does not establish peak-load capacity.
Ask for capacity that scales with demand and balances performance against cost, rather than assuming the largest possible environment is necessary. The AWS CAF notes that managed technologies can automatically scale from zero to peak demands[25], the Generative AI Lens recommends elastic data infrastructure that can handle varying AI workloads and data volumes[1], and the workload assessment lists cost management: balancing the potential benefits with the high costs of implementation and operation[26] among the challenges to plan for.
The takeaway: state data volume, arrival rate, concurrent demand at peaks, and required response speed, and ask for scalable capacity that meets the peak at a justified cost.
Integration with business systems and workflows
For an AI capability intended to operate within an existing business process, integration means obtaining the information it needs and delivering results into the systems and workflows people already use. The AWS generative AI workload assessment lists integration complexity, incorporating generative AI into existing workflows and legacy systems[26], among the main adoption challenges, and asks how mature your existing systems that integrate with new generative AI technologies are[27]; its example answer describes systems built on APIs and standardized on common data formats.
Assess three connections, each a foundational requirement rather than an optional addition after a demonstration:
- Inputs: can the AI obtain the information it needs through authorized data access, from the systems where that information lives?
- Outputs: do results arrive in the tools people already use, such as the case-management or order system, instead of a separate window they must copy from?
- Handoffs: what happens when the AI cannot complete a case, and who receives it?
A standalone tool can prove a capability without proving that the organization can use it in its end-to-end process. A high demonstration score from a standalone application is therefore not integration acceptance; acceptance requires evidence of working connections into the business workflow.
The takeaway: before approving an AI capability, confirm how it gets its inputs, where its outputs land, and how handoffs work in the real process.
Managed AI platforms: Amazon Bedrock and SageMaker AI
When assessing the technology foundation, first decide which platform category the initiative needs, because that choice changes how much of the technology foundation the organization must build. This section covers the two AWS AI platforms listed in the exam guide's AI and Machine Learning category at that strategic level.
- 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[28]. A foundation model is a large model pre-trained on broad data that many applications can build on; AWS explains that data scientists use a foundation model as a starting point to develop ML models[29], a different sense of "foundation" from an AI initiative's data and technology foundations. Since Amazon Bedrock is serverless, you don't have to manage any infrastructure[28].
- Amazon SageMaker AI is a fully managed machine learning service with which data scientists and developers build, train, and deploy ML models[30].
| If the initiative needs | Platform category | What the organization still has to provide |
|---|---|---|
| Generative AI on its own information, such as an assistant that answers from company documents | Managed foundation model access (Amazon Bedrock) | Well-governed source data, prompt and evaluation skills, integration |
| A custom model trained on its own data, such as a demand forecast | Custom machine learning (Amazon SageMaker AI) | Training data, data science and ML engineering skills, model operations |
AWS guidance notes that most organizations are unlikely to undertake fine-tuning or distillation because it is often unnecessary for their use cases[9], and that many business needs can be met effectively using pre-trained foundation models, sometimes with light customization through prompt engineering or RAG[9] (retrieval augmented generation, which supplies the model with relevant company information at query time). Fine-tuning, by contrast, requires considerable investment in terms of technical capability, data curation, and model governance[9]. Funding a custom model-development program for every generative AI use case therefore misreads the need: some initiatives need managed model access, not new model construction.
The takeaway: match the platform category and the supporting skills to the initiative, instead of assuming every AI project needs an internally built model environment.
Protection and observability in the assessment
A usable AI foundation needs four capabilities beyond capacity: protecting sensitive data, governing access, recording relevant activity, and observing service and AI quality. A compute performance benchmark says little about any of them, so it cannot serve as the complete platform acceptance test.
A managed service does not remove the need to assess them. Under the AWS shared responsibility model, security is a shared responsibility between AWS and you[31]: AWS handles security of the cloud by protecting the infrastructure that runs AWS services[31], while for security in the cloud your responsibility is determined by the AWS service that you use[31], and you remain responsible for the sensitivity of your data, your company's requirements, and applicable laws and regulations[31]. Choosing Amazon Bedrock or Amazon SageMaker AI shifts some infrastructure work to AWS; it does not hand over your data decisions.
AWS guidance on the data side of generative AI adoption expects these capabilities as an initiative moves toward production:
- Protection and access: deploy encryption, role-based access control (RBAC), and least-privilege access models[19].
- Activity records: monitor and audit data access, including access through ML models[3].
- Observability: establish data observability practices; track data lineage, provenance, and quality metrics[19], with tools that give visibility into data quality, pipeline performance, and AI model behavior[1].
Your part as a business strategist is to state the business requirements (which data is sensitive, who must be able to see what, what must be recorded, and what quality must be watched) and to obtain evidence from the responsible technical and risk teams. Configuring the infrastructure is not the strategist's task. The AWS AIB-C01 exam guide lists identifying appropriate access controls and data security measures for AI systems[11] under 'Task 3.2: Establish AI governance structures and ensure regulatory compliance'; see Establish AI Governance and Regulatory Compliance.
The takeaway: accept an AI platform only with evidence of protection, access governance, activity records, and observability for the complete application, not just a performance benchmark.
Exam patterns: data and infrastructure scenarios
For practice, consider a stalled or risky AI initiative with one option that fixes the real foundation gap alongside tempting options that add volume, capacity, or access. Use the cues in this practice table to identify the foundation involved.
| Business situation | Suggested approach | Tempting wrong answer, and why it fails |
|---|---|---|
| A data acceptance criterion is needed | Measures per relevant quality dimension | A record-count target: volume does not establish correctness or completeness |
| Required fields are almost all filled, but duplicates and invalid IDs remain | Add duplicate-rate and validity measures | Treat completion rate as readiness: other dimensions still fail |
| Clean historical data comes from a different population | Check relevance and representation of the target setting | Use the clean records as representative: cleanliness is not relevance |
| A live policy assistant uses the archive's publishing schedule | Set refresh frequency from the business change rate | Keep the archive schedule: it may not meet the currency requirement |
| Stores record the same information differently | Standardize collection, definitions, and checks at the source | Correct outputs centrally: the recurring source defect remains |
| A pilot works on one database but must now use other systems | A unified data access and governance strategy | More processing capacity: it connects no disconnected source |
| Records the AI needs are inaccessible | Resolve the silo (integrate, or agree ownership and sharing) | Model retraining: the records stay out of reach |
| Two departments' systems are connected but they still will not share | Agree ownership and sharing rules | A new connection: connectivity does not resolve an organizational cause |
| The data assessment lists only the reporting warehouse | Add documents, messages, and media to the inventory | Treat the warehouse inventory as complete: knowledge sits outside tables |
| Success of the data strategy must be measured | Readiness of targeted business initiatives | Total data collected: volume does not show business support |
| Nobody can say who approves access to a dataset | Name a business data owner and stewards | Default to the storage administrator: custody is not authority |
| Business units want autonomy | Federated governance with shared policies and definitions | Separate departmental rules: incompatible rules block cross-domain use |
| Connected datasets disagree about what a term means | Shared glossary and data dictionary | A common file format: format does not align meaning |
| Teams cannot find data | A data catalog with ownership, meaning, provenance, and quality | Folder names: location explains none of these |
| Collaboration is slow because access takes weeks | Access-request workflow with role-appropriate permissions | Organization-wide read access: availability is not authorization |
| A department sent another a file and calls it a sharing agreement | Agree purposes, conditions, and responsibilities | Treat the transfer as the agreement: transport does not define reuse |
| An earlier AI output must be explained | Lineage and version records | The source department's name: it records neither changes nor version |
| Capacity is approved from average daily volume | Size for concurrent demand at peaks and required response speed | Average volume alone: it hides short peaks |
| A standalone demo scored well | Evidence of working connections into the workflow | Demo score as integration acceptance |
| A generative AI use case, with no ML team | Managed foundation model access (Amazon Bedrock) | A custom model-development program for every use case |
| A platform passed a compute benchmark | Evidence of protection, access, activity records, and observability | Benchmark as the complete acceptance test |
Name the foundation first: data readiness, data strategy and governance, or technology foundation. These are the three foundations from Three foundations, worked backward from the initiative, matching the three skills the AWS AIB-C01 exam guide lists under Task 4.2: Establish data and infrastructure foundations for AI[5].
Distrust volume, capacity, or access alone: adding records, compute, or access without addressing the diagnosed gap does not by itself demonstrate that the specific data readiness (reach and fitness), data strategy and governance, or technology foundation requirement is met.
The takeaway: identify the missing foundation, whether data readiness (reach and fitness), data strategy and governance, or the technology foundation, then choose the option that closes that gap rather than one that adds volume, capacity, or access.
The three foundations of an AI initiative
| Aspect | Data readiness | Data strategy and governance | Technology foundation |
|---|---|---|---|
| Question it answers | Is this data fit for this use case, and can the AI reach it? | Who decides about the data, and how is it found and shared? | Can the platform carry, connect, and protect the workload? |
| What to assess | Fitness: six quality dimensions (timeliness as refresh requirements), relevance, representation, and quality controls at the source. Reach: accessibility across silos, including unstructured information | Ownership and stewardship, governance style, shared definitions, sharing conditions, catalog, lineage, lifecycle | Data volume, arrival rate, concurrent demand, required response speed, integration, platform category, protection, observability |
| Evidence it is in place | Acceptance criteria per relevant dimension, measured on data from the target setting | Named owners and stewards, an access-request workflow, cataloged assets with lineage and lifecycle rules | Peak-load evidence, working connections into business workflows, security and monitoring evidence from responsible teams |
| Tempting wrong fix | Collect more records | Grant everyone read access | Buy more compute or a larger environment |
| Why the wrong fix fails | Volume establishes neither correctness nor coverage | Availability is not authorization for every user and purpose | Capacity reaches no new data and proves neither integration nor protection |
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.
- Data quality dimensions identify different kinds of readiness defects
Data accuracy asks whether values are correct, completeness asks whether required information is present, and consistency asks whether representations agree across records and systems. Timeliness asks whether data is current enough for its intended use. Evaluate the dimensions that matter to the business process so a dataset's strength in one dimension does not conceal a different defect.
Trap Using a high record-count target as the data acceptance criterion because volume does not establish correctness or completeness.
4 questions test this
- A property insurer wants to train a risk model on building records. Every required field is filled in, and the fields use the same formats across systems. A sample inspection finds that clerks often m
- A grocery chain plans an AI tool that suggests same-day restocking. Its shelf data scores well in a quality review for accuracy and completeness. The data reaches the planning system three days after
- A transit agency is buying 20 million rider feedback records for an AI service analysis tool. The vendor proposes that delivery of the full volume should count as acceptance. Agency leaders do not wan
- A retailer's dashboard gives its product dataset an overall quality score of 91 out of 100, which averages six facets. The team building an AI product search notes that shoppers filter on product attr
- Reliable AI inputs require consistent collection and source-level quality controls
When sites collect the same business information differently, standardize collection practices, definitions, and quality checks before expanding an AI initiative. Assign responsibility for recurring defects and improve the collection process at their source. Consistent controls across locations make reliable inputs an ongoing capability instead of a one-time cleanup.
Trap Using centralized output correction to compensate for inconsistent store inputs because it leaves the recurring source defect in place.
4 questions test this
- A hotel group's data team spent a month cleaning guest data before an AI launch, and the first results were strong. Six months later, the same formatting and accuracy defects are back. The team propos
- A bank's regional offices each collect small-business loan data. Each region knows its customers well, but they use different definitions, and an AI credit-review tool gets uneven inputs. The bank wan
- A distributor's AI demand forecasts suffer because the same warehouses submit late and malformed files every month. The analytics team fixes the files itself. No one at the warehouses knows about the
- A restaurant chain's AI tool summarizes guest complaints. Stores record the same complaint in different ways, so a central team corrects the summaries each week. The correction work grows every month.
- Data must represent the people and situations the AI initiative will serve
Data readiness includes relevance to the use case and representation of the intended users, operating conditions, and business situations. A large, internally consistent dataset can still omit important populations or contexts. Evaluate coverage of the target setting and potential representational bias before treating existing records as an adequate foundation.
Trap Using clean historical records as the default representative dataset because cleanliness does not establish relevance to the target population.
4 questions test this
- A national bank tested an AI chat assistant on logs from its large city branches. The logs are clean and well labeled, and the test scores are strong. The bank will offer the assistant to all customer
- A telecom company wants an AI assistant for business customers. The only available conversation history comes from its consumer support team, and it is large and well organized. Business customers ask
- A hospital network has 15 years of records from its teaching hospitals. It wants an AI tool that flags readmission risk at its new community clinics, which serve older patients and more rural resident
- A state benefits agency holds ten years of spotless application records. Planners want to use them to train a tool that helps gig workers and recent migrants, because the records passed each quality t
- Required data freshness depends on how quickly the business information changes
Set data update expectations according to the use case's need for current information. A dynamic knowledge source or time-sensitive decision process needs a different refresh approach from a stable reference collection. Include update frequency and quality monitoring in readiness requirements so data that was suitable at the pilot's start remains usable later.
Trap Using the document archive's publication schedule for a live policy assistant because that schedule may not meet its currency requirement.
4 questions test this
- An airline's AI assistant helps gate agents rebook stranded passengers. It answers from seat availability that is loaded overnight, so it often offers seats that are already gone. Availability changes
- A company's HR assistant answers from a history handbook that rarely changes and from benefits rules that change every month. All sources now refresh each quarter, so benefits answers are sometimes ou
- A pharmacy chain's AI assistant for store staff performed well at launch, using the current drug coverage lists. The readiness checklist looked only at data quality on launch day. Coverage lists chang
- A manufacturer's AI assistant answers questions from engineering standards that change about once a year. The vendor offers continuous real-time synchronization at a premium, and the product manager a
- Uniqueness and validity reveal defects that correct-looking values can conceal
Uniqueness checks for unwanted duplicate records of the same business entity or event; duplicates can inflate counts and give repeated evidence undue weight even when their individual values are correct. Validity asks whether values meet the required formats and business rules, such as an accepted identifier format or a permitted category. Include duplicate rate and the share of values passing validity rules in the data-readiness scorecard, with acceptance criteria tied to the intended business use. Passing a format rule does not establish that a value is factually accurate, and an accurate value does not establish that its record is unique.
Trap Using a high required-field completion rate as sufficient evidence of readiness when duplicate customers and invalid identifiers remain.
4 questions test this
- A car dealer group is preparing vehicle records for an AI pricing tool. Reviewers find that many vehicle identifiers have the wrong number of characters. Some status entries use words that are not on
- A manufacturer's supplier list passed a field audit, and each address and tax number matches the real company. An AI spend analysis still shows more suppliers than the firm actually uses, because some
- A telecom firm will use customer records to personalize renewal offers with AI. Its data vendor reports 99 percent field completion. In a pilot, some customers received two offers, and some plan codes
- A lender's acceptance report for a customer dataset shows that 98 percent of required fields are filled in, so the data passes. Reviewers then find the same customers listed several times and account
- Fragmented enterprise data requires a unified access and governance strategy
When AI must use information across disconnected systems, establish a unified data strategy that connects and governs the required sources. Unified access can work across existing locations; it does not require every asset to reside in one physical repository. Resolve fragmentation in access, formats, and governance as an enterprise foundation rather than leaving each initiative to negotiate an isolated solution.
Trap Using additional processing capacity to expand a single-database pilot because capacity does not connect disconnected information sources.
5 questions test this
- A telecom company runs 25 customer data systems. Its first AI initiative, an agent-assist tool, needs records from four of them. Leaders want the tool live this year and want the next three initiative
- A manufacturer has three AI initiatives that each need supplier, quality, and shipment data. Each project team negotiates its own access with the same system owners and builds its own connections. App
- A logistics company piloted an AI document assistant against one contract database and reached 94% accuracy. The steering committee wants to approve a company-wide rollout across 11 business units, wh
- A retailer's AI merchandising assistant performed well with the sales database alone. After the pilot expanded, it now gives incomplete recommendations because inventory, promotion, and supplier infor
- A global insurer wants an AI claims assistant to use records held in four regional systems. Local regulations require some records to remain in their home regions. Leaders want the assistant live with
- Data democratization means easier authorized access, not unrestricted access
A data sharing framework should make relevant data discoverable and accessible to appropriate users while preserving privacy and use restrictions. Define clear access-request workflows and role-appropriate permissions so routine legitimate use does not depend on ad hoc negotiation. Excessive control can recreate silos, while excessive access exposes information beyond its permitted audience.
Trap Using an organization-wide read permission to accelerate collaboration because broad availability does not establish authorization for every user.
4 questions test this
- A retailer wants to speed up AI experiments. The operations director proposes giving every employee read permission on all company datasets, including payroll and customer identity records. Analysts n
- A logistics firm's AI teams ask individual managers for data, and each manager applies a different test. Similar requests get different answers, and legitimate work waits for weeks. Leaders want routi
- A pharmaceutical company requires a vice president's signature for every data request, including requests for low-risk operational reports. Project teams now keep private spreadsheet extracts that fal
- A media company let every employee copy any dataset into a personal workspace to speed up AI work. Six months later, several copies have gone stale, and some subscriber records sit in workspaces with
Shared business glossaries, data dictionaries, and reference definitions give teams a consistent interpretation of enterprise data. Reconcile meanings such as customer status, product identifiers, and reporting periods before using information together. Technical connectivity solves access, whereas shared semantics make the combined information interpretable and comparable.
Trap Using a common file format as proof of interoperability because matching formats can still carry conflicting business meanings.
5 questions test this
- Sales and service teams at an insurer use the term active policy differently, and an AI retention model needs one meaning. Data engineers can change either field. The service director and sales direct
- A global manufacturer's AI analysts keep asking which definition of on-time delivery is correct, because each plant uses its own. The company has now agreed on one definition. It wants analysts to see
- A distributor bought two companies and connected all three order systems. An AI assistant now reports each customer's total orders, but the same buyer appears under different account numbers in each s
- A bank's AI team joined card and loan data, and every job ran without errors. A review then found that one division reports revenue by booking date and the other by settlement date. The team proposes
- A hospital network combines discharge data from two regions for an AI readmission model. Both regions export the same file format with identical column names. Results differ widely, because one region
- An AI data strategy must account for useful unstructured information
Enterprise AI can need documents, messages, images, or other unstructured material as well as structured business records. Include the relevant formats in data discovery, classification, quality assessment, and authorized access planning. Plan how users and AI applications will find and understand this information's context while preserving appropriate controls.
Trap Using the existing reporting warehouse inventory as the complete data assessment because important business knowledge may reside outside tabular systems.
5 questions test this
- A construction firm is rating its readiness for an AI project assistant. The data team's assessment covers only the finance warehouse, and it reports high readiness. Project staff say that their knowl
- A law firm wants an AI assistant to search 80,000 client contracts and emails. Some items carry confidentiality terms, but no one has recorded which ones. Partners want the assistant available to asso
- A property insurer wants AI to help assess claims. Adjusters rely on policy records, phone call recordings, and damage photos. The data plan covers only text and tables, and leaders assume that other
- A utility's HR assistant will answer employee policy questions from 12,000 documents on an intranet. A sample shows that different departments have posted conflicting versions of the same policy, with
- A media company plans AI tools that search its video, audio, and document archive, which grows by terabytes each week. Its current foundation was sized for nightly reporting on structured tables. Busi
- Data silos restrict usable business context and create reconciliation work
Data silos are isolated systems or departmental collections whose information is difficult for other authorized users or applications to access or share. They limit the business context available to AI and can create duplicate or inconsistent records that require manual transfer and reconciliation. Diagnose whether the isolation comes from disconnected technology or from organizational barriers such as ownership disputes and unwillingness to share, because technical connectivity alone does not resolve the latter.
Trap Funding model retraining to compensate for business records that remain inaccessible to the AI application.
5 questions test this
- A bank's AI relationship assistant gives weak recommendations because account history sits in a departmental system that the assistant cannot access. A vendor proposes to fund model re-training to imp
- A new chief data officer suspects that data silos are slowing AI work but cannot say where they are. Business units describe different problems, and no overview of systems exists. The officer wants an
- A bank replaced its departmental data stores with one unified platform for an AI lending project. Six months later, the platform works as designed, yet most departments still keep key files on their o
- A manufacturer's sales and service systems already exchange data through modern APIs, yet its AI warranty model still lacks service history. The service vice president refuses to release the records,
- A manufacturer's quality data sits in a 25-year-old system with no modern interfaces, so staff export it by hand for an AI analysis tool. The exports are slow and error-prone. Leaders want a dependabl
- Data strategy should work backward from business initiatives
A data strategy should identify which domains, sources, quality levels, and access capabilities are needed for targeted business initiatives. Translate the business requirements into a data roadmap and measures of whether the required information is fit for its intended purpose. This makes the data foundation an enabler of business value rather than a collection program seeking a future purpose.
Trap Using total enterprise data collected as the data strategy's success measure because collection volume does not establish support for business priorities.
4 questions test this
- A manufacturer's data office plans to spend a year listing every dataset in the company before it writes a data strategy. The chief executive has funded three AI initiatives for this year. The executi
- A logistics firm has funded an AI route planning initiative, and the data team is drafting the strategy section for it. The first draft lists the firm's storage platforms and its total data volume. Th
- An insurer's leaders have agreed on four priority AI initiatives and the data that each one needs. Six months later, little has changed. Teams disagree on who acts first, when work should finish, and
- A retailer's data strategy reports success as the number of terabytes collected each quarter, and the figure rises every year. Yet AI teams say that they still lack the data that their priority projec
- Data ownership and stewardship provide different forms of accountability
Data owners define policies for their data, including access conditions and key business definitions, while data stewards apply governance in day-to-day work, identify issues, and escalate them. Technical teams provide the supporting mechanisms, but business responsibilities must still be explicit. Route policy decisions to the responsible owner and recurring quality or governance issues through the stewardship process.
Trap Using the storage administrator as the default business data owner because technical custody does not establish authority over permitted business use.
4 questions test this
- A marketing team asks to use customer service transcripts to train an AI model for campaign targeting. The current policy does not address this use. The storage team can grant access in minutes. The d
- A company is assigning ownership of its employee data ahead of an AI workforce planning project. The storage administrator manages the systems and suggests that this makes them the owner. HR leaders d
- A company's new AI program is confused about roles. Engineers are deciding who can see customer data, business staff think that IT handles all quality issues, and no one follows up on policy breaches.
- A call center's AI assistant keeps giving wrong answers because product codes in its knowledge data are often mislabeled. Support agents see the errors daily, but no one acts on them. The problem seem
When business units need local data expertise and autonomy, federated governance lets domains manage their data while a central function maintains shared policies and common capabilities. Preserve coherence across domains so information can still be linked and used responsibly. Centralized governance may fit common control needs, but local ownership need not mean independent, incompatible standards.
Trap Using separate departmental dictionaries and access rules as the definition of federation because incompatible rules undermine cross-domain use.
4 questions test this
- A consumer goods company's regional units understand local markets and data far better than its head office. All data requests now wait on one central team, and projects stall. Regulators expect consi
- A media group's brands each manage their own audience data, and their local knowledge is strong. Every brand now runs the same data quality checks with separately bought tools, so the group pays for t
- A company is moving to a federated data model. Domain teams will run their own data products, and the products must work together so that AI projects can combine them. Leaders are now deciding what th
- A retailer uses a federated data model. Its supply chain domain has built a delivery performance data product, and other domains want to combine it with their own data for an AI project. The domain wa
- A data sharing framework must define the conditions of use
Data sharing rules should specify who may use which data, for what purposes, and under what protection and compliance conditions. Assess privacy, confidentiality, consent, and applicable use restrictions alongside availability and quality; sensitive information may require protection or exclusion. Make ownership, approval, and ongoing responsibilities clear across organizational boundaries so authorized access is paired with an authorized AI purpose.
Trap Using a successful interdepartmental file transfer as evidence of an adequate sharing agreement because transport does not define permitted reuse.
4 questions test this
- A retailer's divisions want to share customer data with a central AI team. The data includes payment details that the model does not need, along with useful purchase history. The privacy office wants
- A company's HR team successfully sent an employee survey file to the analytics team, which wants to use it to train an AI retention model. The project lead says that the completed transfer shows that
- A retail chain will share sales data with an external supplier's AI team for a joint demand forecast. Both parties have agreed on the data fields. No one has stated who approves changes, who is accoun
- A sales team received a customer support dataset through a secure, approved file transfer. It now says that the data is cleared for an AI churn model. Customers gave this information to get help with
- A data catalog helps users find and understand candidate data
A data catalog inventories data assets and their metadata, including ownership, meaning, provenance, and quality information. It supports discovery and informed access requests instead of forcing each AI team to rediscover sources. Treat the catalog as part of governance: a listing alone does not supply the policies, stewardship, and quality processes that make data trustworthy.
Trap Using folder names as the enterprise data catalog because locations alone do not explain ownership, meaning, or fitness for use.
4 questions test this
- A bank's catalog holds complete technical metadata, such as tables, columns, and data types. Marketing analysts planning an AI project still cannot tell what a field means for the business or whether
- An airline's AI teams each keep private lists of useful datasets, and a list is lost when an analyst moves on. New teams rediscover the same sources and cannot tell who owns them or how reliable they
- A pharmaceutical company's catalog lists every research dataset, but AI scientists cannot tell how to obtain any of them. They email various contacts, and requests stall. The company wants scientists
- An insurer must show auditors which AI data stores contain regulated information. Its catalog already lists the stores. Compliance staff know the rules, and application teams know the data, but neithe
- Lineage and version records make AI data use traceable
Data lineage records where information originated and how it changed, while version records identify the particular dataset state used. Together they support impact assessment, investigation of quality problems, and reproduction of prior results. A dataset can be suitable for operational use without being sufficient evidence of what supported an earlier AI output.
Trap Using the source department's name as the lineage record because ownership metadata does not describe transformations or the dataset state used.
4 questions test this
- A bank's AI credit model uses a dataset that is labeled in its catalog with the source department's name. An examiner asks where the data originated and how it was changed before training. The team ca
- A food company learns that a recalled ingredient table fed several planning datasets for two months. Some AI forecasts and executive reports may have used those datasets. The quality team wants to cor
- A health insurer's AI model produced a coverage recommendation last year, and a regulator now asks for the exact data behind it. The current dataset is suitable for daily operations, but it has been u
- A logistics company's AI delivery-time model gave a disputed estimate to a major customer three months ago. The operations director says that the dataset passes every daily quality check, so it proves
- AI data needs lifecycle rules rather than indefinite accumulation
Data governance should define lifecycle policies for retaining, updating, and disposing of data in line with business needs and compliance obligations. Include AI-related data assets in those policies so old copies do not become unmanaged exceptions. Retention supports legitimate operational and evidence needs, while disposal controls data without a justified use.
5 questions test this
- A hospital network retires an AI triage pilot. The project space still holds training data with patient details, along with the evaluation results that regulators may ask about. No one owns the space
- An insurer's model risk team needs the data behind each approved claims model for seven years. Separately, data scientists hold many abandoned experiment copies with customer details. One executive wa
- A company's lifecycle policy covers its production systems. An audit finds that AI teams hold fine-tuning datasets, evaluation samples, and test copies in project spaces, and no rule applies to them.
- A bank's AI teams keep every extract from finished projects, because the data might be useful someday. Many extracts hold customer details, and no rule says when they should go. Legal wants exposure r
- A retailer applies one five-year retention rule to everything that AI teams use, from public product photos to customer chat logs. Security leaders say that the logs stay too long. Merchandising leade
- Infrastructure requirements must reflect workload demand and service expectations
Assess whether the foundation can handle the expected data volume, arrival rate, concurrent demand, and required response speed. A real-time interaction needs different responsiveness from a scheduled batch task, and a successful low-volume demonstration does not establish peak-load capacity. Ask for a scalable capacity approach that balances performance with cost rather than assuming the largest possible environment is necessary.
Trap Using average daily request volume alone to approve capacity because it can conceal short periods of high concurrent demand.
5 questions test this
- A retailer tested an AI shopping assistant with 40 employees, and responses were fast. The company plans to open it to 300,000 customers on a major sale weekend. The project sponsor says that the test
- A travel company's AI assistant sees heavy demand during holiday booking periods and light demand for most of the year. The finance team objects to paying year-round for peak capacity. The operations
- A card issuer wants an AI model to score each transaction for fraud within seconds, before approval. Transaction data currently reaches its analytics platform in a nightly file. The pilot used last mo
- A health plan will launch an AI enrollment assistant. Planners expect 50,000 requests per day on average, and a vendor recommends sizing capacity to that average. During the six-week open enrollment p
- A bank has two AI initiatives. One scores overnight batches of loan documents for review the next morning. The other helps contact center agents during live calls. A sponsor proposes one set of perfor
- An AI capability needs a workable connection to the business process
Assess how the AI solution will obtain necessary information and deliver results into the systems and workflows people already use. Compatibility, authorized data access, and operational handoffs are foundational requirements, not optional additions after a demonstration. A standalone tool can prove a capability without proving that the organization can use it effectively in its end-to-end process.
Trap Using a standalone application's demonstration score as integration acceptance because it does not establish usable connections to the business workflow.
5 questions test this
- A utility is evaluating an AI scheduling product for field crews. The demonstration impressed leaders. The utility's work order system is 20 years old and lacks the modern interfaces that the vendor's
- A sponsor's business case for an AI forecasting tool includes licenses and a pilot but no funding for connecting the tool to planning systems. The sponsor plans to address integration after the demons
- An insurer's AI tool drafts claim summaries in a separate web application. Adjusters copy each summary into the claims system by hand. Adoption has stalled, and errors have appeared during copying. Th
- A telecom company's AI assistant scored 96% in a standalone demonstration of answering billing questions. The sponsor wants to treat that score as integration acceptance and release it to agents. Agen
- A distributor wants an AI assistant to answer order status questions from customers. The demonstration used a static export of last month's orders. In production, the assistant must reflect current or
- Managed AI platforms can reduce the foundation an enterprise must build itself
When assessing foundational needs, distinguish access to managed foundation models from a broader requirement for custom machine learning capabilities. Amazon Bedrock provides managed access to foundation models for generative AI applications, while Amazon SageMaker AI supports building, training, and deploying machine learning models. Match the platform category and supporting skills to the initiative instead of assuming every AI project needs an internally built model environment.
Trap Using a custom model-development program for every generative AI use case because some initiatives need managed model access rather than new model construction.
5 questions test this
- A logistics company chose a managed foundation model service for a document assistant. The project plan had included budget for model hosting, data preparation, integration, and access governance. The
- A professional services firm wants a generative AI tool that drafts proposals from approved templates. An executive proposes hiring a team to build and train the firm's own foundation model on dedicat
- A bank's AI portfolio includes an employee assistant built on foundation models and a credit risk model that its quantitative team must build and validate itself. The CIO wants one platform mandated f
- A consultant is assessing foundational needs for a client's next AI initiative. The client assumes that every AI project needs an internal model environment. Before recommending a platform category an
- A regional retailer will launch a customer FAQ assistant on managed foundation models. The HR director has drafted a plan to hire five model-training specialists. The assistant will use existing model
- A usable AI foundation needs protection and observability as well as capacity
Foundational assessment should include the ability to protect sensitive data, govern access, record relevant activity, and observe service and AI quality. Computational capacity cannot establish these capabilities, and a managed service does not remove the need to assess them for the complete application. State the business requirements and obtain evidence from the responsible technical and risk teams without treating infrastructure configuration as the business strategist's task.
Trap Using a compute performance benchmark as the complete platform acceptance test because it says little about protection or operational visibility.
5 questions test this
- A business strategist is leading a foundation assessment for an AI claims tool. The strategist is not a security specialist and cannot judge technical controls. Leaders still expect an informed view o
- A retailer's AI assistant has met its uptime and latency targets for three months. Customer complaints about wrong answers are rising, yet operations dashboards stay green. The assessment team is revi
- A vendor offers a benchmark that shows its AI platform handling a bank's expected workload with ease. The steering group wants to use the result as the final acceptance test. The platform will handle
- An insurer reviews an AI platform that lets adjusters ask questions about claim files, which contain medical details. The review covers capacity and latency. Internal audit asks how the insurer will s
- A retailer selects a managed AI service for a customer assistant, which will read order history and loyalty data. The IT lead proposes dropping the data protection review, because the provider secures
Also tested in
References
- https://docs.aws.amazon.com/wellarchitected/latest/generative-ai-lens/data-architecture.html
- https://aws.amazon.com/what-is/data-silos/
- https://aws.amazon.com/what-is/data-governance/
- https://docs.aws.eu/prescriptive-guidance/latest/gen-ai-workload-assessment/data-strategy.html
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain4.html
- https://aws.amazon.com/blogs/big-data/build-a-serverless-data-quality-pipeline-using-deequ-on-aws-lambda/
- https://docs.aws.amazon.com/wellarchitected/latest/devops-guidance/metrics-for-data-lifecycle-management.html
- https://aws.amazon.com/smart-business/resources-for-smb/data-quality-management/
- https://docs.aws.amazon.com/prescriptive-guidance/latest/strategy-data-considerations-gen-ai/lifecycle.html
- https://docs.aws.eu/prescriptive-guidance/latest/gen-ai-workload-assessment/use-cases.html
- https://docs.aws.amazon.com/aws-certification/latest/ai-business-strategist-01/ai-business-strategist-01-domain3.html
- https://docs.aws.amazon.com/wellarchitected/latest/machine-learning-lens/mlops03-bp01.html
- https://docs.aws.amazon.com/whitepapers/latest/overview-aws-cloud-adoption-framework/governance-perspective.html
- https://aws.amazon.com/what-is/structured-data/
- https://aws.amazon.com/what-is/data-mesh/
- https://aws.amazon.com/what-is/data-catalog/
- https://aws.amazon.com/what-is/shadow-it/
- https://csrc.nist.gov/glossary/term/personally_identifiable_information
- https://docs.aws.amazon.com/prescriptive-guidance/latest/strategy-data-considerations-gen-ai/journey.html
- https://docs.aws.amazon.com/wellarchitected/latest/security-pillar/sec_data_classification_define_protection.html
- https://docs.aws.amazon.com/wellarchitected/latest/machine-learning-lens/mlops02-bp04.html
- https://gdpr.eu/what-is-gdpr/
- https://aws.amazon.com/compliance/california-consumer-privacy-act/
- https://aws.amazon.com/what-is/inferencing/
- https://docs.aws.amazon.com/whitepapers/latest/overview-aws-cloud-adoption-framework/platform-perspective.html
- https://docs.aws.eu/prescriptive-guidance/latest/gen-ai-workload-assessment/introduction.html
- https://docs.aws.eu/prescriptive-guidance/latest/gen-ai-workload-assessment/readiness.html
- https://aws.amazon.com/bedrock/faqs/
- https://aws.amazon.com/what-is/foundation-models/
- https://docs.aws.amazon.com/sagemaker/latest/dg/whatis.html
- https://docs.aws.amazon.com/bedrock/latest/userguide/security.html