Domain 1 of 4 · Chapter 2 of 4

Data in Gen AI

Why data decides gen AI outcomes

Picture a support assistant grounded on a knowledge base where half the articles are two product versions out of date and a few of them flatly contradict each other. The model can be state of the art and it will still answer confidently and wrongly, because it can only work with what you give it. That is the rule to carry through this whole topic: a gen AI system is only as good as the data behind it.

Data shapes the answer at two moments

A gen AI system inherits data quality twice. The foundation model, the large general-purpose model introduced in the Core Gen AI Concepts subtopic, was pre-trained on an enormous body of data you did not choose and cannot change. Then, at run time, you shape its answers with your own enterprise data through grounding or fine-tuning, and that data you do own. A leader's leverage sits almost entirely in the second moment, so treating enterprise data readiness as a business decision, not a downstream engineering chore, is the highest-value habit in this domain.

Three axes for judging a dataset

Instead of the vague question 'is our data good', judge any dataset along three independent axes, and let each one drive a different decision:

  • Quality and accessibility: is the data complete, consistent, relevant, reachable, affordable, and in a usable format? This decides whether the data is usable at all.
  • Structure: is it structured, semi-structured, or unstructured? This decides which tools and models fit.
  • Labeling: does it carry the answers (labeled) or not (unlabeled)? This decides how much manual tagging a solution needs.

The figure below draws that same decomposition: any dataset fans out to the three axes, with the concrete checks each axis applies listed underneath it. The rest of this page walks the three axes in turn.

Any datasetQuality & accessibilityStructureLabelingCompleteness, consistency,relevance, availability,cost, formatStructured /semi-structured /unstructuredLabeledvs unlabeled
The three axes for assessing any dataset before committing to a gen AI use case.

Data quality and accessibility

High-quality data is data you can trust and reach at the moment the model needs it. The exam frames data quality through six characteristics, so here is each one at a business level, paired with the failure it causes when it is missing.

The six characteristics

Characteristic What it means What poor quality causes
Completeness The data holds all the information required for its purpose, per Google Cloud's data quality dimensions[1] Gaps produce answers that quietly skip whole cases
Consistency The same fact carries the same value everywhere, again per the data quality dimensions[1] Contradictions leave the model to pick one version at random
Relevance The data actually pertains to the question at hand Off-topic data dilutes and misleads the output
Availability You can reach the data when the model needs it, past silos and permissions Locked-away data cannot ground anything, however clean it is
Cost Acquiring, storing, cleaning, and labeling all cost money and time Budget burned on data no use case will consume
Format The data sits in a machine-usable form An unusable format stalls the project before it starts

Google Cloud's automated data-quality scans measure several of these as scored dimensions[1], including completeness and consistency, which is a useful reminder that they are concrete and checkable, not vague ideals.

Availability is a business problem, not a storage one

Quality and availability (also called accessibility) are different things, and a dataset can have one without the other. A pristine, perfectly consistent table sitting in a system the AI team cannot reach is worth nothing to a gen AI use case, while a reachable but messy dataset can at least be cleaned. Availability failures are usually organizational, meaning silos, unclear ownership, or slow access approvals, so a leader fixes them with governance decisions rather than with a better model.

Structured, semi-structured, unstructured

A row in a sales table (2026-07-01, SKU-42, 3 units, 59.97) and a customer's angry voicemail about that same order are both data, yet a database can query the first instantly and cannot make sense of the second. That gap is the structure axis, and it decides which tools and which kind of model can use the data.

The three kinds

  • Structured data fits a fixed schema of rows and columns: transaction records, CRM entries, inventory tables. It lives naturally in a warehouse like BigQuery[2] or a relational database, where it is easy to query and aggregate.
  • Semi-structured data has some organization but no rigid table shape, and its attributes vary from record to record. JSON and XML documents, log files, and sensor feeds are typical.
  • Unstructured data has no predefined schema at all: emails, PDFs, images, audio, video, and chat logs. On Google Cloud it usually lands in Cloud Storage[3], which holds an object of any format without imposing a schema.

Why this is the headline for gen AI

Here is the business punchline: roughly up to 80% of an organization's data is unstructured[4], and for years most of it went unanalyzed because traditional analytics needed structured tables. Multimodal gen AI changes the economics, because it reads documents, images, audio, and video directly. The value is not a marginal analytics gain, it is unlocking the largest and least-used part of the data estate for search, summarization, and question answering. The figure below sets the three kinds side by side in that order, and the callout hanging under Unstructured is this same unlock: models read that column directly. Structured data keeps its role as the backbone of analytics and the most reliable source to ground a model on, but it is the minority of what an enterprise actually holds.

StructuredFixed rows & columnsSales transactionsCRM recordsInventory tablesSemi-structuredTags vary per recordJSON, XMLLog filesSensor feedsUnstructuredNo schemaEmails, PDFsImages, audio, videoChat logsGen AI unlock:models read it directly
The structure spectrum; unstructured data is the majority of enterprise data and the main gen AI unlock.

Labeled versus unlabeled data

A photo saved as cat.jpg is unlabeled data; the same photo tagged 'cat' so a model can learn the association is labeled data. The label is the answer attached to an example, and whether your data carries one changes the whole cost structure of a project.

What each kind enables

  • Labeled data pairs each example with its correct output: emails marked spam or not spam, images tagged by object, tickets sorted by category. It is what supervised learning trains on (the machine learning approaches themselves live in the Core Gen AI Concepts subtopic). Labels are usually produced by people, so they are the slow, expensive part of classical machine learning.
  • Unlabeled data is the raw material with no answers attached, and it is the vast majority of what any organization collects.

The foundation-model shift

The reason this matters to a leader is a shift in where cost falls. Foundation models[5] are pre-trained on enormous amounts of largely unlabeled data and then adapted to specific tasks. Because the heavy learning already happened on unlabeled data, an organization can get real value by prompting or grounding a pre-trained model as-is, with little or no labeling of its own, and only needs a smaller labeled set if it decides to fine-tune the model or to measure output quality. The figure below traces that path end to end: vast unlabeled data pre-trains the foundation model, which then either serves as-is when you prompt or ground it, or takes a small labeled set to become a fine-tuned model. Classical machine learning asked 'can we afford to label millions of examples'; with foundation models the binding question becomes 'which data can we reach and trust'. Labeling has not disappeared, it moved from a mandatory upfront tax to an optional investment you make only when accuracy demands it.

Vast unlabeled dataFoundation modelPrompt or ground:use the model as-isSmall labeled setFine-tuned modelpre-trainno labelsadd labelsfine-tune
How gen AI uses each kind: pre-training runs on unlabeled data, and labeling is optional, only needed to fine-tune.

Exam-pattern recognition

The Generative AI Leader exam tests this subtopic mostly through classification and business-implication questions. Knowing the traps is half the battle.

Common stems and the right read

  • 'Which of these is an example of unstructured data?' The answer is free text, images, audio, or video; a relational table or spreadsheet is the structured distractor.
  • 'A company wants to use thousands of scanned contracts and support-call recordings.' That is unstructured data, and the business implication is that gen AI can now use it directly where older analytics could not.
  • 'What data quality issue explains a model giving contradictory answers?' Consistency. Missing fields point to completeness, and data the team cannot reach points to availability.
  • 'Why do foundation models reduce the need for labeled data?' Because they are pre-trained on large unlabeled datasets, so most of the learning happened before any of your labels exist.

Traps to avoid

  • Do not call JSON, XML, or log files unstructured. They are semi-structured, which is a favorite distractor.
  • Do not treat a spreadsheet or database table as unstructured. A fixed schema makes it structured.
  • Do not assume more data always means better output. Irrelevant or inconsistent data lowers quality regardless of volume.
  • Do not confuse availability with quality. Data can be perfectly clean and still useless if the AI team cannot reach it.

Data by structure: what gen AI can do with each

PropertyStructuredSemi-structuredUnstructured
SchemaFixed rows and columnsFlexible tags that vary per recordNo predefined schema
Real-world examplesSales transactions, CRM records, inventory tablesJSON and XML, log files, sensor feedsEmails, PDFs, images, audio, video, chat logs
Share of enterprise dataMinorityGrowing middleVast majority (up to ~80%)
Typical Google Cloud homeBigQuery, Cloud SQLBigQuery, Cloud StorageCloud Storage
Gen AI roleAnalytics and reliable groundingEvent and configuration contextThe main unlock: multimodal models read it directly

Decision tree

Data trustworthyand reachable?Fix data readinessfirstLabeled data andneed top accuracy?Fine-tune onthe labeled setContentunstructured?Multimodal gen AI,grounded on your dataStructured analyticsor groundingNoYesYesNoYesNo
Cheat sheet

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.

A model's output is capped by the quality of its data

Gen AI obeys garbage in, garbage out: an answer can be no better than the data behind it, so a state-of-the-art model fed stale, incomplete, or contradictory data still returns confident but wrong answers. For a leader this reframes data readiness as a business decision that gates the use case, not a cleanup chore to delegate later.

Trap Assuming a larger or newer model will compensate for low-quality data; the model cannot exceed what its inputs contain.

1 question tests this
Data shapes gen AI output at both training time and run time

Data influences a gen AI system twice: the huge corpus a foundation model was pre-trained on, which you do not control, and the enterprise data you supply at run time through grounding or fine-tuning, which you do. A leader's leverage lives in the second moment, so effort and governance belong there rather than in second-guessing the training corpus.

Judge a dataset on three axes: quality, structure, and labeling

Assess any dataset three ways, each driving a different decision: quality and accessibility decide whether it is usable at all, structure (structured, semi-structured, or unstructured) decides which tools and models fit, and labeling decides how much manual tagging a solution needs. The three axes are independent, so a dataset can score well on one and badly on another.

Completeness means the data holds everything the task needs

Completeness asks whether a dataset contains all the information required for its intended purpose. Gaps and missing fields lead a model to skip whole cases or answer around the hole, which is why completeness is one of the core data-quality characteristics the exam calls out.

5 questions test this
Consistency means the same fact has the same value everywhere

Consistency is having identical values for a fact across sources, such as a product's revenue matching between the sales database and the usage database. When data is inconsistent the model receives contradictory inputs and effectively picks one at random, which surfaces as answers that disagree with each other.

Trap Blaming contradictory model answers on missing data; contradictions point to a consistency problem, whereas gaps point to completeness.

4 questions test this
Relevance beats volume when judging data quality

Relevance is whether the data actually pertains to the question at hand. More data is not automatically better, because off-topic records dilute and mislead the output, so a smaller on-point dataset often beats a larger noisy one for a given use case.

Trap Assuming that adding more data always improves output; irrelevant data lowers quality regardless of volume.

5 questions test this
Availability is whether you can reach the data, not whether it is clean

Availability, or accessibility, is whether the AI team can actually get to the data when the model needs it, and it is separate from quality. A pristine, consistent table locked inside an unreachable system is useless to a use case, while a reachable but messy dataset can still be cleaned. Availability failures are usually organizational, meaning silos, unclear ownership, or slow access approvals, so a leader fixes them through governance.

Trap Treating unreachable data as a data-quality problem; access is an availability and governance issue, not something a cleaner dataset or a better model solves.

3 questions test this
Data carries real acquisition, storage, and labeling cost

Cost is a data-quality characteristic because acquiring, storing, cleaning, and labeling data all consume money and time. Weighing that cost against the value a use case returns is part of judging feasibility, since data that no use case will consume is wasted spend.

1 question tests this
Format decides whether data is machine-usable at all

Format is whether the data sits in a form a system can actually process. Data trapped in an unusable format stalls a project before it starts, which is why format sits alongside completeness, consistency, relevance, availability, and cost as a data-quality characteristic.

2 questions test this
Structured data fits a fixed schema of rows and columns

Structured data has a predefined schema and lives in tables of rows and columns, such as sales transactions, CRM records, and inventory tables. Its fixed shape makes it easy to query and aggregate, which is why analytics has always leaned on it, and it typically sits in a relational database or a data warehouse.

Trap Calling a spreadsheet or database table unstructured; a fixed schema makes it structured.

6 questions test this
JSON, XML, and log files are semi-structured, not unstructured

Semi-structured data has some organization but no rigid table shape, and its attributes vary from record to record. JSON and XML documents, log files, and sensor feeds are the classic examples, sitting between structured tables and free-form content because their tags describe the data without a fixed schema.

Trap Labeling JSON, XML, or log files as unstructured; their self-describing tags make them semi-structured, a favorite exam distractor.

Unstructured data has no predefined schema

Unstructured data carries no schema at all: emails, PDFs, images, audio, and video. It cannot be dropped into neat rows and columns, which is exactly why traditional analytics struggled with it and why so much of it went unused for years.

11 questions test this
Most enterprise data, up to about 80%, is unstructured

Roughly up to 80% of an organization's data is unstructured, and most of it historically sat unanalyzed because it did not fit structured tables. That share is the whole reason gen AI's ability to use unstructured content matters so much at a business level.

Gen AI's real unlock is reading unstructured data directly

Multimodal gen AI can read documents, images, audio, and video directly, so the headline business value is turning a previously dead archive into something you can search, summarize, and ask questions of. This is a step change from older analytics, which first needed the data reshaped into structured tables.

Trap Assuming gen AI's main value is faster analysis of structured tables; its distinctive unlock is using the far larger pool of unstructured data.

1 question tests this
Structured data is still best for analytics and reliable grounding

Even as the minority of enterprise data, structured data stays the backbone of analytics and the most dependable source to ground a model on, because its fixed schema keeps values precise and queryable. The takeaway is not to abandon structured data but to add the unstructured majority that gen AI newly unlocks.

Unstructured data lands in Cloud Storage; structured in BigQuery

At a business level, Google Cloud stores unstructured objects of any format in Cloud Storage, its object store, while structured data for analytics lives in BigQuery, a serverless data warehouse. Knowing which service pairs with which data type helps a leader scope where an organization's AI-ready data already sits.

6 questions test this
Labeled data pairs each example with its correct answer

Labeled data attaches the right output to each example, such as emails marked spam or not spam, or images tagged by the object they contain, and it is what supervised learning trains on. Because people usually create the labels, producing them is the slow, expensive part of classical machine learning.

7 questions test this
Unlabeled data is raw data with no answers attached

Unlabeled data has no annotations marking the correct output, and it is the vast majority of what any organization collects. On its own it cannot train a classic supervised model, but it is exactly what foundation models learn from during pre-training.

1 question tests this
Foundation models cut the need for large labeled datasets

Foundation models are pre-trained on enormous amounts of largely unlabeled data, so the heavy learning happens before any of your labels exist. That lets an organization get value with little or no labeling and only add a smaller labeled set when it chooses to fine-tune, shifting the question from 'can we afford to label millions of examples' to 'which data can we reach and trust'.

Trap Assuming a foundation model still needs a large hand-labeled dataset before it is useful; pre-training on unlabeled data is what removes that upfront requirement.

Labeled data still matters, mainly for fine-tuning and evaluation

Foundation models reduce but do not eliminate the need for labels: a smaller labeled set is what you use to fine-tune a model to a specific task or to measure output quality. Labeling moved from a mandatory upfront tax to an optional investment made only when accuracy demands it.

References

  1. Auto data quality overview
  2. BigQuery overview
  3. Cloud Storage overview
  4. Unifying data and AI to bring unstructured data analytics to BigQuery Blog
  5. What are foundation models?