Gen AI Leader
Deep Dive
No foundation model, no matter how powerful, can compensate for poor-quality data. This article covers the data quality dimensions the exam tests, the difference between structured and unstructured data, and between labeled and unlabeled data — along with their business implications.
"Garbage in, garbage out" applies more strongly than ever in gen AI: the quality of the data used to train, tune, or ground a model directly determines the quality and reliability of its outputs. Google Cloud identifies six key dimensions.
Completeness
Are important fields, records, or categories missing from the dataset? Incomplete data produces models with systematic blind spots — for example, a credit risk model trained without data from certain customer segments will make biased decisions about them.
Consistency
Does the data follow the same format, units, and conventions across all sources? Dates in different formats, mixed currencies, or inconsistent field names between systems produce silent errors that are hard to detect.
Relevance
Is the data actually related to the business problem you're trying to solve? A massive dataset that's irrelevant to the use case doesn't improve the model — it can even dilute its performance on the actual task.
Availability
Is the data accessible when needed, in the right format and with the right permissions? Data trapped in departmental silos or legacy systems without an API creates delays and blocked projects.
Cost
Collecting, cleaning, labeling, and storing quality data has a real cost in time and money — a business decision is how much to invest in data preparation versus the model's expected benefit.
Format
Is the data in a format the model or training pipeline can process directly (CSV, JSON, plain text, images in a supported format), or does it require prior conversion?
Business implication
Investing in data quality isn't a "technical" step separate from the business: incomplete or inconsistent data produces models with biases, more frequent hallucinations, or wrong decisions that end up costing more (in reputation, regulatory compliance, or loss of customer trust) than the initial investment in preparing the data correctly.
Structured data follows a predefined, organized model — rows and columns, fixed schemas — that makes it easy to search and analyze directly. Unstructured data has no predefined format and requires additional processing (often with gen AI) to extract value from it.
Structured data — real examples
Unstructured data — real examples
Why it matters for gen AI
The vast majority of enterprise data is estimated to be unstructured. Generative AI is especially valuable here because it can extract, summarize, and structure useful information from this "messy" data — for example, with Document AI or OCR — something traditional analytics tools, built for structured data, couldn't do well.
Document AI
Document extractionA Google Cloud service that extracts and structures text, tables, and fields from unstructured documents (scanned PDFs, forms, contracts), turning them into data usable by a model or pipeline.
Labeled data includes a "correct answer" or category assigned alongside each example (for example, an image tagged as "cat"), while unlabeled data is raw data with no annotation attached.
| Type | Example | Typical use |
|---|---|---|
| Labeled | Emails manually marked as "spam"/"not spam"; X-ray images annotated by radiologists as "normal"/"abnormal" | Supervised learning — training the model with the explicit expected outcome |
| Unlabeled | Millions of raw web pages; uncategorized audio transcripts | Pre-training foundation models and unsupervised learning — the model discovers patterns on its own |
Business implication: the cost of labeling
Labeling data usually requires expert human review (radiologists, lawyers, linguists depending on the domain), which makes it one of the largest hidden costs of a supervised ML project. Modern foundation models reduce this dependency because they are pre-trained mainly on huge volumes of unlabeled data, and afterward only a much smaller set of labeled data is needed for task-specific fine-tuning.
Did you get this topic?
Test what you just learned
An insurance company has thousands of vehicle damage photos submitted by customers, with no category or annotation attached. A team of adjusters manually reviews a sample of 2,000 photos and classifies them as 'minor damage,' 'moderate damage,' or 'total loss' before using them to train an automatic classification model. What data transformation does this process describe?
Put what you learned to the test with original Gen AI Leader exam questions.
Start Gen AI Leader mock examUnlimited mock exams with the Pro plan
Log in to track your progress.