1.5 Data Quality, Data Accessibility, and Labeled vs. Unlabeled Data
Key Takeaways
- The exam guide names six data-quality and accessibility characteristics: completeness, consistency, relevance, availability, cost, and format.
- Quality failures propagate: incomplete or inconsistent source data produces confidently wrong grounded answers, which is more damaging than a model that declines to answer.
- Accessibility is an organizational problem, not a storage problem - data that exists but is locked in a silo, an unreadable format, or an unclear permission model is not usable for gen AI.
- Labeled data carries a known correct answer and enables supervised training and evaluation; unlabeled data is the raw material for pretraining, embeddings, and retrieval.
- Retrieval-augmented generation reduces the need for labeled training data but raises the bar on data quality, because the model now quotes the corpus directly.
1.5 Data Quality, Data Accessibility, and Labeled vs. Unlabeled Data
Executive Summary: Generative AI does not repair bad data; it amplifies it and speaks it fluently. The exam guide asks leaders to explain the characteristics and importance of data quality and data accessibility using a specific vocabulary - completeness, consistency, relevance, availability, cost, and format - and to distinguish labeled from unlabeled data. These are business-risk concepts, not storage trivia: each characteristic maps to a distinct failure mode a leader will be asked to prevent.
The Six Characteristics, and What Each One Costs You
| Characteristic | Definition | Failure mode in a gen AI system |
|---|---|---|
| Completeness | Required fields and records are present, with no systematic gaps | A grounded assistant answers confidently from the 60% of contracts that were digitized and silently omits the 40% still on paper |
| Consistency | The same entity is represented the same way across systems and time | "Acme Corp.", "ACME Corporation", and "Acme Inc." are treated as three customers, so retrieval returns a third of the relevant history |
| Relevance | The data actually bears on the decision being supported | A policy chatbot grounded on the entire shared drive surfaces a 2011 draft memo alongside the current policy |
| Availability | The data can be reached, at the freshness the use case requires | A pricing agent grounded on a warehouse refreshed nightly quotes yesterday's inventory as if it were live |
| Cost | The economic burden of acquiring, storing, moving, and processing the data | Egress, storage, and re-embedding costs make a corpus refresh uneconomic, so the index quietly goes stale |
| Format | The data is in a structure that downstream systems can parse | Critical knowledge lives in scanned PDFs and screenshots that no text pipeline can read without Document AI or Cloud Vision |
The compounding point a leader must make: these characteristics interact. A corpus that is complete but inconsistent retrieves poorly. A corpus that is relevant but stale is worse than no corpus, because the answers look authoritative. Grounding a model on defective data converts a visible failure (the model says "I do not know") into an invisible one (the model cites a wrong document and the reader believes it).
Data Accessibility Is an Organizational Problem
Executives often assume that because data exists somewhere in the enterprise, it is available. In practice four barriers dominate:
- Silos and ownership. The data sits in a business unit that has no incentive to share it and no budget line for supporting a new consumer.
- Permissions. Even where access is technically possible, nobody can state who is entitled to see which rows. Gen AI makes this urgent: an assistant that inherits over-broad access will happily summarize a salary spreadsheet for the wrong employee. Access control must be enforced at retrieval time, so the assistant only ever sees what the asking user is entitled to see.
- Format. Scanned documents, images, audio recordings, and proprietary binary formats need a conversion step. Google Cloud addresses this with Document AI for structured document parsing, Cloud Vision for images, and Speech-to-Text for audio.
- Cost and gravity. Very large datasets are expensive to move. This is why architectures increasingly bring the model to the data - for example querying through BigQuery - rather than exporting the data to the model.
Labeled vs. Unlabeled Data
Labeled data carries the known correct answer alongside the input: a claim plus its adjudicated category, an email plus spam or not spam, a support transcript plus its resolution code.
Unlabeled data is raw content with no attached answer: the text of every contract, an archive of call recordings, a decade of product manuals.
| Labeled data | Unlabeled data | |
|---|---|---|
| Cost to produce | High - requires human annotation or a historical process that recorded outcomes | Low - it is the exhaust of normal operations |
| Volume typically available | Thousands to hundreds of thousands of records | Millions of documents; effectively the whole enterprise archive |
| Primary gen AI use | Supervised fine-tuning, evaluation sets, safety benchmarks | Pretraining, embeddings, retrieval corpora for RAG |
| What it enables | Measuring whether the system is right | Giving the system something true to say |
The strategic insight the exam rewards: generative AI shifted the balance of value toward unlabeled data. A traditional supervised project could not start without labels. A retrieval-augmented assistant can deliver value against an entirely unlabeled document corpus, because a pretrained foundation model already supplies the language capability and the corpus only needs to supply the facts.
But labels do not become optional. You still need a labeled set to evaluate. Without a few hundred question-and-verified-answer pairs, an organization has no defensible way to state whether its assistant is accurate, and no way to detect regression when the underlying model is upgraded. Budget for an evaluation set even when no training labels are required.
Worked Scenario: A Grounded Policy Assistant
A 14,000-employee utility wants an HR assistant grounded on its policy library.
- Completeness check: 22 of 190 policies exist only as printed binders in regional offices. Those 22 include the field-crew overtime rules - the most frequently asked topic. Digitizing them is the highest-value data task in the project.
- Consistency check: three policies use "supervisor," "line manager," and "crew lead" for the same role. Without an alias mapping, retrieval misses two of every three relevant passages.
- Relevance check: the shared drive holds nine superseded versions of the leave policy. Filtering to current, approved versions matters more than adding new documents.
- Availability check: policies change quarterly, so the index needs a quarterly refresh with a named owner - otherwise the assistant confidently quotes retired rules.
- Cost check: re-embedding 190 policies quarterly is trivially cheap; re-embedding 40 million customer emails would not be, which is a reason to scope the corpus tightly.
- Format check: the binders are scanned to PDF images. Document AI converts them to structured, machine-readable text before ingestion.
- Labels: none are needed for training. The team still writes 150 verified question-and-answer pairs as an evaluation set to measure accuracy before launch and after every model upgrade.
Strategic Leadership Guidance: Exam Tips and Common Pitfalls
[!TIP] Exam Tip: When a scenario blames the model for wrong answers but describes a stale, partial, or inconsistent source corpus, the correct remediation is a data action - digitize, deduplicate, filter to current versions, refresh the index - not a bigger model or a better prompt.
Pitfall 1: Confusing unlabeled with unusable. Unlabeled data is the primary fuel for retrieval-augmented generation and embeddings. It is the most abundant and most underused asset most enterprises hold.
Pitfall 2: Skipping the evaluation set because no training labels are needed. RAG removes the need for training labels, not for measurement labels.
Pitfall 3: Treating access control as a later phase. Permissions must be enforced when documents are retrieved, so each user only sees what they are entitled to. Retrofitting this after launch is far more expensive than designing it in.
A grounded HR assistant gives confident but wrong answers about overtime rules. Investigation shows the model and prompts are sound, but the field-crew overtime policies were never digitized and exist only as printed binders. Which data characteristic has failed, and what is the correct remediation?
An enterprise plans a retrieval-augmented assistant over 80,000 unlabeled internal documents. A director argues the project cannot proceed until the documents are manually labeled. What is the accurate response?
Three source systems refer to the same customer as "Acme Corp.", "ACME Corporation", and "Acme Inc." A grounded assistant consistently returns only part of that customer's history. Which characteristic is the root cause?