Data Quality Dimensions: Accuracy, Completeness, Consistency

Overview
In the PMI Cognitive Project Management for AI (PMI-CPMAI) methodology, the Identify Data Needs phase (Phase II, Data Understanding, in the CRISP-DM-aligned approach) is where the team decides whether the data available can support the AI solution defined during Business Understanding. Data quality is central to that decision. Of the many data quality dimensions, three are tested most often and matter most in practice: Accuracy, Completeness, and Consistency. This guide explains why they matter, what they are, how they are assessed, and how to handle exam questions about them.

Why It Is Important
1. AI is data-driven, not code-driven. In traditional software, behavior comes from programmed logic. In AI and machine learning, behavior comes from patterns learned from data. Poor data quality directly becomes poor model behavior. This is the classic garbage in, garbage out principle.
2. Most AI project failures are data failures. PMI-CPMAI stresses that a large share of AI projects fail or stall because of data issues, not algorithm issues. Assessing quality early prevents expensive rework later.
3. Go/No-Go decisions. CPMAI is iterative and favors early validation. If data quality is insufficient, the team may need to loop back to Business Understanding, reduce the scope, acquire new data, or stop the project. Each of these is far cheaper than discovering the problem after model deployment.
4. Trustworthy AI. Inaccurate or incomplete data can introduce bias, unfairness, and unreliable predictions, which creates ethical, regulatory, and reputational risk.
5. Cost and schedule estimation. Data preparation often consumes most of an AI project's effort, commonly cited as around 80%. Knowing the quality level helps the project manager plan realistic timelines and resources.

What They Are

1. Accuracy
Definition: The degree to which data correctly represents the real-world entity or event it describes. Accurate data is true and correct.
Examples of accuracy problems:
- A customer's birthdate recorded as 1890 instead of 1980
- Sensor readings drifting because of a faulty calibration
- Mislabeled training images, such as a cat labeled as a dog
- Typos in product prices
AI impact: Inaccurate labels or features teach the model wrong patterns. Mislabeled training data is especially harmful in supervised learning, because the model learns from errors as if they were truth.
Key idea: Accuracy asks, Is the data correct?

2. Completeness
Definition: The degree to which all required data is present. Completeness covers missing values in records, missing records, and missing attributes or features needed for the AI use case.
Examples of completeness problems:
- 30% of customer records lack an income field
- Data covers only weekdays, with no weekend transactions
- A fraud dataset has no examples of a new fraud type
- Under-representation of certain demographic groups
AI impact: Missing values reduce usable training data and may require imputation. Missing segments or classes create blind spots and bias, so the model cannot generalize to cases it never saw.
Key idea: Completeness asks, Is all the needed data there? This depends on the use case. Data can be complete for one purpose and incomplete for another.

3. Consistency
Definition: The degree to which data is uniform and non-contradictory across records, datasets, systems, and time. Consistent data follows the same formats, definitions, units, and rules.
Examples of consistency problems:
- Dates stored as MM/DD/YYYY in one system and DD/MM/YYYY in another
- 'NY', 'New York', and 'N.Y.' used for the same state
- A customer shown as active in the CRM but closed in billing
- Temperatures mixed between Celsius and Fahrenheit
- Labeling guidelines applied differently by different annotators
AI impact: Inconsistent data confuses models, inflates the number of categories, creates false patterns, and makes it hard to integrate multiple sources.
Key idea: Consistency asks, Does the data agree with itself and with other sources?

Related Dimensions (for context)
Other dimensions also appear in data quality frameworks and on the exam:
- Timeliness or currency: is the data up to date?
- Validity: does the data conform to allowed formats and ranges?
- Uniqueness: are there duplicates?
- Relevance: is the data pertinent to the problem?
Accuracy, completeness, and consistency are the core three to master.

How It Works in Practice (the CPMAI Approach)
Step 1: Link data to the business problem. Start from the AI pattern and objective defined in Phase I, such as prediction, recognition, or anomaly detection. Determine which data is required, at what volume, and at what quality threshold.
Step 2: Inventory and profile data sources. Identify internal and external sources. Use data profiling to compute statistics such as null counts, value distributions, ranges, distinct values, and format patterns.
Step 3: Assess each dimension.
- Accuracy: Compare samples against a trusted source or ground truth. Run range and plausibility checks. Audit labels, for example with inter-annotator agreement. Consult subject matter experts.
- Completeness: Measure the percentage of missing values per field. Check coverage of time periods, classes, geographies, and populations. Compare required features against available features.
- Consistency: Check format standardization, units, and naming conventions. Run cross-system reconciliation and referential integrity checks. Review whether labeling guidelines were applied the same way.
Step 4: Document findings. Record issues in a data quality report and the data requirements documentation. Quantify each problem, for example "12% missing in field X."
Step 5: Decide and plan. Determine whether quality is good enough to proceed. If not, choose a response:
- Cleanse the data (correct, standardize, deduplicate)
- Impute missing values
- Acquire or augment data, such as buying data or generating synthetic data
- Re-scope the problem
- Return to Business Understanding, which is CPMAI's iterative loop-back
Remediation tasks themselves are typically carried out in Phase III, Data Preparation.
Step 6: Establish ongoing governance. Data quality is not a one-time activity. Deployed models face data drift, so monitoring quality continues through operationalization.

Distinguishing the Three: Quick Comparison
- Wrong value present means Accuracy issue.
- Value or record absent means Completeness issue.
- Values conflict or use different formats or definitions means Consistency issue.
Example: A customer's age is blank, so completeness. It says 250, so accuracy (and validity). It says 35 in one system and 53 in another, so consistency. If you cannot tell which value is correct, the issue is consistency first. Once you check the source of truth, it becomes accuracy.

Exam Tips: Answering Questions on Data Quality Dimensions: Accuracy, Completeness, Consistency
1. Identify the symptom keyword.
- Words such as incorrect, wrong, erroneous, mislabeled, does not reflect reality, typo point to Accuracy.
- Words such as missing, null, blank, gaps, not captured, under-represented, partial point to Completeness.
- Words such as conflicting, mismatched, different formats, different units, contradictory across systems, inconsistent labeling point to Consistency.

2. Choose the most specific dimension. Scenarios may touch several dimensions. Pick the one that describes the root problem described in the question stem.

3. Favor early assessment. CPMAI answers usually favor assessing data quality before model building. Be wary of options like "start training and see how it performs" or "let the algorithm handle it."

4. Favor iteration and loop-back. If data quality is insufficient, the correct CPMAI response is often to go back to Business Understanding or Data Understanding, adjust scope, or acquire more data. Pushing forward is rarely right.

5. Do not over-engineer. Avoid answers that pick a more complex algorithm to compensate for poor data. Better data beats better algorithms is a recurring CPMAI theme.

6. Remember context dependency. Completeness and quality thresholds are defined relative to the AI use case. An option that ties quality requirements to business objectives is usually strong.

7. Labels matter. For supervised learning scenarios, mislabeled data is an accuracy issue. Different annotators applying rules differently is a consistency issue. A recommended fix is clear labeling guidelines plus quality checks.

8. Bias connection. Under-representation of groups or classes is usually framed as a completeness or coverage issue that leads to bias. Look for answers about collecting representative data.

9. Involve stakeholders. Subject matter experts and data owners are key to verifying accuracy and resolving inconsistencies. Answers involving collaboration with data stewards are often correct.

10. Watch for phase alignment. Identifying and assessing quality belongs to Data Understanding (Identify Data Needs). Fixing it through cleansing, transformation, or imputation belongs to Data Preparation. If a question asks what to do first, assessment or profiling usually comes before cleansing.

11. Eliminate extremes. Options like "discard all records with any missing value" or "cancel the project immediately" are usually wrong unless the scenario clearly justifies them. Prefer measured, documented, risk-based responses.

Sample Question Walkthrough
Scenario: A retail AI team merges sales data from three regional systems. Product categories are named differently in each system, causing the model to treat identical products as distinct. Which data quality dimension is most affected?
A) Accuracy B) Completeness C) Consistency D) Timeliness
Answer: C) Consistency. The data is not wrong or missing. It is represented differently across sources. The remedy is standardization and master data mapping.

Key Takeaways
- Accuracy means correct, completeness means present, consistency means uniform and agreeing.
- Assess quality early, in the Identify Data Needs and Data Understanding phase.
- Quality thresholds depend on the business objective and the AI pattern.
- Poor quality triggers CPMAI's iterative response: remediate, acquire, re-scope, or loop back.
- On the exam, match symptoms to dimensions, prefer data-first and iterative answers, and avoid technology-only fixes.