The six dimensions of data quality that determine AI success

“Is our data good enough for AI?” has no single answer. Data can be complete and wrong, accurate and out of date, or clean in every system and still disagree between them. Each of those is a different way to fail, and each needs its own measure. Here are six, what each means for AI, and how to weight them for the use case in front of you.

Understanding these dimensions helps organisations diagnose specific readiness gaps rather than treating data quality as a monolithic problem. Each dimension requires different assessment techniques and different remediation approaches. An organisation might excel on some dimensions while failing on others, and knowing which dimensions are problematic is essential for effective remediation planning.

What is AI-ready data?

AI-ready data is not the same as clean data. It is data fit for a specific use: enough history, consistent definitions, a representative sample and someone accountable for keeping it that way. A database can pass every traditional quality check and still be unsuitable for a model, if every customer comes from one industry or one period, or if the behaviour you want to predict was never recorded. AI readiness is a relationship between the data and the use case, not a fixed property of the data: data that is ready for one application can be inadequate for another. The six dimensions below are how to test that relationship.

Dimension 1: completeness

Completeness measures whether sufficient data exists to support the intended analysis. This is perhaps the most intuitive dimension, but it is often measured incorrectly. Organisations track field completion rates without understanding whether those rates are adequate for AI applications.

Field-level completeness

What percentage of records have values for each required field? A customer churn model that needs email engagement data will fail if most customer records lack email addresses. A demand forecasting model that needs promotional calendar data will fail if promotional events are inconsistently recorded.

Field-level completeness should be measured for each field required by the AI use case, not for all fields in the database. A database might have high overall completion rates while critical fields for the AI application are sparsely populated.

Record-level completeness

What percentage of records have all required fields populated? A record missing three of ten required fields might be unusable even if individual field completion rates look acceptable. As a simple illustration, if 90% of records have field A populated and 90% have field B populated, but different records are missing each field, only 81% of records have both fields.

For AI applications, partial records often cannot be used. The model needs complete feature vectors, meaning all input features must be present. Record-level completeness directly determines how many training examples are available.

Coverage completeness

Does the data cover all scenarios the model must handle? If your training data lacks examples of enterprise customers, your model will perform poorly on enterprise prospects. If your training data is all from North America but you are deploying globally, the model may fail in other regions.

Coverage completeness is particularly challenging when the target outcome is rare. In fraud detection, fraudulent transactions might be a tiny fraction of total volume. You need sufficient examples of fraud across different fraud types, amounts, and contexts to train an effective model.

Temporal completeness

Does historical data extend far enough back with sufficient continuity? A customer churn model might need two to three years of behavioural history. A demand forecasting model for seasonal products might need multiple years of seasonal data. Gaps from system migrations or changed collection practices can undermine model training.

Temporal completeness also means consistent collection over time. If data collection practices changed, older data might be missing fields that newer data has, or vice versa.

Why completeness matters for AI

For AI, completeness requirements are often more demanding than for traditional reporting. A business intelligence dashboard might work fine with patchy coverage. A machine learning model often needs near-complete coverage of the fields it depends on to avoid biased predictions.

Incomplete data does not just reduce the amount of training data available. It can introduce systematic bias if incompleteness is correlated with the outcome being predicted. If customers who churn are more likely to have incomplete records, training on complete records only will produce a model that systematically underestimates churn risk.

Dimension 2: consistency

Consistency measures whether the same real-world entity is represented the same way across all systems and over time. Enterprise data typically lives in multiple systems that were implemented at different times with different standards. Reconciling these systems for AI applications reveals inconsistencies that traditional reporting might tolerate.

Cross-system consistency

Does the same customer have the same identifier in your CRM, ERP, and marketing automation system? Does revenue mean the same thing in your sales system and your finance system? Cross-system consistency is particularly challenging in enterprises with multiple applications accumulated through organic growth and acquisitions.

Entity resolution, the process of determining whether records in different systems refer to the same real-world entity, is often more difficult than organisations expect. Name variations, address changes, and missing identifiers make matching uncertain. And uncertain matching means uncertain features for AI models.

Temporal consistency

Have definitions changed over time? If your definition of an active customer changed in 2023, comparing pre-2023 and post-2023 data will produce misleading results. If product categories were reorganised in 2022, sales trends that span that reorganisation are unreliable.

Temporal inconsistency is particularly insidious because it can be invisible in the data itself. The numbers look reasonable, but they mean different things in different periods. AI models learn patterns from this data without understanding the definitional shifts.

Definitional consistency

Are business rules applied consistently? If some sales reps classify deals differently than others, your win rate predictions will be unreliable. If some support agents categorise tickets differently, your support analytics will be noisy.

Human judgment introduces variability that machines cannot easily account for. Training data that reflects inconsistent human decisions teaches the model to reproduce that inconsistency.

Why consistency matters for AI

Machine learning algorithms are pattern-matching engines. They find patterns in whatever data you give them. If the data contains inconsistencies, the model learns those inconsistencies as if they were real patterns. The result is predictions that seem plausible but are fundamentally unreliable.

Cross-system inconsistency creates particular problems when AI applications need to combine data from multiple sources. A customer 360 view that cannot reliably link CRM data to product usage data to support interactions will produce fragmented features that miss important signals.

Dimension 3: accuracy

Accuracy measures whether data correctly represents the real-world entities and events it describes. Data can be complete, consistent, and formatted correctly while still being wrong.

An address field might be complete and consistently formatted but contain the wrong address. A transaction might be recorded in the system but with the wrong amount. A customer might be flagged as churned when they actually renewed through a different channel. A product might be categorised incorrectly. A date might be wrong due to timezone handling issues.

Measuring accuracy

Measuring accuracy is challenging because it requires comparison against authoritative sources or validation through business process review. Unlike completeness or consistency, accuracy cannot be assessed by examining the data alone.

Common approaches include comparison against authoritative external sources, such as validating addresses against postal databases. Sampling and manual verification provides ground truth for a subset of records. Process review identifies where errors could be introduced. Trend analysis detects anomalies that suggest accuracy issues.

Why accuracy matters for AI

For AI, accuracy directly impacts model quality. Inaccurate training data teaches wrong patterns. If even a small share of your customer addresses are wrong, your model learns geographic patterns based partly on noise.

Inaccurate prediction inputs produce wrong outputs. Even if the model is well-trained, feeding it inaccurate data at prediction time produces unreliable predictions.

Inaccurate outcome labels make it impossible to evaluate model performance. If you cannot reliably know what actually happened, you cannot know whether the model predicted it correctly.

Dimension 4: timeliness

Timeliness measures whether data is sufficiently current for its intended use. Different AI use cases have vastly different timeliness requirements, and understanding these requirements is essential to assessing readiness.

Currency

How old is the data? Customer contact information that is two years old is likely outdated for engagement purposes. Competitive pricing data that is a week old might be stale for dynamic pricing applications. Inventory levels that are hours old might be useless for real-time availability promises.

Update frequency

How often is data refreshed? Daily sales data might be adequate for demand planning but insufficient for real-time inventory optimisation. Weekly customer activity updates might be fine for monthly churn prediction but inadequate for real-time intervention triggers.

Latency

How long between an event occurring and data being available? Same-day fraud detection requires near-real-time data pipelines. An hour of latency might be acceptable for some applications but fatal for others.

Why timeliness matters for AI

For training data, timeliness affects whether historical patterns remain relevant. If customer behaviour has fundamentally changed, older data might teach outdated patterns. The COVID-19 pandemic rendered many pre-2020 behavioural models obsolete.

For prediction inputs, timeliness determines whether predictions are actionable. A fraud alert that arrives an hour after the transaction is useless. A churn prediction based on data that is two weeks old might miss recent warning signs.

Dimension 5: validity

Validity measures whether data conforms to defined formats, ranges, and constraints. This is the dimension most closely aligned with traditional data quality, but it remains essential for AI applications.

A date field containing text values is invalid. A price field with negative values might be invalid depending on business rules. An email field without an @ symbol is invalid. A percentage field with values above 100 is invalid. A state code that does not match any actual state is invalid.

Types of validity checks

Data type validity ensures fields contain the expected data types. Numeric fields should contain numbers. Date fields should contain dates.

Range validity ensures values fall within expected ranges. Ages should be positive and below some reasonable maximum. Prices should be non-negative.

Format validity ensures values conform to expected patterns. Phone numbers should match expected formats. Email addresses should contain @ symbols.

Referential validity ensures foreign keys reference existing records. A customer ID in an order should correspond to an actual customer.

Business rule validity ensures data complies with domain-specific rules. An order ship date should not precede the order date.

Why validity matters for AI

Invalid data can cause technical failures. Model training might crash when encountering unexpected data types. Runtime predictions might fail when feature engineering encounters invalid values.

Invalid data can also produce nonsensical predictions without causing technical failures. A model might happily process a negative age or an impossible date, producing predictions that are technically valid but substantively meaningless.

Dimension 6: governance

Governance measures the organisational capability to maintain data quality over time. The other five dimensions are point-in-time measurements. Governance determines whether quality will persist or degrade.

Without governance, data quality improvements are temporary. Quality degrades as systems change, processes evolve, and personnel turn over. Governance provides the organisational infrastructure to maintain quality continuously.

Ownership clarity

Is someone accountable for each critical data element? Data without clear ownership inevitably degrades. When no one is responsible for data quality, no one ensures quality is maintained. Ownership should include authority to enforce quality standards and accountability for quality outcomes.

Documentation adequacy

Are data definitions, business rules, and system integrations documented? Without documentation, institutional knowledge walks out the door with employee turnover. New team members make decisions that inadvertently violate undocumented assumptions. Documentation should cover data definitions, business rules, quality standards, and integration logic.

Quality monitoring

Are there automated checks that detect quality degradation? Without monitoring, problems are discovered only when AI models fail or business users complain. By then, significant damage may have occurred. Monitoring should cover all critical quality dimensions with alerting that enables rapid response.

Change management

Is there a process to assess data impact when systems or business processes change? Unmanaged change is a primary cause of data quality problems. A system upgrade, a process change, or a new data source can introduce quality issues that propagate undetected until they undermine AI applications.

Why governance matters for AI

AI systems are not one-time projects. They require ongoing maintenance as data patterns shift, business requirements evolve, and model performance degrades. Governance provides the organisational infrastructure for this ongoing maintenance.

Without governance, AI models degrade over time as the data they depend on degrades. Model retraining on degraded data propagates quality problems into future predictions. The result is AI systems that work initially but become unreliable over time.

Weighting dimensions for AI use cases

Each dimension matters differently depending on the AI application. Assessment should score each dimension independently, but the overall readiness score must weight dimensions based on the specific use case requirements.

A recommendation engine might tolerate some inconsistency but require high completeness. Without sufficient examples of user-item interactions, recommendations will be poor regardless of how consistent the data is.

A fraud detection model might prioritise timeliness over perfect accuracy. A slightly noisy signal delivered in real-time is more valuable than a perfect signal delivered hours later.

A demand forecasting model might prioritise temporal consistency over completeness. Better to have consistent data over a shorter period than inconsistent data over a longer period.

Effective assessment requires understanding both the general requirements of AI and the specific requirements of the intended use case. Generic data quality scores provide false confidence or unnecessary alarm. Use-case-specific assessment provides actionable insight.