Data quality metrics are numbers that tell you whether a dataset is fit for the job you want it to do. The core ones are completeness, accuracy, consistency, timeliness, validity and uniqueness. Each is a simple ratio or time difference that you calculate per table, per field and per period, then compare against a target.
The hard part is not the formulas. It is choosing the few metrics that matter for your use case, setting a threshold for each, and deciding who acts when a number drops. This guide covers all three, with worked examples you can adapt this week.
TL;DR
- Start with six dimensions: completeness, accuracy, consistency, timeliness, validity and uniqueness. Turn each into a percentage or a time measure on specific fields.
- Pick metrics by use case. A finance report and a machine learning training set need different checks, and measuring too many dimensions at once is a common mistake.
- A metric without a target and a performance band is only a number. Set thresholds and name an owner for each one.
- For AI work, add label quality, representational bias and a single source of truth to the standard list.
- According to Gartner, 59% of organizations do not measure data quality at all, so even a small scorecard puts you ahead.
Table of Contents
- Dimensions, metrics and rules: three terms, three jobs
- The core data quality metrics, with formulas
- Metrics that depend on your use case
- Data quality metrics for AI and machine learning
- Turn metrics into a framework: targets, bands and owners
- Where to run the checks
- Perspective: measure less, act more
- How Requestum can help
- Common mistakes to avoid
- FAQ
- Sources
Dimensions, metrics and rules: three terms, three jobs
These words get mixed up, and the confusion slows teams down.
- A dimension is a category of quality, such as completeness. The Frontiers survey of measurement tools describes data quality as a concept with multiple dimensions, each referring to one aspect of quality.
- A metric is the number you compute for a dimension on a specific asset, for example the share of customer records with a non-empty email.
- A rule is an actionable check that defines what good data looks like. GOV.UK guidance adds that rules should carry priority levels, so you know which failures need immediate remediation and which are acceptable for now.
The flow is simple. Pick a dimension, write a rule on a field, compute a metric from the rule's pass rate, and compare it to a target.
The core data quality metrics, with formulas
The formulas below are standard in practice. Adjust the denominators to your own definitions of a "record" or a "value".
| Dimension | What it asks | Typical metric |
|---|---|---|
| Completeness | Is the required data present? | Non-null values ÷ expected values × 100 |
| Accuracy | Does it match reality? | Verified-correct values ÷ checked values × 100 |
| Consistency | Do systems agree? | Matching records across systems ÷ compared records × 100 |
| Timeliness | Is it fresh enough? | Time data became available minus time the event occurred |
| Validity | Does it follow format and business rules? | Records passing rules ÷ total records × 100 |
| Uniqueness | Is each entity stored once? | Distinct entities ÷ total records × 100 |
Completeness
If a customer_signups table has 10,000 rows and 8,500 have a phone number, phone completeness is 85%. Always measure completeness on fields that are required for a purpose. A nullable "middle name" column at 20% is not a problem. A shipping address at 85% is.
Accuracy
Accuracy is the hardest to automate, because you need a trusted reference. Common references are a sample audited by hand, a government or vendor lookup, or a source system that you treat as the authority. Measure on a sample and report the sampling size alongside the percentage.
Consistency
Consistency catches the case where billing says a customer is active and the CRM says churned. Compare the same attribute across systems on a shared key, and count the mismatches. A rising mismatch rate often points to a broken integration, not bad data entry. Our guide on data integration strategy covers how to align use cases and governance so these mismatches stop at the source.
Timeliness
Measure data latency: the timestamp when data is available in the target system minus the timestamp of the event. Then compare it with what the consumer needs. A daily sales report can tolerate a few hours. A fraud model cannot.
Validity
Validity checks format and domain: dates parse, country codes belong to a list, order totals are not negative. This is where IBM's data science guidance is useful. It stresses that profiling tools are valuable for quick analysis but are often insufficient for complex business logic, so teams should add rule-based validation that reflects industry constraints, such as patient records linking correctly or IoT sensors reporting within physically possible ranges.
Uniqueness
Duplicates inflate counts and split customer histories. Measure the share of records that are duplicates of another, using exact keys first and fuzzy matching on name and address second.
Metrics that depend on your use case
Standardization work backs a point that experienced teams already know. ETSI's technical report on data quality metrics frames data quality as context-dependent. It advises aligning measurement criteria with business objectives, regulatory requirements and the intended use of the data asset, and avoiding generic metrics where possible. The same report defines 18 data quality metrics with formal definitions, which is a useful catalog if you outgrow the basic six.
In practice that means different priorities for different assets:
- Financial reporting: accuracy, consistency across ledgers, and timeliness at month-end close.
- Customer operations: uniqueness, completeness of contact fields, validity of addresses.
- Real-time operations such as routing or IoT: timeliness, validity of sensor ranges, completeness of event streams.
- Machine learning: the standard six plus label quality and bias, covered next.
A frequent pitfall is trying to measure every dimension on every table. Pick two to four per asset, based on the decisions that asset supports. Metrics also need to be metadata-aware: without business rules and lineage, a number such as "92% valid" cannot be interpreted.
Data quality metrics for AI and machine learning
Standard metrics are not enough when data feeds a model. The Japan AI Safety Institute's data quality management guidebook lists accuracy, completeness, consistency and timeliness as baseline quality metrics. For AI applications it also points to label quality, representational bias and single-source-of-truth concepts to prevent skewed model performance.
Here is what that looks like in a working scorecard:
- Label agreement rate: the share of items on which independent annotators give the same label. Low agreement means the model is learning noise.
- Label error rate on an audited sample: a reviewer re-checks a random slice of labeled items.
- Class and segment balance: the distribution of examples across classes, regions, or customer groups, compared to the population the model will serve.
- Source consistency: whether the same feature is computed one way, from one governed source, in training and in production.
For teams that label their own data, our article on hybrid automation and active learning for AI data labeling shows how to keep label quality measurable while cutting annotation effort.
The business case is direct. If model inputs are skewed, the outputs are too, and decisions built on them inherit the problem. The same logic applies to marketing: analytics only pays off when it rests on consistent, trusted data sources.
Turn metrics into a framework: targets, bands and owners
Measuring is half the job. Gartner's guidance notes that without performance bands and clear targets, metrics lack actionable context, and it recommends heat maps to show where issues cluster, which helps build the case for investment in automated cleaning and validation tools.
The GOV.UK framework gives a clean sequence for each data asset: understand the purpose of the data and its importance, create data quality rules, then set targets and performance bands. Its implementation guide adds that your rules should align with user needs and business goals, and that you should focus the assessment on the data's main use rather than secondary uses.
A five-step starting plan
- Choose three to five critical data assets. Pick the tables behind revenue, compliance or a model in production.
- Write the purpose of each in one sentence. "Used to calculate monthly recurring revenue" is a purpose. "Customer data" is not.
- Select two to four dimensions per asset and write one rule per field you care about.
- Set bands. For example, green at 98% or higher, amber from 95% to 98%, red below 95%. These numbers are examples, not standards: set yours with the people who rely on the data.
- Name an owner and a response time. A red metric with no owner stays red.
What a scorecard row looks like
| Asset | Field | Dimension | Rule | Target | Current | Owner |
|---|---|---|---|---|---|---|
| Orders | ship_country | Validity | Value in ISO country list | 99% | 97.4% | Fulfillment ops |
| Customers | email | Completeness | Non-empty | 95% | 91.2% | CRM manager |
| Events stream | event_ts | Timeliness | Available within 5 min | 99% | 99.6% | Data engineering |
The values above are illustrative. They show the structure, not benchmarks from any company.
Where to run the checks
Metrics are most useful when they run automatically, close to where data enters or moves. Three places pay off first:
- At ingestion: reject or quarantine records that fail validity rules before they reach the warehouse.
- In transformation: test uniqueness and referential integrity as part of each pipeline run.
- In monitoring: track freshness and volume on a schedule, and alert when a metric crosses into the red band.
Our pieces on data engineering pipelines and SLAs and data warehouse design explain how to build these checkpoints into the architecture, including why you should declare the grain of a table before choosing tools. For ownership and escalation, the 90-day data governance framework is a practical companion.
If you work in health IT, the ONC Project Playbook recommends applying data quality dimensions to each attribute when setting targets, thresholds and metrics, which fits the attribute-level approach above.
Perspective: measure less, act more
We build custom software, AI and data science solutions, and we see the same pattern on many projects. A team arrives with a long list of dimensions and a dashboard nobody opens. The data is rarely the only problem. The metrics were never tied to a decision.
Our advice is to begin from the decision. If a model will route vehicles, ask which fields would change a route if they were wrong, and measure those first. If an operations team wants to automate routine work, ask which records the automation reads, and put validity and completeness rules on them before writing a line of automation code. A discovery phase and business analysis are where this thinking belongs, because the data rules come from business rules, not from a tool.
We also push back on one habit: treating automated profiling as the finish line. Profiling shows you what the data looks like. Only a rule written with someone who knows the business can say whether it is right.
— Dmitry
How Requestum can help
If your metrics point to problems in pipelines, integrations or model inputs, we can help with the build. Our AI development services are listed at $30,000–$60,000 one-off, and our data science team can design the measurement layer behind your models. For application-level checks, our software testing and QA services cover the reliability side. Not sure where to start? Our business analysis work turns fuzzy data complaints into a prioritized list of rules.
Common mistakes to avoid
- Measuring everything. Fewer metrics, tied to decisions, get acted on.
- Using one threshold for all data. Context decides what is acceptable.
- Ignoring lineage. A failing metric is easier to fix when you know which upstream system caused it.
- Reporting averages only. A table at 97% overall can hide a field at 60%. Report per field.
- Skipping the baseline. Measure once before you change anything, or you cannot show improvement.
FAQ
What are the most important data quality metrics?
Start with completeness, accuracy, consistency, timeliness, validity and uniqueness. Which of them matter most depends on how the data is used, so choose two to four per asset rather than all six everywhere.
How do I measure data quality?
Write a rule for each field you care about, such as "value is not empty" or "value belongs to this list". Calculate the pass rate on a schedule, compare it with a target, and track the trend. For accuracy, audit a sample against a trusted reference.
What is the difference between a data quality framework and a data quality metric?
A framework is the operating model: purpose, rules, targets, performance bands, owners and escalation. A metric is one measurement inside it. Without the framework, metrics are numbers that nobody is responsible for.
How is data quality different for AI projects?
You still need the standard metrics, but you also need to measure label quality, representational bias and whether training and production use the same governed source. Skewed inputs produce skewed model performance.
Sources
- Gartner: Data Quality, Best Practices for Accurate Insights
- GOV.UK: Data quality issues framework
- GOV.UK: Implementing a data quality action plan
- ETSI TR 104 180: Development and identification of Data Quality Metrics
- Japan AI Safety Institute: Data Quality Management Guidebook
- IBM Data Science Best Practices: Quality Measurements
- Frontiers: A Survey of Data Quality Measurement and Monitoring Tools
- ONC Project Playbook: Data quality assessment