Who is accountable when a sales dashboard double-counts customers, a support system uses the wrong renewal date, or a financial report contains a transaction that should never have entered the data set? The answer cannot simply be “IT” or “everyone.” Data quality is a shared operational responsibility, but each important data product needs a named business owner who can decide what correct data means, a steward who can measure and coordinate quality, and the teams that create and process records must fix problems at their source.
That distinction matters because organizations can buy validation software, schedule a hundred automated checks and still have no one empowered to decide whether conflicting records should be corrected, excluded, merged or preserved. A practical accountability model starts with decisions and consequences, not software labels.
Start with the business decision the data supports
Data is not high quality in the abstract. A phone number can be perfectly formatted yet obsolete; a customer address can be sufficiently accurate for region-level analysis but unusable for delivering an order. Before assigning ownership, ask what decision or operation depends on the data, what failure would cost, and how current and precise the record must be.
For example, suppose the sales team reports 12,000 active subscribers while the billing system shows 11,400 paying accounts. Neither figure is automatically wrong. The first might include trials, suspended accounts or multiple subscriptions; the second might exclude late payments. The first quality task is to agree on the definition of “active paying subscriber,” the observation date, and the source of truth for each field. Only then do duplicate detection and automated completeness checks become meaningful.
Who owns which part of data quality?
Organizations use different job titles, so treat this as a workable division of responsibility rather than a universal organizational chart. Microsoft Purview’s governance model, for example, distinguishes business-oriented owners from stewards who maintain standards, quality and discoverability.
| Role | Accountability or task | Practical evidence |
|---|---|---|
| Business data owner | Approves the data product’s meaning, acceptable error thresholds, priority and risk decision. | A named owner for customer, finance, product or another domain; approved definitions and exception decisions. |
| Data steward | Coordinates rules, glossary terms, lineage, issue triage and remediation across teams. | Rule catalog, quality scorecard, issue log and escalation history. |
| Data producer or application owner | Prevents invalid records where they are created and repairs source-system defects. | Input validation, source fixes, release tests and data contracts. |
| Data engineering and platform team | Makes quality controls, access, monitoring, lineage and recovery reliable in pipelines and storage. | Automated test runs, pipeline alerts, schema checks, backfills and recovery records. |
| Analyst or data consumer | Uses documented definitions, questions anomalies and avoids presenting uncertain figures as settled facts. | Reconciled reports, metric annotations and feedback to owners. |
| Governance, security and privacy specialists | Set cross-domain requirements and advise on legal, retention, classification and access constraints. | Policies, review findings and approved controls; not an implied replacement for business ownership. |
One person may hold several roles in a small organization. The essential property is that each critical issue has one accountable decision-maker and a clear path to the team that can make the technical change. A central data office can define common standards without becoming the owner of every business meaning.
Measure the dimensions that could actually fail
A useful quality scorecard separates different failure modes rather than compressing everything into one opaque percentage.
- Completeness: Are required values present? Measure missing invoice identifiers or missing shipping postcodes against records where the field should exist. An optional field should not count as a defect merely because it is blank.
- Validity: Does the value satisfy a documented rule, such as a recognized country code, nonnegative quantity or permitted transaction state? Valid values are not necessarily accurate.
- Uniqueness: Are two records representing the same business entity where only one should exist? Distinguish true duplicates from legitimate multiple accounts.
- Consistency: Do systems agree under the same definition and time window? A nightly snapshot and a streaming dashboard will differ unless their cutoffs align.
- Timeliness: Is the data recent enough for its use? A daily inventory feed might be adequate for planning and dangerously stale for a live stock promise.
- Accuracy: Does a recorded value correspond to the real entity or event? Testing accuracy usually needs a trusted reference or source confirmation, not just SQL syntax.
For each metric, record a denominator, a threshold, a measurement interval and an owner. “99% complete” is meaningless without knowing which fields and records were assessed. Do not declare a whole data product healthy because one easy metric scores well.
Use a concrete defect to connect ownership to remediation
Imagine an online retailer discovering that order totals in its revenue dashboard are higher than the finance ledger. A join between orders and order_items multiplied an order-level total by the number of items. The pipeline ran successfully, all required fields were populated and the dashboard looked plausible. This is a semantic error, not a missing-value problem.
- The finance data owner defines the correct metric: paid order value after recorded refunds, for a specific accounting period.
- The steward opens a tracked issue, identifies affected dashboards and captures the incorrect-versus-correct totals.
- The data engineer fixes the aggregation grain so item-level rows are aggregated before joining to order-level measures.
- The analyst reconciles a representative set of orders with source transactions and documents the reporting impact.
- The producer and engineer add a regression test so future schema or query changes cannot silently multiply totals again.
This is why a metric cannot be trusted merely because the pipeline is green. Production pipeline quality controls should test business invariants as well as system availability.
Put guardrails where mistakes enter the system
Quality checks should exist at more than one point. Application forms can reject impossible dates or duplicate identifiers; ingestion can quarantine records with unexpected schemas; transformations can enforce referential and business rules; curated products can reconcile key totals against trusted source systems. The right controls depend on the failure mode.
A simple PostgreSQL-style example illustrates complementary tests:
-- Find orders whose customer reference has no matching customer.
SELECT o.order_id
FROM orders AS o
LEFT JOIN customers AS c ON c.customer_id = o.customer_id
WHERE c.customer_id IS NULL;
-- Detect repeated external transaction identifiers.
SELECT external_transaction_id, COUNT(*) AS occurrences
FROM payments
WHERE external_transaction_id IS NOT NULL
GROUP BY external_transaction_id
HAVING COUNT(*) > 1;
The first test identifies referential mismatches; a properly configured enforced foreign key can also prevent this class of invalid insert. The second identifies candidates for investigation, not automatic deletion: retries, partial settlements and system-generated duplicates require different treatment. In PostgreSQL, primary keys, foreign keys, unique constraints and check constraints are useful integrity controls, but they cannot establish every business meaning. Indexing or transaction rules cannot rescue a wrong definition.
Resolve disputes with a data contract and a source-of-truth rule
A data contract records the schema, expected meaning, allowed values, freshness expectations and change responsibilities of a shared data interface. It should specify who is alerted when upstream producers make a breaking change and what consumers are allowed to assume. Contracts work best when paired with a living business glossary, versioned transformations and a known owner.
“Source of truth” is not necessarily a single database for every field. Finance may own posted payment status, the customer system may own contact preferences, and an identity service may own canonical customer IDs. The governance decision is to document field-level authority, reconciliation rules and how late-arriving updates are handled.
For an example of connecting quality tests to the data product people actually use, see data quality and observability in Fabric; the same ownership principles apply beyond that platform.
Make lineage and monitoring part of accountability
When a dashboard changes, the steward needs to trace it through transformations to source fields. Lineage turns a vague complaint into a bounded investigation: which upstream table changed, which scheduled jobs consumed it and which reports may need correction? Monitoring should alert on meaningful failures such as unexpected row-count shifts, missing daily partitions, schema incompatibilities and reconciliation gaps. It should not bury owners under thousands of threshold alerts without a triage plan.
Microsoft Purview’s catalog and data-quality documentation describe profiling, rule configuration, scoring, monitoring and owner/steward workflows. Those features support a program; they do not automatically establish the organization’s authority model. A useful starting point is a data-governance implementation in which policy, metadata and operational remediation are linked.
How to run an incident when a number is wrong
- Contain: Label suspect dashboards or exports, assess whether downstream decisions should pause and preserve the affected data versions.
- Assign: The domain owner accepts or rejects the business risk; the steward coordinates investigation and keeps stakeholders informed.
- Diagnose: Determine whether the issue is input, identity matching, transformation logic, late-arriving data, permissions or report interpretation.
- Correct: Repair the responsible system or transformation; plan backfills and reconciliations without overwriting audit history.
- Verify: Compare representative records and aggregate totals against trusted independent evidence; rerun dependent tests.
- Prevent: Add regression checks, clear ownership and a useful alert threshold; record the root cause and close the incident only after validation.
A data incident is not closed because a ticket changes status. It is closed when the actual consumer-facing result is correct, the affected period is understood and future recurrence is meaningfully less likely.
Privacy and security are related, not identical
Good data governance also covers access, retention, lawful use and protection of personal data. But a secure, access-controlled system can contain incorrect balances, and an accurate data set can be exposed inappropriately. The security function may approve access patterns and incident escalation while the business owner remains accountable for data meaning. Where sensitive personal information is involved, collect only what is needed for the task and apply approved retention and correction procedures; do not create a “quality” justification for broad unnecessary data access.
A practical first-month improvement plan
Week 1: Identify one critical data product, its business decision, owner, steward, source applications and known consumer reports. Agree on five plain-language definitions before collecting more metrics.
Week 2: Profile representative records, quantify the most damaging two or three defects, identify their upstream causes and establish baselines. Set thresholds from actual business impact instead of copying a generic 99% target.
Week 3: Implement one source prevention rule and at least one downstream reconciliation. Create a visible issue workflow with severity, affected products and named remediators.
Week 4: Review exceptions with the owner, validate corrected outcomes, document lineage and decide which automated checks deserve ongoing operation. Extend to other data products only when ownership and definitions remain clear.
The accountable answer
Data quality succeeds when responsibility is shared for producing and checking records but unambiguous for decisions. Business owners approve meaning and acceptable risk. Stewards coordinate the standards and evidence. Producers and engineers prevent and repair technical defects. Analysts surface anomalies and use documented definitions. A governance team makes the system consistent. Name those people, prove quality with real business examples and make corrections at the source: that is more durable than assigning data quality vaguely to “IT.”