What is data quality?
Data quality is the measurable degree to which data is fit for its intended use. What counts as sufficient quality depends heavily on how and why the data is used. The same customer data might be perfectly adequate for a monthly report and completely unsuitable for a real-time recommendation engine. There is no such thing as absolute data quality — only quality relative to one or more specific use cases.
This has a practical implication: a one-time cleanup is never enough. Data and datasets are dynamic and change continuously, so their quality must be measured and maintained on an ongoing basis.
What are the dimensions of data quality?
Data quality is typically measured across five dimensions:
Completeness: Are all required data fields populated? Missing or inconsistent values are the most common and the most easily measurable problem when assessing data quality.
Correctness: Do the data points reflect reality? An outdated delivery address, for example, may be complete but is still out of date — and therefore incorrect.
Consistency: Do data points align across different systems? According to Gartner, inconsistencies between CRM, ERP, and shop systems are the hardest quality problem to resolve, because they stem from siloed environments.
Timeliness: Is the data still valid at the point of use? Data begins to go stale the moment it is captured. Keeping data current is therefore an ongoing quality concern.
Uniqueness: Does each entity exist exactly once? Duplicates — the same customer stored under three different spellings — distort every analysis.
Without these five dimensions, there is no way to determine whether your data is good or poor. For a detailed look at each dimension: The Dimensions of Data Quality.
What does poor data quality cost?
According to Gartner, poor data quality costs organizations an average of around $12.9 million per year (2020 study, widely cited to this day). The costs are spread across many areas: poor decisions based on faulty data, duplicated work, lost revenue, compliance risks, and damaged customer trust.
The 1-10-100 rule illustrates just how much timing matters when it comes to fixing errors. A data error costs roughly one unit to fix at the point of capture, around ten times as much if it is only caught once it is in the system, and up to a hundred times as much if it makes it unchecked to a decision or to the customer. Catching errors early costs a fraction of what a late correction does.
There is also a cost that is rarely quantified: time. Anaconda found in 2020 that data professionals spend an average of around 45% of their time just preparing data before they can work with it at all.
For a detailed look with case studies and sources: What Poor Data Quality Costs.
How do you recognize poor data quality?
Poor data quality rarely becomes obvious all at once. It shows up as constant friction: reports vary depending on the source, campaigns fall flat, and teams stop trusting their own numbers. A reliable warning sign is when staff start manually double-checking data before every decision.
In retail, the typical problem areas are product master data (incomplete attributes, conflicting categories), customer data (duplicates, outdated contacts), and the gaps between PIM, ERP, and shop systems.
Why is data quality a prerequisite for AI?
AI doesn't make poor data less harmful — it makes it more dangerous. An agent acts on the data it receives. If that data is flawed, so are its outputs, and faster than any human can catch. According to Gartner, poor or AI-unready data is one of the most common reasons AI initiatives are abandoned — not the model, but the data foundation underneath it. Anyone who wants to deploy agents in production needs to start with data quality.
We see this in practice. A large German retail company had to delay the go-live of Microsoft Copilot in 2023 by several months because the necessary governance was missing for a data estate of around 70 terabytes — even though the licenses had already been purchased.
Why the data foundation is the most common breaking point is explored in Why AI Projects Fail Because of Data Quality. What makes data AI-ready beyond pure quality is covered at the next level: AI Foundation.
How do you improve data quality?
A one-time cleanup is not enough, because data starts going stale again immediately afterwards. Sustained improvement comes from three things: preventing errors at the point of capture, measuring quality continuously rather than checking it once, and clearly defining who is responsible for which data.
The specific steps are covered in the individual articles in this cluster:
How good is our data, really? Assessing Data Quality How do you monitor data quality on an ongoing basis? Data Quality Monitoring Who owns data quality? Data Governance How do you clean up product data in retail? Cleaning Product Data
Data quality in Foundation Ascent, the prodct maturity model
Foundation Ascent describes an organization's journey from a solid data foundation to AI agents in productive use — across four sequential stages that cannot be skipped. Data quality belongs to the first stage.
Stage 1 The Foundation. Data quality, together with data sovereignty, forms the foundation on which all subsequent Foundation Ascent stages are built. An organization reaches this first stage when it:
- measures — rather than estimates — the five quality dimensions for its business-critical data assets,
- ensures data quality at the source rather than correcting it downstream,
- has defined ownership for its most important data objects (data ownership),
- keeps its data both controlled and compliant. More on this: Data Sovereignty.
Three further stages build on the foundation. In Stage 2, Enablement, the clean and controlled data is made usable for AI — through the right architecture, reliable pipelines, and sound governance. In Stage 3, Activation, organizational knowledge is made accessible so that AI systems and agents can draw on it directly. In Stage 4, Autonomy, agents and automation move into productive use — as the result of a solid foundation, not as an isolated quick fix.
Each stage supports the next. Until the foundation is in place, none of the stages above it can be built sustainably — and that is precisely the most common reason AI initiatives stall later on.
Where an organization stands today can be identified using the accompanying AI Readiness Score. Full overview: Foundation Ascent, the prodct Maturity Model